Vision
YOLO + Tracking
YOLO11 detection with multi-object tracking and overlay.
-
gst-launch-1.0 filesrc location=video.mp4 \
+python pyml-launch.py filesrc location=video.mp4 \
! decodebin ! videoconvertscale \
! video/x-raw,width=640,height=480 \
! pyml_yolo model-name=yolo11m \
@@ -1467,7 +1467,7 @@ YOLO + Tracking
Vision
Torch Compile
Object detection with torch.compile for optimized steady-state throughput.
-gst-launch-1.0 filesrc location=video.mp4 \
+python pyml-launch.py filesrc location=video.mp4 \
! decodebin ! videoconvert ! videoscale \
! video/x-raw,width=640,height=480 \
! pyml_objectdetector \
@@ -1480,7 +1480,7 @@ Torch Compile
Vision
Pose Estimation
YOLO pose with skeleton visualization on frame.
-gst-launch-1.0 filesrc location=video.mp4 \
+python pyml-launch.py filesrc location=video.mp4 \
! decodebin ! videoconvert \
! videoscale ! video/x-raw,width=640,height=480 \
! pyml_yolo_pose model-name=yolo11n-pose \
@@ -1491,7 +1491,7 @@ Pose Estimation
Vision
Depth Estimation
DepthAnything V2 monocular depth with colormap.
-gst-launch-1.0 filesrc location=video.mp4 \
+python pyml-launch.py filesrc location=video.mp4 \
! decodebin ! videoconvert \
! videoscale ! video/x-raw,width=640,height=480 \
! pyml_depth device=cuda \
@@ -1503,7 +1503,7 @@ Depth Estimation
Vision
Zero-Shot CLIP
Classify video frames with custom text labels.
-gst-launch-1.0 filesrc location=video.mp4 \
+python pyml-launch.py filesrc location=video.mp4 \
! decodebin ! videoconvert \
! videoscale ! video/x-raw,width=640,height=480 \
! pyml_clip device=cuda \
@@ -1516,7 +1516,7 @@ Zero-Shot CLIP
Audio
Transcribe + Translate
Korean audio to English text transcription.
-gst-launch-1.0 filesrc location=audio.wav \
+python pyml-launch.py filesrc location=audio.wav \
! decodebin ! audioconvert \
! pyml_whispertranscribe \
device=cuda language=ko \
@@ -1527,7 +1527,7 @@ Transcribe + Translate
Audio
VAD + Transcription
Voice activity gating before Whisper for lower latency.
-gst-launch-1.0 filesrc location=audio.wav \
+python pyml-launch.py filesrc location=audio.wav \
! decodebin ! audioconvert \
! audioresample \
! audio/x-raw,format=S16LE,\
@@ -1540,7 +1540,7 @@ VAD + Transcription
Language
LLM Text Generation
Run a HuggingFace LLM as a GStreamer element.
-gst-launch-1.0 filesrc location=prompt.txt \
+python pyml-launch.py filesrc location=prompt.txt \
! pyml_llm device=cuda \
model-name="microsoft/phi-2" \
! fakesink
@@ -1549,7 +1549,7 @@ LLM Text Generation
Language
Stable Diffusion
Text prompt to PNG image generation.
-gst-launch-1.0 filesrc location=prompt.txt \
+python pyml-launch.py filesrc location=prompt.txt \
! pyml_stablediffusion device=cuda \
! pngenc \
! filesink location=output.png
@@ -1558,7 +1558,7 @@ Stable Diffusion
Multi-Stream
Mux/Demux Pipeline
Batch two video streams through one model.
-gst-launch-1.0 \
+python pyml-launch.py \
filesrc location=cam1.mp4 ! decodebin \
! videoconvertscale \
! video/x-raw,width=640,height=480 \
@@ -1619,7 +1619,7 @@ Up and Running in Minutes
# Set plugin path and run
export GST_PLUGIN_PATH=$PWD/plugins:$GST_PLUGIN_PATH
-gst-launch-1.0 filesrc location=data/people.mp4 ! decodebin \
+python pyml-launch.py filesrc location=data/people.mp4 ! decodebin \
! videoconvert ! videoscale ! video/x-raw,width=640,height=480 \
! pyml_yolo model-name=yolo11m device=cuda ! pyml_overlay \
! videoconvert ! autovideosink
diff --git a/extern/rzv2h/CMakeLists.txt b/extern/rzv2h/CMakeLists.txt
new file mode 100644
index 0000000..5f09e2f
--- /dev/null
+++ b/extern/rzv2h/CMakeLists.txt
@@ -0,0 +1,54 @@
+# Build the `drpai_runtime` Python extension for RZ/V2H.
+#
+# This mirrors the SDK's apps/CMakeLists.txt (same TVM includes, same V2H
+# runtime libraries) but produces a Python module instead of an executable.
+# It MUST be configured with the SDK cross-toolchain and built inside the
+# RZ/V2H DRP-AI TVM SDK Docker. See README.md.
+#
+# Required env: TVM_ROOT (root of rzv_drp-ai_tvm), SDK (Yocto cross SDK)
+# Required -D : PYBIND11_INCLUDE_DIR, PYTHON_INCLUDE_DIR (target aarch64 python)
+cmake_minimum_required(VERSION 3.16)
+project(drpai_runtime CXX)
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+if(NOT DEFINED ENV{TVM_ROOT})
+ message(FATAL_ERROR "TVM_ROOT not set — source the DRP-AI TVM SDK env first")
+endif()
+set(TVM_ROOT "$ENV{TVM_ROOT}")
+
+set(DRPAI_APPS "${TVM_ROOT}/apps" CACHE PATH "rzv_drp-ai_tvm/apps directory")
+set(PYBIND11_INCLUDE_DIR "" CACHE PATH "pybind11 include directory")
+set(PYTHON_INCLUDE_DIR "" CACHE PATH "target python3 include dir")
+set(LIBMERA_RT_PATH ${TVM_ROOT}/obj/build_runtime/v2h/lib)
+
+add_library(drpai_runtime MODULE
+ drpai_runtime_pybind.cpp
+ ${DRPAI_APPS}/MeraDrpRuntimeWrapper.cpp
+)
+set_target_properties(drpai_runtime PROPERTIES PREFIX "" SUFFIX ".so")
+
+target_include_directories(drpai_runtime PRIVATE
+ ${DRPAI_APPS}
+ ${TVM_ROOT}/tvm/include
+ ${TVM_ROOT}/setup/include
+ ${TVM_ROOT}/tvm/3rdparty/dlpack/include
+ ${TVM_ROOT}/tvm/3rdparty/dmlc-core/include
+ ${TVM_ROOT}/tvm/3rdparty/compiler-rt
+ ${PYBIND11_INCLUDE_DIR}
+ ${PYTHON_INCLUDE_DIR}
+)
+
+add_definitions(-DMERA_DRP_RUNTIME)
+target_compile_definitions(drpai_runtime PUBLIC KDLDRPAI)
+target_link_directories(drpai_runtime PRIVATE ${LIBMERA_RT_PATH})
+target_link_libraries(drpai_runtime PRIVATE
+ mera2_runtime
+ mera2_plan_io
+ drp_tvm_rt
+ pthread
+)
+set_target_properties(drpai_runtime PROPERTIES
+ LINK_FLAGS "-Wl,-rpath,${LIBMERA_RT_PATH} -Wl,-rpath-link,${LIBMERA_RT_PATH}")
+
+target_compile_options(drpai_runtime PRIVATE -O3 -mtune=cortex-a55 -Wall -fvisibility=hidden)
diff --git a/extern/rzv2h/README.md b/extern/rzv2h/README.md
new file mode 100644
index 0000000..de8d0bc
--- /dev/null
+++ b/extern/rzv2h/README.md
@@ -0,0 +1,80 @@
+# Object detection on Renesas RZ/V2H (DRP-AI NPU)
+
+This runs `pyml_objectdetector` on the **RZ/V2H** DRP-AI NPU, using a YOLO11
+model compiled with the **DRP-AI TVM** compiler (powered by EdgeCortix MERA).
+
+It is the decomposed, metadata-passing pipeline used elsewhere in this repo —
+detector -> (tracker) -> overlay, but the detector's inference runs on the NPU:
+
+```
+... ! pyml_objectdetector engine-name=drpai model-name= device=drpai
+ input-format=nchw post-process=anchor_free
+ ! pyml_tracker ! pyml_overlay ! ...
+```
+## Prerequisites
+
+- RZ/V2H EVK with the **RZ/V2H AI SDK v6.00** Yocto image (provides the DRP-AI
+ driver, `/dev/drpai0`, GStreamer, and Python 3).
+- The **DRP-AI TVM** package (`rzv_drp-ai_tvm`) and its SDK Docker, with the
+ environment sourced so `TVM_ROOT`, `SDK` (cross SDK), and the DRP-AI
+ translator are set. (`PRODUCT=V2H`.)
+- `pybind11` headers available to the cross build.
+
+## 1 — Convert the model (in the SDK Docker)
+
+```bash
+./convert_yolo11_v2h.sh yolo11m 640
+```
+
+This exports YOLO11->ONNX (input node `images`, `1x3x640x640`) and runs the V2H
+DRP-AI TVM compiler. See the script for the exact commands.
+
+## 2 — Build the Python binding (in the SDK Docker)
+
+Source the SDK env first (so `TVM_ROOT`/`SDK` are set and CXX is the aarch64
+cross compiler), then:
+
+```bash
+cd rzv2h
+cmake -B build \
+ -DCMAKE_TOOLCHAIN_FILE="$TVM_ROOT/apps/toolchain/runtime.cmake" \
+ -DPYBIND11_INCLUDE_DIR="$(python3 -m pybind11 --includes | sed 's/-I//;q')" \
+ -DPYTHON_INCLUDE_DIR="$SDK/sysroots/aarch64-poky-linux/usr/include/python3.12"
+cmake --build build -j
+```
+
+Adjust `python3.12` to the AI SDK image's Python version, and point
+`PYBIND11_INCLUDE_DIR` at a real pybind11 headers dir if the one-liner doesn't
+resolve in the container.
+
+## 3 — Deploy to the board
+
+Copy onto the RZ/V2H (e.g. under `/home/weston`):
+
+- this repo's `plugins/` (the gst-python-ml elements),
+- `build/drpai_runtime.so`,
+- the compiled `yolo11m_drpai_v2h/` deploy dir,
+- a COCO label file if you overlay class names.
+
+```bash
+export GST_PLUGIN_PATH=/home/weston/gst-python-ml/plugins:$GST_PLUGIN_PATH
+export PYTHONPATH=/home/weston/rzv2h/build:$PYTHONPATH
+gst-inspect-1.0 pyml_objectdetector
+```
+
+## 4 — Run on the board
+
+File -> annotated file (run as a user that can open `/dev/drpai0`, often root):
+
+```bash
+gst-launch-1.0 filesrc location=clip.mp4 ! decodebin ! videoconvert ! videoscale \
+ ! "video/x-raw,format=RGB,width=640,height=640" \
+ ! pyml_objectdetector engine-name=drpai model-name=yolo11m_drpai_v2h device=drpai \
+ input-format=nchw post-process=anchor_free \
+ ! pyml_tracker tracker-type=bytetrack \
+ ! videoconvert ! "video/x-raw,format=RGBA" ! pyml_overlay \
+ ! videoconvert ! autovideosink
+```
+
+Live camera (MIPI/USB): swap `filesrc ! decodebin` for the camera source
+(`v4l2src` / the EVK's ISP source), keeping the `640x640` caps into the detector.
diff --git a/extern/rzv2h/convert_yolo11_v2h.sh b/extern/rzv2h/convert_yolo11_v2h.sh
new file mode 100755
index 0000000..06d8412
--- /dev/null
+++ b/extern/rzv2h/convert_yolo11_v2h.sh
@@ -0,0 +1,58 @@
+#!/usr/bin/env bash
+# Compile YOLO11 (ONNX) -> RZ/V2H DRP-AI (INT8) deploy dir, using the REAL
+# mera2 + DRP-AI Translator i8 + DRP-AI Quantizer flow.
+#
+# RUN INSIDE the drpai-tvm-v2h container (built via rzv2h/sdk_eval/build_image.sh),
+# with this repo mounted at /work. RZ/V2H uses the DRP-AI INT8 accelerator, so
+# quantization is MANDATORY and calibration images are required — this is why
+# the plain FP compile_onnx_model.py does NOT work for V2H.
+#
+# Usage (inside container):
+# ./rzv2h/convert_yolo11_v2h.sh [MODEL.onnx] [OUT_DIR] [CALIB_DIR] [IMGSZ]
+# Defaults assume the repo is at /work and the ONNX is exported already
+# (e.g. `yolo export model=models/yolo11m/yolo11m.pt format=onnx imgsz=640` on a
+# host with ultralytics — the container has no ultralytics).
+set -euo pipefail
+
+ONNX="${1:-/work/models/yolo11m/yolo11m.onnx}"
+OUT="${2:-/work/rzv2h/yolo11m_drpai_v2h}"
+CALIB="${3:-/work/rzv2h/calib}"
+IMGSZ="${4:-640}"
+
+: "${TVM_ROOT:?run inside the drpai-tvm-v2h container (TVM_ROOT unset)}"
+export PRODUCT=V2H
+export SDK="$(find /opt/ -name sysroots -type d | head -1)/../"
+export TRANSLATOR="$(find /opt/ -name python_api -type d | head -1)/../../"
+: "${QUANTIZER:?QUANTIZER env not set (expected from the image)}"
+export PATH="$TVM_ROOT/tutorials:$PATH" # so run_drp_compiler.sh resolves
+chmod +x "$TVM_ROOT"/tutorials/*.sh 2>/dev/null || true # SDK ships them non-+x
+
+[[ -f "$ONNX" ]] || { echo "ONNX not found: $ONNX (export it first)"; exit 1; }
+[[ -d "$CALIB" ]] || { echo "calibration image dir not found: $CALIB"; exit 1; }
+
+# The stock quant script preprocesses calibration images as ImageNet (224 +
+# mean/std) — wrong for YOLO (needs IMGSZ, /255, RGB, CHW). Patch that one line.
+python3 - "$TVM_ROOT/tutorials/compile_onnx_model_quant.py" "$IMGSZ" <<'PYEOF'
+import sys
+p, sz = sys.argv[1], int(sys.argv[2])
+s = open(p).read()
+old = "input_data = pre_process_imagenet_pytorch(image, mean, stdev, need_transpose=True)"
+new = ("input_data = (cv2.resize(image,(%d,%d))[:,:,::-1]"
+ ".astype('float32')/255.0).transpose(2,0,1)" % (sz, sz))
+if old in s:
+ open(p, "w").write(s.replace(old, new)); print("[patch] calibration preprocessing ->", sz)
+else:
+ print("[patch] calibration line already patched / not found")
+PYEOF
+
+rm -rf "$OUT"
+cd "$TVM_ROOT/tutorials"
+python3 compile_onnx_model_quant.py "$ONNX" \
+ -o "$OUT" -i images -s "1,3,${IMGSZ},${IMGSZ}" \
+ -t "$SDK" -d "$TRANSLATOR" -c "$QUANTIZER" --images "$CALIB"
+
+echo
+echo "Done. RZ/V2H DRP-AI (INT8) deploy dir: $OUT"
+echo " sub_0000__CPU_DRP_TVM/{deploy.so,deploy.json,deploy.params} (aarch64 + DRP-AI)"
+echo " preprocess/ (DRP-AI pre-processing runtime objects)"
+echo "Copy $OUT to the board; load sub_0000__CPU_DRP_TVM with the MERA runtime."
diff --git a/extern/rzv2h/drpai_runtime_pybind.cpp b/extern/rzv2h/drpai_runtime_pybind.cpp
new file mode 100644
index 0000000..5d2d0b9
--- /dev/null
+++ b/extern/rzv2h/drpai_runtime_pybind.cpp
@@ -0,0 +1,149 @@
+// drpai_runtime_pybind.cpp
+// Copyright (C) 2024-2026 Collabora Ltd. — LGPL (see COPYING).
+//
+// pybind11 binding around the Renesas DRP-AI TVM runtime
+// (MeraDrpRuntimeWrapper, powered by EdgeCortix MERA(TM)) for RZ/V2H.
+//
+// Exposes a minimal `drpai_runtime.Runtime` class to Python so the pure-Python
+// `drpai_engine.py` can drive the DRP-AI NPU:
+//
+// import drpai_runtime
+// rt = drpai_runtime.Runtime()
+// rt.load("/path/to/deploy_dir") # deploy.so/json/params
+// rt.set_input(0, nchw_float32_numpy)
+// rt.run()
+// out0 = rt.get_output(0) # numpy (float32, fp16 upcast)
+//
+// Build with CMake against the board's DRP-AI TVM runtime — see CMakeLists.txt
+// and README.md. This compiles only inside the RZ/V2H DRP-AI TVM SDK and runs
+// only on the board (it talks to /dev/drpai0).
+
+#include
+#include
+#include
+
+#include
+#include
+#include
+#include
+
+#include
+#include
+#include
+#include
+
+#include "MeraDrpRuntimeWrapper.h"
+
+namespace py = pybind11;
+
+static float fp16_to_fp32(uint16_t h) {
+ uint32_t sign = static_cast(h & 0x8000) << 16;
+ uint32_t exp = (h >> 10) & 0x1F;
+ uint32_t mant = h & 0x3FF;
+ uint32_t f;
+ if (exp == 0) {
+ if (mant == 0) {
+ f = sign;
+ } else {
+ exp = 127 - 15 + 1;
+ while ((mant & 0x400) == 0) {
+ mant <<= 1;
+ exp--;
+ }
+ mant &= 0x3FF;
+ f = sign | (exp << 23) | (mant << 13);
+ }
+ } else if (exp == 0x1F) {
+ f = sign | 0x7F800000 | (mant << 13); // Inf / NaN
+ } else {
+ f = sign | ((exp - 15 + 127) << 23) | (mant << 13);
+ }
+ float out;
+ std::memcpy(&out, &f, sizeof(out));
+ return out;
+}
+
+static uint64_t get_drpai_start_addr() {
+ int fd = open("/dev/drpai0", O_RDWR);
+ if (fd < 0) {
+ throw std::runtime_error("Failed to open /dev/drpai0 (run on the board, as root?)");
+ }
+ drpai_data_t drpai_data;
+ int ret = ioctl(fd, DRPAI_GET_DRPAI_AREA, &drpai_data);
+ close(fd);
+ if (ret == -1) {
+ throw std::runtime_error("ioctl(DRPAI_GET_DRPAI_AREA) failed");
+ }
+ return drpai_data.address;
+}
+
+class Runtime {
+ public:
+ Runtime() : rt_() {}
+
+ bool load(const std::string& model_dir) {
+ model_dir_ = model_dir;
+ return rt_.LoadModel(model_dir, get_drpai_start_addr());
+ }
+
+ void set_input(int index,
+ py::array_t data) {
+ rt_.SetInput(index, static_cast(data.data()));
+ }
+
+ void run() { rt_.Run(); }
+
+ int num_input() { return rt_.GetNumInput(model_dir_); }
+ int num_output() { return rt_.GetNumOutput(); }
+
+ py::array get_output(int index) {
+ auto out = rt_.GetOutput(index);
+ InOutDataType dtype = std::get<0>(out);
+ const void* ptr = std::get<1>(out);
+ int64_t size = std::get<2>(out);
+
+ switch (dtype) {
+ case InOutDataType::FLOAT16: {
+ const uint16_t* src = reinterpret_cast(ptr);
+ py::array_t result(size);
+ float* dst = static_cast(result.request().ptr);
+ for (int64_t i = 0; i < size; ++i) dst[i] = fp16_to_fp32(src[i]);
+ return result;
+ }
+ case InOutDataType::FLOAT32: {
+ py::array_t result(size);
+ std::memcpy(result.request().ptr, ptr, size * sizeof(float));
+ return result;
+ }
+ case InOutDataType::INT32: {
+ py::array_t result(size);
+ std::memcpy(result.request().ptr, ptr, size * sizeof(int32_t));
+ return result;
+ }
+ case InOutDataType::INT64: {
+ py::array_t result(size);
+ std::memcpy(result.request().ptr, ptr, size * sizeof(int64_t));
+ return result;
+ }
+ default:
+ throw std::runtime_error("Unsupported DRP-AI output data type");
+ }
+ }
+
+ private:
+ MeraDrpRuntimeWrapper rt_;
+ std::string model_dir_;
+};
+
+PYBIND11_MODULE(drpai_runtime, m) {
+ m.doc() = "pybind11 binding for the Renesas DRP-AI TVM runtime (RZ/V2H)";
+ py::class_(m, "Runtime")
+ .def(py::init<>())
+ .def("load", &Runtime::load, py::arg("model_dir"),
+ "Load a DRP-AI TVM deploy directory (deploy.so/json/params).")
+ .def("set_input", &Runtime::set_input, py::arg("index"), py::arg("data"))
+ .def("run", &Runtime::run)
+ .def("num_input", &Runtime::num_input)
+ .def("num_output", &Runtime::num_output)
+ .def("get_output", &Runtime::get_output, py::arg("index"));
+}
diff --git a/extern/rzv2h/emulation/drpai_runtime.py b/extern/rzv2h/emulation/drpai_runtime.py
new file mode 100644
index 0000000..6fdbe25
--- /dev/null
+++ b/extern/rzv2h/emulation/drpai_runtime.py
@@ -0,0 +1,123 @@
+# drpai_runtime.py — off-board stand-in for the native pybind `drpai_runtime`.
+# Copyright (C) 2024-2026 Collabora Ltd. — LGPL (see COPYING).
+#
+# Same interface as the C++ binding (Runtime.load / set_input / run /
+# num_output / get_output), with two backends auto-selected by what's in the
+# model directory and what's importable:
+#
+# 1. MERA / TVM graph_executor — if the dir has deploy.so/json/params AND a
+# `tvm` runtime is importable (i.e. inside the Renesas DRP-AI TVM SDK
+# container, or on the board). This runs the REAL MERA/TVM runtime — the
+# faithful "test through the TVM runtime". On the board the deploy.so runs
+# on the DRP-AI NPU / Arm CPU; in the SDK container it runs on whatever the
+# module was compiled for (aarch64 needs QEMU; an x86-target build runs
+# natively for functional check).
+#
+# 2. ONNX Runtime (CPU) — fallback look-alike for plain x86 dev boxes
+# with no SDK: runs the same yolo11m.onnx that feeds the DRP-AI compiler so
+# the engine's preprocess/reshape/decode path is exercised. Validates our
+# code, NOT the DRP-AI/MERA runtime.
+#
+# get_output() always returns a FLAT array, matching the C++ GetOutput buffer,
+# so the engine's reshape-to-(1, 4+nc, anchors) path is genuinely tested.
+
+import glob
+import os
+
+import numpy as np
+
+
+class Runtime:
+ def __init__(self):
+ self._backend = None
+ # tvm backend
+ self._mod = None
+ self._dev = None
+ self._input_name = os.getenv("DRPAI_INPUT_NAME", "images")
+ # onnx backend
+ self._sess = None
+ self._ort_input = None
+ self._feed = None
+ self._outputs = None
+
+ def load(self, model_dir):
+ deploy_so = os.path.join(model_dir, "deploy.so")
+ if os.path.isfile(deploy_so) and self._try_load_tvm(model_dir, deploy_so):
+ return True
+ return self._try_load_onnx(model_dir)
+
+ # ---- backend 1: real MERA / TVM graph_executor ----
+ def _try_load_tvm(self, model_dir, deploy_so):
+ try:
+ import tvm
+ from tvm.contrib import graph_executor
+ except ImportError:
+ return False
+ try:
+ lib = tvm.runtime.load_module(deploy_so)
+ with open(os.path.join(model_dir, "deploy.json")) as f:
+ graph = f.read()
+ self._dev = tvm.cpu(0)
+ self._mod = graph_executor.create(graph, lib, self._dev)
+ with open(os.path.join(model_dir, "deploy.params"), "rb") as f:
+ self._mod.load_params(bytearray(f.read()))
+ self._backend = "tvm"
+ print(
+ f"[drpai_runtime] MERA/TVM graph_executor backend "
+ f"(deploy.so, input='{self._input_name}') — real runtime"
+ )
+ return True
+ except Exception as e:
+ print(f"[drpai_runtime] TVM backend load failed ({e}); trying ONNX")
+ return False
+
+ # ---- backend 2: ONNX Runtime look-alike ----
+ def _try_load_onnx(self, model_dir):
+ try:
+ import onnxruntime as ort
+ except ImportError:
+ print("[drpai_runtime] no TVM and no onnxruntime — cannot load")
+ return False
+ onnx_files = sorted(glob.glob(os.path.join(model_dir, "*.onnx")))
+ if not onnx_files:
+ print(f"[drpai_runtime] no deploy.so and no .onnx in {model_dir!r}")
+ return False
+ self._sess = ort.InferenceSession(
+ onnx_files[0], providers=["CPUExecutionProvider"]
+ )
+ self._ort_input = self._sess.get_inputs()[0].name
+ self._backend = "onnx"
+ print(
+ f"[drpai_runtime] ONNX Runtime EMULATION backend ({onnx_files[0]}, "
+ f"input='{self._ort_input}') — NOT the NPU/MERA runtime"
+ )
+ return True
+
+ def set_input(self, index, data):
+ arr = np.ascontiguousarray(data, dtype=np.float32)
+ if self._backend == "tvm":
+ import tvm
+
+ self._mod.set_input(self._input_name, tvm.nd.array(arr, self._dev))
+ else:
+ self._feed = arr
+
+ def run(self):
+ if self._backend == "tvm":
+ self._mod.run()
+ else:
+ self._outputs = self._sess.run(None, {self._ort_input: self._feed})
+
+ def num_input(self):
+ return 1
+
+ def num_output(self):
+ if self._backend == "tvm":
+ return self._mod.get_num_outputs()
+ return len(self._outputs) if self._outputs is not None else 0
+
+ def get_output(self, index):
+ # Flat buffer, like the C++ GetOutput; the engine reshapes it.
+ if self._backend == "tvm":
+ return self._mod.get_output(index).numpy().reshape(-1).astype(np.float32)
+ return np.asarray(self._outputs[index], dtype=np.float32).reshape(-1)
diff --git a/extern/rzv2h/emulation/run_emulated.sh b/extern/rzv2h/emulation/run_emulated.sh
new file mode 100755
index 0000000..da44de2
--- /dev/null
+++ b/extern/rzv2h/emulation/run_emulated.sh
@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+# Run the DRP-AI object-detection pipeline on the DEV BOX using the emulated
+# drpai_runtime (CPU/ONNX Runtime stand-in) — same engine code as the board,
+# but no NPU. For validating the integration before deploying to RZ/V2H.
+#
+# Usage: ./run_emulated.sh [INPUT.mp4] [OUTPUT.mp4]
+set -euo pipefail
+HERE="$(cd "$(dirname "$0")" && pwd)"
+REPO="$(cd "$HERE/../../.." && pwd)"
+cd "$REPO"
+
+source .venv/bin/activate
+export GST_PLUGIN_PATH="$REPO/plugins:${GST_PLUGIN_PATH:-}"
+export PYTHONPATH="$HERE:${PYTHONPATH:-}" # resolves `import drpai_runtime` to the fake
+
+IN="${1:-08fd33_4.mp4}"
+OUT="${2:-${IN%.*}_drpai_emu.mp4}"
+DEPLOY="$HERE/yolo11m_drpai_v2h_emu" # dir containing yolo11m.onnx
+
+if [[ ! -f "$DEPLOY/yolo11m.onnx" ]]; then
+ echo "Missing $DEPLOY/yolo11m.onnx — export it first:" >&2
+ echo " yolo export model=yolo11m.pt format=onnx imgsz=640 opset=12 simplify=True" >&2
+ echo " mkdir -p $DEPLOY && cp yolo11m.onnx $DEPLOY/" >&2
+ exit 1
+fi
+
+echo "EMULATED DRP-AI run: '$IN' -> '$OUT' (CPU/ONNX, not the NPU)"
+gst-launch-1.0 -e \
+ filesrc location="$IN" ! decodebin ! videoconvert ! videoscale \
+ ! "video/x-raw,format=RGB,width=640,height=640" \
+ ! pyml_objectdetector engine-name=drpai model-name="$DEPLOY" device=drpai \
+ input-format=nchw post-process=anchor_free \
+ ! pyml_tracker tracker-type=bytetrack \
+ ! videoconvert ! "video/x-raw,format=RGBA" \
+ ! pyml_football_overlay show-ids=false show-labels=false \
+ ! videoconvert ! openh264enc ! h264parse ! mp4mux ! filesink location="$OUT"
+echo "Done: $OUT"
diff --git a/extern/rzv2h/sdk_eval/README.md b/extern/rzv2h/sdk_eval/README.md
new file mode 100644
index 0000000..13fa30f
--- /dev/null
+++ b/extern/rzv2h/sdk_eval/README.md
@@ -0,0 +1,113 @@
+# Faithful DRP-AI TVM eval (real mera2 / MERA runtime)
+
+This is the most faithful test short of running on hardware: the **real**
+`mera2` compile and the **real** MERA/TVM runtime, instead of the ONNX-RT
+look-alike in [../emulation](../emulation). It composes with the same
+`engine-name=drpai` + `drpai_runtime` shim we use everywhere else.
+
+## Read this first — what's gated, and the aarch64 catch
+
+Two things make this unable to run on a plain x86 box out of the box:
+
+1. **License-gated downloads (Renesas account required).** The stack build needs
+ the **DRP-AI Translator i8** and the **RZ/V2H AI SDK** (`RTK0EF0180F06000SJ.zip`).
+ There is **no public prebuilt image**; you download these and build Renesas'
+ `Dockerfile`. I cannot fetch them for you.
+2. **The compile targets aarch64, not x86.** Even `compile_cpu_only_onnx_model.py`
+ uses `target = "llvm ... -mtriple=aarch64-linux-gnu"` and the SDK's aarch64
+ cross-g++. So `deploy.so` runs on the board's Arm CPU / NPU — to execute it
+ off-board you either run on the **board**, under **QEMU-aarch64**, or compile
+ with an **x86 `llvm` target** for a pure functional check (see below).
+
+If you don't have the downloads, the ONNX-RT emulation in `../emulation`
+already validates all of *our* code (engine preprocess/reshape/decode +
+pipeline). What's left to validate here is mera2-compile success and runtime
+numerics — both inherently need Renesas assets or hardware.
+
+## Steps
+
+### 1. Build the SDK image (host, needs the two downloads)
+
+```bash
+mkdir -p rzv2h/sdk_eval/assets
+# put both Renesas downloads in rzv2h/sdk_eval/assets/ :
+# DRP-AI_Translator_i8-*-Linux-x86_64-Install and RTK0EF0180F06000SJ.zip
+cd rzv2h/sdk_eval && ./build_image.sh
+```
+
+`build_image.sh` fetches the repo `Dockerfile`, assembles a clean build context
+(Dockerfile + the toolchain `.sh` it unzips from the AI SDK zip + the Translator
+installer), and runs `docker build --build-arg PRODUCT=V2H -t drpai-tvm-v2h`.
+The Dockerfile (`FROM ubuntu:22.04`) defaults `PRODUCT=V2H` and builds the TVM
+fork itself, so the build takes a while.
+
+To fetch just the Dockerfile by hand:
+`wget https://raw.githubusercontent.com/renesas-rz/rzv_drp-ai_tvm/main/Dockerfile`
+
+### 2. Compile YOLO11 with the real mera2 (inside the container)
+
+```bash
+docker run -it --rm -v "$PWD":/workspace/gst-python-ml drpai-tvm-v2h bash
+# inside:
+cd /workspace/gst-python-ml
+./rzv2h/convert_yolo11_v2h.sh yolo11m 640 # real mera2.from_onnx + mera2.drp.build
+# -> yolo11m_drpai_v2h/{deploy.so,deploy.json,deploy.params} (aarch64)
+```
+
+For a **host x86 functional check** instead of the board artifact, compile with a
+native target (edit a copy of `tutorials/compile_onnx_model.py` to
+`target = "llvm"` and drop the aarch64 cross-compiler), producing an x86
+`deploy.so` the MERA/TVM `graph_executor` can run natively.
+
+### 3. Run through the real MERA/TVM runtime
+
+The [../emulation/drpai_runtime.py](../emulation/drpai_runtime.py) shim
+auto-selects the **MERA/TVM `graph_executor`** backend as soon as the model dir
+has `deploy.so/json/params` and `tvm` is importable (true inside this
+container). The engine code is unchanged.
+
+```bash
+export GST_PLUGIN_PATH=/workspace/gst-python-ml/plugins:$GST_PLUGIN_PATH
+export PYTHONPATH=/workspace/gst-python-ml/rzv2h/emulation:$PYTHONPATH
+# (x86 deploy.so) run natively; (aarch64 deploy.so) run under qemu-aarch64
+gst-launch-1.0 filesrc location=08fd33_4.mp4 ! decodebin ! videoconvert ! videoscale \
+ ! "video/x-raw,format=RGB,width=640,height=640" \
+ ! pyml_objectdetector engine-name=drpai model-name=yolo11m_drpai_v2h device=drpai \
+ input-format=nchw post-process=anchor_free \
+ ! pyml_tracker ! videoconvert ! "video/x-raw,format=RGBA" \
+ ! pyml_football_overlay ! videoconvert ! autovideosink
+```
+
+The shim prints which backend it picked:
+`[drpai_runtime] MERA/TVM graph_executor backend ... — real runtime`.
+
+## On the actual board
+
+Two ways to run the same pipeline on the RZ/V2H:
+
+- **Python graph_executor** — copy the `deploy.so/json/params` + the emulation
+ shim; if the board image has the MERA/TVM python runtime, it Just Works (the
+ shim's TVM backend), NPU included.
+- **C++ pybind binding** — build [../drpai_runtime_pybind.cpp](../drpai_runtime_pybind.cpp)
+ per [../README.md](../README.md); the native `drpai_runtime.so` takes
+ precedence over this shim on `PYTHONPATH`.
+
+## Verified results (RZ/V2H AI SDK v6.00 + DRP-AI Translator i8 v1.11)
+
+Both paths were run end-to-end driving the `drpai-tvm-v2h` image on an x86 host:
+
+- **x86 MERA/TVM runtime test** — `compile_x86_cpu.py` compiled YOLO11 via the
+ MERA-fork TVM (native `llvm`), and `x86_runtime_check.py` ran it through the
+ real `graph_executor`: output matched ONNX to **max|Δ| = 6.2e-3**, **22 = 22
+ detections** (label `person`). Confirms compile + MERA/TVM runtime + the
+ `drpai_runtime` shim + our decoder, no NPU needed.
+- **Real INT8 NPU compile** — `../convert_yolo11_v2h.sh` (quantized flow)
+ produced the RZ/V2H deploy dir: `[Finish DRP-AI Translator for V2H]`,
+ `sub_0000__CPU_DRP_TVM/{deploy.so (65 MB),deploy.json,deploy.params}` +
+ `preprocess/` (DRP-AI pre-processing objects). aarch64 — runs on the board.
+
+SDK gotchas the scripts now handle automatically: `run_drp_compiler.sh` ships
+non-executable and off-PATH (`chmod +x` + add tutorials to PATH); the quant
+script preprocesses calibration as ImageNet-224 instead of 640 (patched). And
+V2H **requires** the INT8 quantized flow — the plain FP `compile_onnx_model.py`
+drives a legacy translator path the i8 v1.11 layout lacks.
diff --git a/extern/rzv2h/sdk_eval/_probe_sysroot.sh b/extern/rzv2h/sdk_eval/_probe_sysroot.sh
new file mode 100644
index 0000000..0344994
--- /dev/null
+++ b/extern/rzv2h/sdk_eval/_probe_sysroot.sh
@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Probe the RZ/V2H board rootfs (via the cross-SDK aarch64 sysroot) for the
+# GStreamer + Python stack our pipeline needs. Run inside drpai-tvm-v2h.
+# Target rootfs sysroot (NOT the x86_64-pokysdk-linux cross-compiler dir).
+SR=$(ls -d /opt/*/*/sysroots/*-poky-linux 2>/dev/null | grep -v pokysdk | head -1)
+[ -d "$SR" ] || SR=$(ls -d /opt/*/sysroots/*-poky-linux 2>/dev/null | grep -v pokysdk | head -1)
+echo "sysroot = $SR"
+echo "--- python3 ---"; ls -d "$SR"/usr/lib/python3* 2>/dev/null | head -1
+echo "--- gstreamer core ---"; ls "$SR"/usr/lib/libgstreamer-1.0.so.* 2>/dev/null
+grep -h "Version" "$SR"/usr/lib/pkgconfig/gstreamer-1.0.pc 2>/dev/null
+echo "--- gst-python loader (libgstpython) ---"; find "$SR" -name 'libgstpython*' 2>/dev/null | head
+echo "--- GstAnalytics (lib + typelib) ---"
+find "$SR" -iname '*gstanalytics*' 2>/dev/null | head
+ls "$SR"/usr/lib/girepository-1.0/ 2>/dev/null | grep -iE 'Analytics|GstApp|GstBase|^Gst-' | head
+echo "--- python modules on target: gi / numpy / cairo / cv2 ---"
+for m in gi numpy cairo cv2; do
+ hit=$(find "$SR" -maxdepth 7 -path '*python3*' -iname "${m}" 2>/dev/null | head -1)
+ echo "$m: ${hit:-MISSING}"
+done
+echo "--- tvm / mera python runtime on target? ---"
+find "$SR" -iname '*tvm*' -o -iname '*mera*' 2>/dev/null | grep -i python | head
+echo "--- gstreamer plugins present (count) ---"
+ls "$SR"/usr/lib/gstreamer-1.0/*.so 2>/dev/null | wc -l
diff --git a/extern/rzv2h/sdk_eval/build_image.sh b/extern/rzv2h/sdk_eval/build_image.sh
new file mode 100755
index 0000000..fc7f073
--- /dev/null
+++ b/extern/rzv2h/sdk_eval/build_image.sh
@@ -0,0 +1,78 @@
+#!/usr/bin/env bash
+# Build the Renesas DRP-AI TVM (Mera2) Docker image for RZ/V2H.
+#
+# This is the *faithful* compile/runtime stack (real mera2 + MERA runtime).
+# It needs two downloads that require a Renesas account login — put them in
+# ./assets first (this script cannot download them for you). Both arrive as
+# ZIPs and can be dropped in as-is:
+#
+# DRP-AI Translator i8 (ZIP, contains DRP-AI_Translator_i8-*-Linux-x86_64-Install)
+# https://www.renesas.com/software-tool/drp-ai-translator-i8 (Downloads tab)
+# RZ/V2H AI SDK (RTK0EF0180F*SJ.zip)
+# https://www.renesas.com/us/en/software-tool/rzv2h-ai-software-development-kit
+#
+# The repo Dockerfile COPYs every ./*.sh in the context and runs it, plus
+# ./DRP-AI_Translator*-Install. So we assemble a CLEAN context holding only:
+# Dockerfile + the SDK toolchain installer (.sh, from the AI SDK zip) + the
+# Translator installer (from the Translator zip).
+set -euo pipefail
+cd "$(dirname "$0")"
+ASSETS="${ASSETS:-./assets}"
+CTX="${CTX:-./context}"
+PRODUCT="${PRODUCT:-V2H}"
+TAG="${TAG:-drpai-tvm-v2h}"
+
+mkdir -p "$ASSETS"
+TMPS=()
+cleanup() { for d in "${TMPS[@]:-}"; do [[ -n "$d" ]] && rm -rf "$d"; done; }
+trap cleanup EXIT
+
+# --- DRP-AI Translator i8: accept an extracted *-Install or the downloaded zip ---
+TR=$(ls "$ASSETS"/DRP-AI_Translator*-Linux*-x86_64-Install 2>/dev/null | head -n1 || true)
+if [[ -z "$TR" ]]; then
+ TRZIP=$(ls "$ASSETS"/*[Tt]ranslator*i8*.zip "$ASSETS"/*DRP-AI_Translator*.zip 2>/dev/null | head -n1 || true)
+ if [[ -n "$TRZIP" ]]; then
+ t=$(mktemp -d); TMPS+=("$t")
+ unzip -o -q "$TRZIP" -d "$t"
+ TR=$(find "$t" -iname "DRP-AI_Translator*-Linux*-x86_64-Install" | head -n1 || true)
+ fi
+fi
+
+# --- RZ/V2H AI SDK zip (any v6.x build number) ---
+ZIP=$(ls "$ASSETS"/RTK0EF0180F*SJ.zip 2>/dev/null | head -n1 || true)
+
+if [[ -z "$TR" || -z "$ZIP" ]]; then
+ echo "Missing gated downloads in $ASSETS (Renesas login required):" >&2
+ [[ -z "$TR" ]] && echo " - DRP-AI Translator i8 (zip or extracted *-Install)" >&2
+ [[ -z "$ZIP" ]] && echo " - RZ/V2H AI SDK (RTK0EF0180F*SJ.zip)" >&2
+ exit 1
+fi
+
+# Clean build context.
+rm -rf "$CTX" && mkdir -p "$CTX"
+wget -nc https://raw.githubusercontent.com/renesas-rz/rzv_drp-ai_tvm/main/Dockerfile \
+ -O "$CTX/Dockerfile"
+cp "$TR" "$CTX/"
+
+# Unzip the AI SDK and extract its Yocto toolchain installer (.sh) into context.
+s=$(mktemp -d); TMPS+=("$s")
+unzip -o -q "$ZIP" -d "$s"
+# The Yocto toolchain installer is the big *toolchain*.sh (e.g.
+# ai_sdk_setup/rz-vlp-...-rzv2h-evk-toolchain-5.0.11.sh). Pick the largest
+# match so we don't grab a small board/flash helper script by mistake.
+SDK_SH=$(find "$s" -iname "*toolchain*.sh" -printf '%s\t%p\n' | sort -rn | head -n1 | cut -f2-)
+[[ -n "$SDK_SH" ]] || { echo "No toolchain .sh found inside $ZIP" >&2; exit 1; }
+cp "$SDK_SH" "$CTX/"
+
+echo "Build context ready in $CTX:"
+ls -1 "$CTX"
+echo
+echo "Building image '$TAG' (PRODUCT=$PRODUCT) — builds the TVM fork, takes a while..."
+docker build --build-arg PRODUCT="$PRODUCT" -t "$TAG" "$CTX"
+
+cat </deploy.{so,json,params} — loadable by tvm.contrib.graph_executor, i.e.
+# by the drpai_runtime shim's TVM backend.
+#
+# Run inside the drpai-tvm-v2h container:
+# python3 compile_x86_cpu.py [input_name] [C,H,W]
+import os
+import sys
+
+import onnx
+import tvm
+from tvm import relay
+from tvm.relay import transform
+from tvm.relay.build_module import build as _build, bind_params_by_name
+from tvm.relay.param_dict import save_param_dict
+from tvm.ir.transform import Sequential, PassContext
+
+model_file = sys.argv[1]
+out_dir = sys.argv[2]
+input_name = sys.argv[3] if len(sys.argv) > 3 else "images"
+chw = [int(x) for x in (sys.argv[4].split(",") if len(sys.argv) > 4 else [3, 640, 640])]
+input_shape = [1] + chw
+
+os.makedirs(out_dir, exist_ok=True)
+print(f"[x86 compile] {model_file} input {input_name}={input_shape} -> {out_dir}")
+
+onnx_model = onnx.load_model(model_file)
+mod, params = relay.frontend.from_onnx(onnx_model, {input_name: input_shape})
+if params:
+ mod["main"] = bind_params_by_name(mod["main"], params)
+
+with PassContext(opt_level=3):
+ mod = Sequential([
+ transform.SimplifyInference(),
+ transform.FoldConstant(),
+ transform.FoldExplicitPadding(),
+ transform.BackwardFoldScaleAxis(),
+ transform.ForwardFoldScaleAxis(),
+ transform.FoldConstant(),
+ transform.DynamicToStatic(),
+ transform.RemoveUnusedFunctions(),
+ ])(mod)
+
+target = "llvm" # native host (x86), no aarch64 cross target
+with PassContext(opt_level=3):
+ graph, lib, all_params = _build(mod, target=target, target_host=target, params=params)
+
+lib.export_library(os.path.join(out_dir, "deploy.so")) # default host compiler -> x86 .so
+with open(os.path.join(out_dir, "deploy.json"), "w") as f:
+ f.write(graph)
+with open(os.path.join(out_dir, "deploy.params"), "wb") as f:
+ f.write(save_param_dict(all_params))
+print(f"[x86 compile finished] -> {out_dir}/deploy.so,deploy.json,deploy.params")
diff --git a/extern/rzv2h/sdk_eval/x86_runtime_check.py b/extern/rzv2h/sdk_eval/x86_runtime_check.py
new file mode 100644
index 0000000..d1ee523
--- /dev/null
+++ b/extern/rzv2h/sdk_eval/x86_runtime_check.py
@@ -0,0 +1,34 @@
+#!/usr/bin/env python3
+# x86_runtime_check.py — run a real input through the MERA/TVM graph_executor
+# (via the drpai_runtime shim's TVM backend) and check parity against the
+# known-good ONNX output. Run INSIDE the drpai-tvm-v2h container; needs the
+# x86 deploy dir + the pre-saved input/onnx-reference .npy files.
+import sys
+import numpy as np
+
+sys.path.insert(0, "/work/rzv2h/emulation") # drpai_runtime shim (TVM backend)
+sys.path.insert(0, "/work/plugins/python") # utils.detection_decoder (pure numpy)
+
+import drpai_runtime
+from utils.detection_decoder import decode
+
+DEPLOY = "/work/rzv2h/yolo11m_x86_cpu"
+x = np.load("/work/rzv2h/_x86test_input.npy").astype(np.float32)
+ref = np.load("/work/rzv2h/_x86test_onnxout.npy").astype(np.float32).reshape(-1)
+
+rt = drpai_runtime.Runtime()
+assert rt.load(DEPLOY), "drpai_runtime.load failed"
+rt.set_input(0, x)
+rt.run()
+out = np.asarray(rt.get_output(0), dtype=np.float32).reshape(-1)
+
+n = min(out.size, ref.size)
+maxdiff = float(np.max(np.abs(out[:n] - ref[:n]))) if n else float("nan")
+print(f"TVM out size={out.size} ref size={ref.size} max|TVM-ONNX|={maxdiff:.3e}")
+
+tvm_det = decode(out.reshape(1, 84, 8400), "anchor_free")[0]
+onnx_det = decode(ref.reshape(1, 84, 8400), "anchor_free")[0]
+print(f"detections TVM={len(tvm_det['boxes'])} ONNX={len(onnx_det['boxes'])}")
+if len(tvm_det["boxes"]):
+ print("TVM labels:", sorted(set(int(c) for c in tvm_det["labels"])))
+print("PASS" if maxdiff < 1e-2 and len(tvm_det["boxes"]) == len(onnx_det["boxes"]) else "CHECK")
diff --git a/extern/rzv2h/yocto/README.md b/extern/rzv2h/yocto/README.md
new file mode 100644
index 0000000..ccfbdb4
--- /dev/null
+++ b/extern/rzv2h/yocto/README.md
@@ -0,0 +1 @@
+# Custom RZ/V2H image for the gst-python-ml pipeline
diff --git a/extern/rzv2h/yocto/meta-gst-python-ml/conf/include/gstreamer-1.24.inc b/extern/rzv2h/yocto/meta-gst-python-ml/conf/include/gstreamer-1.24.inc
new file mode 100644
index 0000000..cfa0a3d
--- /dev/null
+++ b/extern/rzv2h/yocto/meta-gst-python-ml/conf/include/gstreamer-1.24.inc
@@ -0,0 +1,25 @@
+# Pin GStreamer to 1.24 across the stack so GstAnalytics is available.
+#
+# scarthgap's oe-core ships GStreamer 1.22.x. The 1.24 recipes must be present
+# in the build (see README: copy the gstreamer1.0* recipes from oe-core
+# styhead/master into recipes-multimedia/gstreamer/ of this layer, or layer in
+# a newer meta-oe). These PREFERRED_VERSION lines then select them.
+#
+# require this from local.conf:
+# require ${TOPDIR}/../layers/meta-gst-python-ml/conf/include/gstreamer-1.24.inc
+
+GST_124 ?= "1.24.%"
+
+PREFERRED_VERSION_gstreamer1.0 = "${GST_124}"
+PREFERRED_VERSION_gstreamer1.0-plugins-base = "${GST_124}"
+PREFERRED_VERSION_gstreamer1.0-plugins-good = "${GST_124}"
+PREFERRED_VERSION_gstreamer1.0-plugins-bad = "${GST_124}"
+PREFERRED_VERSION_gstreamer1.0-plugins-ugly = "${GST_124}"
+PREFERRED_VERSION_gstreamer1.0-libav = "${GST_124}"
+PREFERRED_VERSION_gstreamer1.0-python = "${GST_124}"
+PREFERRED_VERSION_gstreamer1.0-rtsp-server = "${GST_124}"
+PREFERRED_VERSION_gstreamer1.0-vaapi = "${GST_124}"
+
+# GstAnalytics + the object-detection / tracking metas live in -plugins-bad.
+# Make sure analytics isn't disabled by a PACKAGECONFIG override.
+PACKAGECONFIG:append:pn-gstreamer1.0-plugins-bad = " analytics"
diff --git a/extern/rzv2h/yocto/meta-gst-python-ml/conf/layer.conf b/extern/rzv2h/yocto/meta-gst-python-ml/conf/layer.conf
new file mode 100644
index 0000000..262f68a
--- /dev/null
+++ b/extern/rzv2h/yocto/meta-gst-python-ml/conf/layer.conf
@@ -0,0 +1,15 @@
+# meta-gst-python-ml — adds the runtime stack gst-python-ml needs on RZ/V2H.
+#
+# The RZ/V2H AI SDK v6.00 image is Yocto scarthgap (5.0.11) with GStreamer
+# 1.22.x. gst-python-ml requires GStreamer >= 1.24 (for GstAnalytics, the
+# metadata type every pyml_* element uses), the gst-python plugin loader, and
+# numpy/pycairo/pygobject/opencv. This layer carries those additions.
+BBPATH .= ":${LAYERDIR}"
+BBFILES += "${LAYERDIR}/recipes-*/*/*.bb ${LAYERDIR}/recipes-*/*/*.bbappend"
+
+BBFILE_COLLECTIONS += "gst-python-ml"
+BBFILE_PATTERN_gst-python-ml = "^${LAYERDIR}/"
+BBFILE_PRIORITY_gst-python-ml = "20"
+
+LAYERDEPENDS_gst-python-ml = "core openembedded-layer"
+LAYERSERIES_COMPAT_gst-python-ml = "scarthgap styhead"
diff --git a/extern/rzv2h/yocto/meta-gst-python-ml/recipes-core/packagegroups/packagegroup-gst-python-ml.bb b/extern/rzv2h/yocto/meta-gst-python-ml/recipes-core/packagegroups/packagegroup-gst-python-ml.bb
new file mode 100644
index 0000000..4a9cef9
--- /dev/null
+++ b/extern/rzv2h/yocto/meta-gst-python-ml/recipes-core/packagegroups/packagegroup-gst-python-ml.bb
@@ -0,0 +1,26 @@
+SUMMARY = "Runtime stack for gst-python-ml on RZ/V2H (GStreamer 1.24 + Python)"
+LICENSE = "MIT"
+
+inherit packagegroup
+
+RDEPENDS:${PN} = " \
+ gstreamer1.0 \
+ gstreamer1.0-plugins-base \
+ gstreamer1.0-plugins-good \
+ gstreamer1.0-plugins-bad \
+ gstreamer1.0-libav \
+ gstreamer1.0-python \
+ python3-core \
+ python3-pygobject \
+ python3-numpy \
+ python3-pycairo \
+ python3-opencv \
+"
+# Notes:
+# - gstreamer1.0-python provides the libgstpython.so plugin loader that runs
+# the pyml_* .py elements. It is NOT in the stock AI SDK image.
+# - GstAnalytics (used by base_objectdetector / tracker / overlay) ships in
+# gstreamer1.0-plugins-bad once GStreamer is >= 1.24 with the analytics
+# PACKAGECONFIG enabled (see conf/include/gstreamer-1.24.inc).
+# - The DRP-AI MERA/TVM *Python* runtime is not a stock Yocto package; install
+# it onto the image separately (see ../README.md "DRP-AI runtime on board").
diff --git a/extern/rzv2h/yocto/meta-gst-python-ml/recipes-multimedia/gst-python-ml/gst-python-ml_git.bb b/extern/rzv2h/yocto/meta-gst-python-ml/recipes-multimedia/gst-python-ml/gst-python-ml_git.bb
new file mode 100644
index 0000000..3b3245b
--- /dev/null
+++ b/extern/rzv2h/yocto/meta-gst-python-ml/recipes-multimedia/gst-python-ml/gst-python-ml_git.bb
@@ -0,0 +1,35 @@
+SUMMARY = "gst-python-ml elements (pyml_*) + DRP-AI engine for RZ/V2H"
+DESCRIPTION = "Installs the pure-Python GStreamer elements and sets GST_PLUGIN_PATH/PYTHONPATH."
+LICENSE = "LGPL-2.1-or-later"
+LIC_FILES_CHKSUM = "file://COPYING;md5="
+
+# Point this at your gst-python-ml source. Examples:
+# SRC_URI = "git://github.com/collabora/gst-python-ml.git;branch=main;protocol=https"
+# SRCREV = ""
+# or a local checkout via: SRC_URI = "file:///path/to/gst-python-ml"
+SRC_URI = "git://github.com/collabora/gst-python-ml.git;branch=master;protocol=https"
+SRCREV = "${AUTOREV}"
+S = "${WORKDIR}/git"
+
+# Pure-Python elements: nothing to compile.
+do_compile[noexec] = "1"
+
+PYML_DIR = "${datadir}/gst-python-ml"
+
+do_install() {
+ install -d ${D}${PYML_DIR}
+ cp -r ${S}/plugins ${D}${PYML_DIR}/plugins
+
+ # Environment so GStreamer finds the .py elements and Python finds the pkg.
+ install -d ${D}${sysconfdir}/profile.d
+ cat > ${D}${sysconfdir}/profile.d/gst-python-ml.sh <",
)
- src_template = Gst.PadTemplate.new(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- VIDEO_SRC_CAPS.copy(),
- )
-
- sink_template = Gst.PadTemplate.new(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.ALWAYS,
- VIDEO_SINK_CAPS.copy(),
- )
- __gsttemplates__ = (src_template, sink_template)
+ if backend.BACKEND == "gst":
+ src_template = Gst.PadTemplate.new(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ VIDEO_SRC_CAPS.copy(),
+ )
+
+ sink_template = Gst.PadTemplate.new(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.ALWAYS,
+ VIDEO_SINK_CAPS.copy(),
+ )
+ __gsttemplates__ = (src_template, sink_template)
rules = GObject.Property(
type=str,
@@ -219,32 +222,11 @@ def do_set_caps(self, incaps, outcaps):
return True
def _read_detections(self, buf):
- """Extract detections from upstream GstAnalytics od_mtd."""
- detections = []
- meta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
+ """Extract detections from upstream object-detection metadata."""
+ meta = analytics.get_relation_meta(buf)
if not meta:
- return detections
-
- count = GstAnalytics.relation_get_length(meta)
- for index in range(count):
- ret, od_mtd = meta.get_od_mtd(index)
- if not ret or od_mtd is None:
- continue
- label_quark = od_mtd.get_obj_type()
- label = GLib.quark_to_string(label_quark)
- presence, x, y, w, h, score = od_mtd.get_location()
- if presence:
- detections.append(
- {
- "label": label,
- "score": score,
- "x": x,
- "y": y,
- "w": w,
- "h": h,
- }
- )
- return detections
+ return []
+ return analytics.read_objects(meta)
def _check_rule(self, rule, detection):
"""Check if a detection matches an alert rule."""
@@ -393,10 +375,9 @@ def do_stop(self):
return True
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(AlertTransform)
- __gstelementfactory__ = ("pyml_alert", Gst.Rank.NONE, AlertTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_alert", AlertTransform)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_alert' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/anomaly.py b/plugins/python/anomaly.py
index 50586e7..c80f67e 100644
--- a/plugins/python/anomaly.py
+++ b/plugins/python/anomaly.py
@@ -17,24 +17,16 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import ctypes
- import json
-
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- from gi.repository import Gst, GObject
-
from video_transform import VideoTransform
from utils.format_converter import FormatConverter
- from utils.muxed_buffer_processor import MuxedBufferProcessor
- from engine.pytorch_engine import PyTorchEngine
+ from engine.anomaly_engine import AnomalyEngine
from engine.engine_factory import EngineFactory
+ from backend import frameio, GObject
+ from tasks.anomaly import AnomalyTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -44,130 +36,7 @@
ANOMALY_META_HEADER = b"GST-ANOMALY:"
-class AnomalyEngine(PyTorchEngine):
- """
- PyTorch engine for anomaly detection using a PatchCore-like approach.
-
- Uses a pretrained feature extractor (WideResNet50 or ResNet) to extract
- patch-level features and compare them against a reference distribution.
-
- Supports torchvision backbone models:
- wide_resnet50_2
- resnet50
- resnet18
- """
-
- def do_load_model(self, model_name, **kwargs):
- try:
- import torch
- import torchvision.models as models
-
- model_fn = getattr(models, model_name, None)
- if model_fn is None:
- raise ValueError(f"Unknown backbone model: {model_name}")
-
- self.backbone = model_fn(weights="DEFAULT")
- # Remove the final FC layer to get feature maps
- self.feature_layers = torch.nn.Sequential(
- *list(self.backbone.children())[:-2]
- )
- self.execute_with_stream(lambda: self.feature_layers.to(self.device))
- self.feature_layers.eval()
-
- self.reference_features = None
- self._transform = None
- self.logger.info(f"Anomaly backbone '{model_name}' loaded on {self.device}")
- except Exception as e:
- raise ValueError(f"Failed to load anomaly backbone '{model_name}': {e}")
-
- def load_reference(self, reference_path):
- """Load precomputed reference features from a .npy file."""
- import numpy as np
-
- try:
- self.reference_features = np.load(reference_path)
- self.logger.info(
- f"Loaded reference features from '{reference_path}': "
- f"shape={self.reference_features.shape}"
- )
- except Exception as e:
- self.logger.warning(f"Failed to load reference features: {e}")
-
- def _get_transform(self):
- if self._transform is None:
- from torchvision import transforms
-
- self._transform = transforms.Compose(
- [
- transforms.ToPILImage(),
- transforms.Resize((224, 224)),
- transforms.ToTensor(),
- transforms.Normalize(
- mean=[0.485, 0.456, 0.406],
- std=[0.229, 0.224, 0.225],
- ),
- ]
- )
- return self._transform
-
- def do_forward(self, frames, threshold=0.5):
- import numpy as np
- import torch
-
- is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
- if not is_batch:
- frames = frames[np.newaxis]
-
- transform = self._get_transform()
- results = []
- for frame in frames:
- try:
- tensor = transform(frame.astype(np.uint8)).unsqueeze(0).to(self.device)
-
- with torch.no_grad():
- features = self.feature_layers(tensor)
-
- # Global average pool to get a feature vector
- feat_vec = features.mean(dim=[2, 3]).squeeze(0).cpu().numpy()
-
- # Compute anomaly score against reference distribution
- anomaly_score = 0.0
- if self.reference_features is not None:
- distances = np.linalg.norm(
- self.reference_features - feat_vec, axis=-1
- )
- anomaly_score = float(distances.min())
-
- # Generate a spatial anomaly heatmap from feature map distances
- feat_map = features.squeeze(0).cpu().numpy()
- heatmap = np.linalg.norm(feat_map, axis=0)
- heatmap = (heatmap - heatmap.min()) / (
- heatmap.max() - heatmap.min() + 1e-8
- )
-
- is_anomaly = anomaly_score >= threshold
-
- results.append(
- {
- "score": anomaly_score,
- "is_anomaly": is_anomaly,
- "heatmap": heatmap,
- }
- )
- except Exception as e:
- self.logger.error(f"Anomaly inference error on frame: {e}")
- results.append(
- {
- "score": 0.0,
- "is_anomaly": False,
- "heatmap": None,
- }
- )
-
- return results[0] if not is_batch else results
-
-
-class AnomalyTransform(VideoTransform):
+class AnomalyTransform(VideoTransform, AnomalyTask):
"""
GStreamer element for anomaly detection in video frames.
@@ -181,6 +50,8 @@ class AnomalyTransform(VideoTransform):
Anomaly scores are always attached as a GST-ANOMALY: memory chunk (JSON).
"""
+ META_HEADER = ANOMALY_META_HEADER
+
__gstmetadata__ = (
"Anomaly Detection",
"Transform",
@@ -230,120 +101,26 @@ def engine_name(self):
def engine_name(self, value):
raise ValueError("'engine_name' is read-only for pyml_anomaly")
- def do_transform_ip(self, buf):
- try:
- # Load reference features on first transform if path is set
- if not self._reference_loaded and self.reference_path and self.engine:
- self.engine.load_reference(self.reference_path)
- self._reference_loaded = True
-
- processor = MuxedBufferProcessor(
- self.logger, self.width, self.height, 30, 1
- )
- frames, _, num_sources, fmt = processor.extract_frames(buf, self.sinkpad)
- if frames is None:
- return Gst.FlowReturn.ERROR
-
- frame = frames[0] if frames.ndim == 4 else frames
- result = self._do_forward(frame)
- if result is None:
- return Gst.FlowReturn.OK
-
- self._apply_anomaly(buf, result, fmt, frame)
- return Gst.FlowReturn.OK
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Score the primary frame against the reference features."""
+ if not self._reference_loaded and self.reference_path and self.engine:
+ self.engine.load_reference(self.reference_path)
+ self._reference_loaded = True
- except Exception as e:
- self.logger.error(f"Anomaly detection transform error: {e}")
- return Gst.FlowReturn.ERROR
+ frame = frames[0] if frames.ndim == 4 else frames
+ result = self.forward(frame)
+ if result is None:
+ return
- def _do_forward(self, frame):
- if self.engine:
- return self.engine.do_forward(frame, threshold=self.threshold)
- return None
+ output, blob = self.decode(frame, result, fmt)
+ frameio.write_result(target, output, blob, self.META_HEADER)
- def _apply_anomaly(self, buf, result, fmt, frame):
- """Overlay heatmap on frame and append anomaly metadata."""
- import cv2
- import numpy as np
- is_anomaly = result.get("is_anomaly", False)
- heatmap = result.get("heatmap")
-
- # Draw heatmap overlay before appending read-only metadata memory
- if self.draw_heatmap and is_anomaly and heatmap is not None:
- H, W = frame.shape[:2]
- heatmap_resized = cv2.resize(heatmap, (W, H))
- heatmap_uint8 = (heatmap_resized * 255).astype(np.uint8)
- heatmap_color = cv2.applyColorMap(heatmap_uint8, cv2.COLORMAP_JET)
- heatmap_rgb = cv2.cvtColor(heatmap_color, cv2.COLOR_BGR2RGB)
-
- overlay = cv2.addWeighted(frame, 0.6, heatmap_rgb, 0.4, 0)
-
- # Draw anomaly score text
- score = result.get("score", 0.0)
- text = f"ANOMALY: {score:.3f}"
- cv2.putText(
- overlay,
- text,
- (12, 36),
- cv2.FONT_HERSHEY_SIMPLEX,
- 1.0,
- (255, 0, 0),
- 2,
- cv2.LINE_AA,
- )
-
- output = self._convert_rgb_to_format(overlay, fmt)
- if output is not None:
- success, map_info = buf.map(Gst.MapFlags.WRITE)
- if success:
- try:
- frame_bytes = np.ascontiguousarray(output).tobytes()
- dst = (ctypes.c_char * map_info.size).from_buffer(map_info.data)
- ctypes.memmove(
- dst, frame_bytes, min(len(frame_bytes), map_info.size)
- )
- finally:
- buf.unmap(map_info)
-
- # Append anomaly metadata (without the numpy heatmap)
- meta = {
- "score": result.get("score", 0.0),
- "is_anomaly": is_anomaly,
- }
- meta_bytes = ANOMALY_META_HEADER + json.dumps(meta).encode("utf-8")
- tmp = Gst.Buffer.new_allocate(None, len(meta_bytes), None)
- tmp.fill(0, meta_bytes)
- buf.append_memory(tmp.get_memory(0))
-
- @staticmethod
- def _convert_rgb_to_format(rgb, fmt):
- """Convert an RGB numpy array to the target GStreamer video format."""
- import cv2
- import numpy as np
-
- if fmt == "RGB":
- return rgb
- elif fmt == "BGR":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
- elif fmt == "RGBA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- elif fmt == "BGRA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- elif fmt == "ARGB":
- rgba = cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- return np.roll(rgba, 1, axis=-1)
- elif fmt == "ABGR":
- bgra = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- return np.roll(bgra, 1, axis=-1)
- else:
- return rgb
-
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(AnomalyTransform)
- __gstelementfactory__ = ("pyml_anomaly", Gst.Rank.NONE, AnomalyTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_anomaly", AnomalyTransform
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_anomaly' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/backend/__init__.py b/plugins/python/backend/__init__.py
new file mode 100644
index 0000000..2ea5f1d
--- /dev/null
+++ b/plugins/python/backend/__init__.py
@@ -0,0 +1,82 @@
+# Pluggable element backend
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""Pluggable multimedia-framework backend for the ML elements.
+
+The ML logic (engines, tasks, base classes) is framework-agnostic. The pieces
+that DO depend on the framework are isolated behind this package:
+
+ * the element base classes `BaseTransform` / `BaseAggregator` / `VideoTransform`
+ * the `analytics` metadata interface (add/get/remove detections)
+
+Select the active backend with the `PYML_BACKEND` environment variable
+(default ``"gst"``). To add another backend, create a sibling package exposing
+the same names (reusing `backend.core.MLEngineMixin` and implementing
+`backend.analytics.AnalyticsBackend`) and add a branch below.
+"""
+
+import os
+
+from backend.analytics import AnalyticsBackend # noqa: F401 (re-exported)
+from backend.frameio import FrameIO # noqa: F401 (re-exported)
+from backend.core import MLEngineMixin # noqa: F401 (re-exported)
+
+BACKEND = os.environ.get("PYML_BACKEND", "gst").lower()
+
+if BACKEND == "gst":
+ from backend.gst import (
+ BaseTransform,
+ BaseAggregator,
+ VideoTransform,
+ analytics,
+ frameio,
+ FlowReturn,
+ GObject,
+ )
+elif BACKEND == "g2g":
+ from backend.g2g import (
+ BaseTransform,
+ BaseAggregator,
+ VideoTransform,
+ analytics,
+ frameio,
+ FlowReturn,
+ GObject,
+ )
+else:
+ raise ImportError(
+ f"Unknown PYML_BACKEND={BACKEND!r}; supported backends: 'gst', 'g2g'"
+ )
+
+
+def register_gst_element(name, cls, type_name=None):
+ """Register `cls` with the GObject type system and return its factory tuple.
+
+ Leaf elements call this behind a `BACKEND == "gst"` guard: under g2g the
+ element base classes are plain Python objects, so type_register would fail.
+ """
+ from backend.gst import Gst, GObject
+
+ GObject.type_register(cls, type_name)
+ return (name, Gst.Rank.NONE, cls)
+
+
+__all__ = [
+ "BACKEND",
+ "register_gst_element",
+ "BaseTransform",
+ "BaseAggregator",
+ "VideoTransform",
+ "analytics",
+ "frameio",
+ "FlowReturn",
+ "GObject",
+ "AnalyticsBackend",
+ "FrameIO",
+ "MLEngineMixin",
+]
diff --git a/plugins/python/backend/analytics.py b/plugins/python/backend/analytics.py
new file mode 100644
index 0000000..9d624b7
--- /dev/null
+++ b/plugins/python/backend/analytics.py
@@ -0,0 +1,85 @@
+# AnalyticsBackend
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Framework-agnostic interface for analytics metadata.
+
+Detection / classification / tracking results are attached to a buffer as
+analytics metadata. GStreamer expresses this via `GstAnalytics` relation
+metadata. Elements should go through this interface instead of calling the
+framework API directly, so the same element code can run on a different backend
+later.
+
+The `meta` handle and the detection handles returned by `add_*` are opaque to
+callers: pass them back into the same backend (e.g. into `relate`). The read
+side (`read_objects`) returns plain dicts so consumers need no framework types.
+"""
+
+from abc import ABC, abstractmethod
+
+
+class AnalyticsBackend(ABC):
+ """Attach to and read analytics metadata on a buffer/frame."""
+
+ @abstractmethod
+ def add_relation_meta(self, buf):
+ """Attach (or fetch the existing) analytics relation meta. Returns an
+ opaque meta handle, or None on failure."""
+
+ @abstractmethod
+ def get_relation_meta(self, buf):
+ """Return the attached relation meta handle, or None if absent."""
+
+ @abstractmethod
+ def remove_relation_meta(self, buf):
+ """Remove any attached relation meta. Returns True if something was
+ removed."""
+
+ @abstractmethod
+ def relation_length(self, meta):
+ """Number of relations (detections) currently held by `meta`."""
+
+ @abstractmethod
+ def quark(self, label):
+ """Intern a string label into the backend's id space (a GQuark for
+ GStreamer). Accepts a str; ints pass through unchanged."""
+
+ @abstractmethod
+ def add_object(self, meta, label, x, y, w, h, score):
+ """Add an object-detection box (x, y, w, h) with a class `label` and
+ confidence `score`. Returns an opaque detection handle, or None."""
+
+ @abstractmethod
+ def add_classification(self, meta, index, label):
+ """Add a single-class classification result tagged with a stream
+ `index` and a `label`. Returns an opaque handle, or None."""
+
+ @abstractmethod
+ def add_tracking(self, meta, track_id, timestamp=None):
+ """Add a tracking record for `track_id`. `timestamp` defaults to the
+ backend's current running time. Returns an opaque handle, or None."""
+
+ @abstractmethod
+ def relate(self, meta, src, dst):
+ """Relate two handles (e.g. a detection to its tracking record).
+ Returns True on success."""
+
+ @abstractmethod
+ def read_objects(self, meta):
+ """Read back the object detections held by `meta`. Returns a list of
+ dicts with keys: ``label`` (str), ``x``, ``y``, ``w``, ``h``, ``score``.
+ Only present (valid-location) detections are returned."""
diff --git a/plugins/python/backend/core.py b/plugins/python/backend/core.py
new file mode 100644
index 0000000..7e92c4b
--- /dev/null
+++ b/plugins/python/backend/core.py
@@ -0,0 +1,303 @@
+# MLEngineMixin
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Portable, framework-agnostic ML element logic.
+
+`MLEngineMixin` holds everything an ML element needs that does NOT depend on
+the multimedia framework: the engine/model lifecycle and the backing storage
+for the common tunable properties.
+
+A framework backend builds a concrete element by combining this mixin with its
+own element base class (`GstBase.BaseTransform`, `GstBase.Aggregator`, ...) and
+declaring the properties in whatever form the framework requires
+(`@GObject.Property` for GStreamer, etc). The property getters/setters read and
+write the `self._` fields initialised here, so the shared logic stays
+identical across backends.
+
+This module imports no `gi`; keep it that way so the same mixin can load under a
+non-GStreamer backend.
+"""
+
+from engine.engine_manager import EngineManager
+from log.logger_factory import LoggerFactory
+
+#: A tunable the element stores and mirrors onto the engine under the same name:
+#: ``(name, type, default, blurb)``.
+ENGINE_TUNABLES = (
+ ("batch_size", int, 1, "Number of items to process in a batch"),
+ ("frame_stride", int, 1, "How often to process a frame"),
+ ("device_queue_id", int, 1, "ID of the DeviceQueue from the pool to use"),
+)
+
+
+def _engine_tunable(gobject, name, value_type, default, blurb):
+ field = f"_{name}"
+
+ def get(self):
+ return getattr(self, field)
+
+ def set(self, value):
+ setattr(self, field, value)
+ if self.engine:
+ setattr(self.engine, name, value)
+
+ # pygobject takes the property's name off the getter
+ get.__name__ = name
+ return gobject.Property(get, type=value_type, default=default, blurb=blurb).setter(
+ set
+ )
+
+
+def ml_property_namespace(gobject):
+ """The tunables every ML element takes, declared with `gobject`'s `Property`.
+
+ Returned as a namespace for an element base to unpack into its class body,
+ not as a mixin to inherit: pygobject installs a property only when it sits
+ in the class's own dict, so an inherited one leaves a gst element with none
+ of these. Both backends build their bases from this call, so a knob cannot
+ exist on one and not the other.
+ """
+
+ @gobject.Property(type=str)
+ def device(self):
+ "Device to run the inference on (cpu, cuda, cuda:0, cuda:1, etc.)"
+ return self.mgr.device
+
+ @device.setter
+ def device(self, value):
+ self.mgr.set_device(value)
+ # TODO: why is this needed, for example for yolo?
+ if self.mgr.engine_name:
+ self.initialize_engine()
+
+ @gobject.Property(type=str)
+ def model_name(self):
+ "Name of the pre-trained model or local model path"
+ return self._model_name
+
+ @model_name.setter
+ def model_name(self, value):
+ self._model_name = value
+
+ @gobject.Property(type=str)
+ def engine_name(self):
+ "Machine Learning Engine to use : pytorch, tflite, tensorflow, onnx, openvino, tvm, tinygrad, mlx, executorch, llamacpp, candle, jax, or custom engine name"
+ return self.mgr.engine_name
+
+ @engine_name.setter
+ def engine_name(self, value):
+ self.mgr.engine_name = value
+
+ @gobject.Property(type=str, default="auto")
+ def input_format(self):
+ "Input tensor layout: auto, nhwc, or nchw"
+ return self.engine.input_format if self.engine else "auto"
+
+ @input_format.setter
+ def input_format(self, value):
+ if self.engine:
+ self.engine.input_format = value
+
+ @gobject.Property(type=str, default="auto")
+ def post_process(self):
+ "Post-processing format for raw engine output (auto, none, or a key from detection_decoder)"
+ return self.engine.post_process if self.engine else "auto"
+
+ @post_process.setter
+ def post_process(self, value):
+ if self.engine:
+ self.engine.post_process = value
+
+ @gobject.Property(type=bool, default=False)
+ def compile(self):
+ "Enable torch.compile optimization for the model"
+ return self._compile
+
+ @compile.setter
+ def compile(self, value):
+ self._compile = value
+ if value:
+ self.kwargs["compile"] = True
+ else:
+ self.kwargs.pop("compile", None)
+
+ namespace = {
+ "device": device,
+ "model_name": model_name,
+ "engine_name": engine_name,
+ "input_format": input_format,
+ "post_process": post_process,
+ "compile": compile,
+ }
+ for name, value_type, default, blurb in ENGINE_TUNABLES:
+ namespace[name] = _engine_tunable(gobject, name, value_type, default, blurb)
+ return namespace
+
+
+class FrameProcessingMixin:
+ """The per-frame work a video element does, shared by every backend.
+
+ The backend driver extracts the frame and handles the framework's error
+ signalling; what is left is the same on all of them, so it lives here once:
+ infer, turn the result into a frame and/or a metadata blob, write both back.
+ An element whose task class follows that contract supplies only its
+ `META_HEADER`; one that does not overrides `process_frames`.
+ """
+
+ #: Tag on the metadata blob this element appends, e.g. ``b"GST-DEPTH:"``.
+ #: `None` for an element that appends none.
+ META_HEADER = None
+
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Infer over the extracted frame(s) and write the result to `target`.
+
+ Raises on a hard failure; the driver maps that to its own error return.
+ """
+ from backend import frameio
+
+ result = self.forward(frames)
+ if result is None:
+ raise RuntimeError(f"{type(self).__name__}: inference returned None")
+ # Inference sees the whole batch, but one buffer carries one frame back:
+ # the overlay and the blob describe the primary source.
+ frame = frames[0] if num_sources > 1 else frames
+ if num_sources > 1 and isinstance(result, list):
+ if not result:
+ return
+ result = result[0]
+ output, blob = self.decode(frame, result, fmt)
+ frameio.write_result(target, output, blob, self.META_HEADER)
+
+
+class PayloadProcessingMixin:
+ """The per-buffer work a non-video element does, shared by every backend.
+
+ These families read one media type and write another (text in / text out,
+ audio in / text out), so a buffer is opaque bytes rather than a frame. The
+ backend driver maps the input buffer and turns whatever comes back into
+ output buffers; the middle, bytes to bytes, is the same everywhere and is
+ what the element supplies.
+ """
+
+ def process_payload(self, payload: bytes) -> list[bytes]:
+ """Turn one input buffer's bytes into the payloads to send onward.
+
+ Zero or more, because these families are not all 1:1: an element that
+ accumulates across buffers returns an empty list until it has something
+ to say, and one that chunks its output returns several at once.
+ """
+ raise NotImplementedError(
+ f"{type(self).__name__} must implement process_payload"
+ )
+
+ def payload_duration_ns(self, payload_size):
+ """How long a payload of this many bytes plays for.
+
+ `None`, the usual answer, means the output covers the same stretch of
+ the stream as the input did, so it keeps the input's timing. An element
+ that generates media of its own length (speech from a text buffer) works
+ it out from the byte count instead.
+ """
+ return None
+
+
+class MLEngineMixin:
+ """Engine/model lifecycle and the start hook, shared by every ML element."""
+
+ def on_start(self):
+ """One-time setup before the first buffer, on any backend.
+
+ For an element with its own background work to start (an inference
+ thread, a warm-up). Framework lifecycle virtuals like `do_start` only
+ exist on GStreamer, so anything a hosted element also needs goes here.
+ """
+
+ def _ensure_started(self):
+ """Run `on_start` once, for a backend with no start virtual of its own."""
+ if not getattr(self, "_started", False):
+ self._started = True
+ self.on_start()
+
+ def _ml_init(self):
+ """Initialise shared ML state. Call this from the element's __init__."""
+ self.logger = LoggerFactory.get(LoggerFactory.LOGGER_TYPE_GST)
+ self.mgr = EngineManager(self.logger)
+ self.kwargs = {}
+ self._batch_size = 1
+ self._frame_stride = 1
+ self._model_name = None
+ self._device_queue_id = 0
+ self._system_prompt = None
+ self._prompt = None
+ self._compile = False
+
+ @property
+ def engine(self):
+ return self.mgr.engine
+
+ # Engine / model lifecycle. None of this touches the framework, so every
+ # backend reuses it verbatim.
+ def initialize_engine(self):
+ if not self.engine and self.mgr.engine_name:
+ self.mgr.initialize_engine()
+ self.engine.batch_size = self._batch_size
+ self.engine.frame_stride = self._frame_stride
+ if self._device_queue_id:
+ self.engine.device_queue_id = self._device_queue_id
+ if not self.engine:
+ self.logger.error(f"Unsupported ML engine: {self.mgr.engine_name}")
+
+ def do_load_model(self):
+ self.initialize_engine()
+ if self.engine is None:
+ self.logger.error(
+ f"Cannot load model {self._model_name}: engine not initialized"
+ )
+ return
+ if self._model_name is None:
+ self.logger.warning("Cannot load model as model name is not set")
+ return
+ self.mgr.do_load_model(self._model_name, **self.kwargs)
+
+ def get_model(self):
+ """Gets the model from the engine."""
+ self.initialize_engine()
+ if self.engine is None:
+ self.logger.error(
+ f"Cannot get model {self._model_name}: engine not initialized"
+ )
+ return None
+ if self.engine:
+ return self.engine.get_model()
+ return None
+
+ def set_model(self, model):
+ """Sets the model in the engine."""
+ self.initialize_engine()
+ if self.engine is None:
+ self.logger.error("Cannot load model: engine not initialized")
+ return False
+ self.engine.model = model
+ self.logger.info("Model set successfully in the engine.")
+
+ def get_tokenizer(self):
+ self.initialize_engine()
+ if self.engine is None:
+ self.logger.error("Cannot get tokenizer: engine not initialized")
+ return None
+ return self.mgr.get_tokenizer()
diff --git a/plugins/python/backend/frameio.py b/plugins/python/backend/frameio.py
new file mode 100644
index 0000000..25c96b6
--- /dev/null
+++ b/plugins/python/backend/frameio.py
@@ -0,0 +1,72 @@
+# FrameIO
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Framework-agnostic interface for frame buffer I/O.
+
+Video transform elements read pixel data out of a buffer as a numpy frame, run
+inference, and either write a frame back or append an opaque metadata blob.
+GStreamer expresses these via buffer mapping (`Gst.Buffer.map`), the muxed
+buffer processor, and `append_memory`. Elements go through this interface
+instead of touching those directly, so the per-buffer task logic that produces
+numpy frames / metadata stays framework-free.
+
+The `target` and `source` arguments are opaque handles (a Gst buffer and the
+sink pad on the gst backend); callers only pass them through.
+"""
+
+from abc import ABC, abstractmethod
+
+
+class FrameIO(ABC):
+ """Read frames from and write frames/metadata back to a buffer/frame."""
+
+ @abstractmethod
+ def read_frames(self, target, source, width, height, framerate=(30, 1)):
+ """Extract frame(s) from `target`. Returns a tuple
+ ``(frames, num_sources, fmt)`` where ``frames`` is a numpy array shaped
+ (H, W, C) for a single source or (N, H, W, C) for a batch (or None on
+ failure), ``num_sources`` is the source count, and ``fmt`` is the pixel
+ format string (e.g. "RGB")."""
+
+ @abstractmethod
+ def read_frame(self, target, source, width, height):
+ """Read a single (H, W, C) RGB frame from `target`, or None on failure.
+ Use this for plain (non-muxed) video buffers; use `read_frames` when the
+ buffer may carry batched/muxed sources."""
+
+ @abstractmethod
+ def write_frame(self, target, frame):
+ """Write an (H, W, C) uint8 numpy ``frame`` back into ``target`` in
+ place. Returns True on success."""
+
+ @abstractmethod
+ def append_blob(self, target, header, payload):
+ """Append an opaque metadata blob (``header`` bytes followed by
+ ``payload`` bytes) to ``target``. Returns True on success."""
+
+ def write_result(self, target, output=None, blob=None, header=None):
+ """Write back what an element produced: a frame, a blob, or both.
+
+ The tail nearly every video element shares. Either half may be absent:
+ an element that only annotates returns no frame, one that only redraws
+ returns no blob.
+ """
+ if output is not None:
+ self.write_frame(target, output)
+ if blob is not None:
+ self.append_blob(target, header, blob)
diff --git a/plugins/python/backend/g2g/__init__.py b/plugins/python/backend/g2g/__init__.py
new file mode 100644
index 0000000..46cde56
--- /dev/null
+++ b/plugins/python/backend/g2g/__init__.py
@@ -0,0 +1,36 @@
+# g2g backend
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""g2g backend: element bases plus the frame I/O and analytics implementations.
+
+The counterpart of `backend.gst`, targeting the glass2glass (`g2g`) host instead
+of GStreamer. Elements are plain Python objects the host drives via
+`g2g_process` / `g2g_process_batch`; there is no GObject type system or GstBase
+element, so `GObject` / `FlowReturn` are lightweight shims (see `shims.py`) that
+let leaf element code load unchanged. Selected by `PYML_BACKEND=g2g`.
+
+Unlike the gst backend, no `gi` is imported anywhere here, so a leaf element can
+be hosted with no GStreamer present at all.
+"""
+
+from backend.g2g.shims import GObject, FlowReturn # noqa: F401
+from backend.g2g.analytics import analytics # noqa: F401
+from backend.g2g.frameio import frameio # noqa: F401
+from backend.g2g.transform import BaseTransform # noqa: F401
+from backend.g2g.video_transform import VideoTransform # noqa: F401
+from backend.g2g.aggregator import BaseAggregator # noqa: F401
+
+__all__ = [
+ "BaseTransform",
+ "BaseAggregator",
+ "VideoTransform",
+ "analytics",
+ "frameio",
+ "FlowReturn",
+ "GObject",
+]
diff --git a/plugins/python/backend/g2g/aggregator.py b/plugins/python/backend/g2g/aggregator.py
new file mode 100644
index 0000000..b63fde5
--- /dev/null
+++ b/plugins/python/backend/g2g/aggregator.py
@@ -0,0 +1,52 @@
+# BaseAggregator (g2g backend)
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""g2g backend for the `aggregator` element family (N inputs -> 1 output).
+
+The host calls `g2g_process_batch([buf, ...], w, h, fmt, sink)` with one
+`FrameBuffer` per input. This base reads each into a frame and hands them to
+`process_frames`, the same per-frame hook a video transform fills in: on gst the
+N-source case arrives muxed into one buffer instead, but what the element is
+given is the same, so it spells the same.
+"""
+
+import numpy as np
+
+from backend.core import FrameProcessingMixin
+from backend.g2g.analytics import analytics
+from backend.g2g.frameio import as_rgb, frameio
+from backend.g2g.transform import BaseTransform
+
+
+class BaseAggregator(BaseTransform, FrameProcessingMixin):
+ """Base for g2g ML aggregator elements (input format may differ from output)."""
+
+ def __init__(self):
+ super().__init__()
+ self.width = 0
+ self.height = 0
+
+ def g2g_process_batch(self, buffers, width, height, fmt, sink):
+ self.width = width
+ self.height = height
+ frameio.bind(sink, fmt)
+ analytics.bind(sink)
+ self._ensure_model()
+ self._ensure_started()
+ frames = []
+ for buf in buffers:
+ frame = frameio.read_frame(buf, None, width, height)
+ if frame is not None:
+ frames.append(frame)
+ if not frames:
+ return None
+ # (H, W, C) for one source and (N, H, W, C) for several, which is what
+ # `read_frames` hands a video transform.
+ batch = frames[0] if len(frames) == 1 else np.stack(frames, axis=0)
+ self.process_frames(as_rgb(batch, fmt), len(frames), fmt, buffers[0])
+ return None
diff --git a/plugins/python/backend/g2g/analytics.py b/plugins/python/backend/g2g/analytics.py
new file mode 100644
index 0000000..2e3605e
--- /dev/null
+++ b/plugins/python/backend/g2g/analytics.py
@@ -0,0 +1,162 @@
+# G2gAnalyticsBackend (g2g backend)
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""Analytics metadata over the g2g host's `MetaSink`.
+
+GStreamer attaches a `GstAnalyticsRelationMeta` to the buffer and relates
+detections, classifications and tracking records. The g2g host instead hands the
+element a flat, write-only `MetaSink` per frame with `add_object(label, x, y, w,
+h, score)` / `add_classification(label, score)` / `add_blob(...)`; the host then
+materializes those into the frame's `AnalyticsMeta`. This backend maps the rich
+`AnalyticsBackend` interface the leaf task code uses onto that flat sink:
+
+ * the "relation meta" handle is a thin wrapper over the bound sink that counts
+ how many records were staged (so `relation_length` works);
+ * string labels are interned to the `u32` ids the sink expects (`quark`);
+ * `add_*` return the sink's own staging handle, which `relate` passes back to
+ pair a detection with its tracking id;
+ * the sink stages straight into the host frame with no read path, so
+ `read_objects` cannot see its own records back.
+"""
+
+import threading
+
+from backend.analytics import AnalyticsBackend
+
+
+class _RelationMeta:
+ """The g2g stand-in for a buffer's relation meta: the bound sink plus a count
+ of staged records (GstAnalytics tracks relations; the flat sink only counts)."""
+
+ def __init__(self, sink):
+ self.sink = sink
+ self.count = 0
+
+
+class _Bound:
+ """What one element has staged: the frame's sink, and its label id space."""
+
+ def __init__(self):
+ self.sink = None
+ self.meta = None
+ self.labels = {} # str -> u32 id
+ self.next_id = 0
+ self.published_names = -1
+
+
+class G2gAnalyticsBackend(AnalyticsBackend):
+ """`AnalyticsBackend` mapping detections/classifications onto a `MetaSink`."""
+
+ def __init__(self):
+ # Per-thread because the host runs one thread per element: one shared
+ # binding would stage an element's detections on another's frame.
+ self._threads = threading.local()
+
+ @property
+ def _bound(self):
+ bound = getattr(self._threads, "bound", None)
+ if bound is None:
+ bound = _Bound()
+ self._threads.bound = bound
+ return bound
+
+ def bind(self, sink):
+ """Bind this frame's sink (called per frame by the g2g element bases).
+ A fresh relation-meta is created lazily on first `add_relation_meta`."""
+ bound = self._bound
+ bound.sink = sink
+ bound.meta = None
+ # Each frame gets its own sink, so the names have to be sent again.
+ bound.published_names = -1
+
+ def quark(self, label):
+ """Intern a string label into the `u32` id space the sink expects; ints
+ pass through unchanged (matches the GStreamer GQuark contract)."""
+ if isinstance(label, int):
+ return label
+ bound = self._bound
+ qid = bound.labels.get(label)
+ if qid is None:
+ qid = bound.next_id
+ bound.labels[label] = qid
+ bound.next_id += 1
+ return qid
+
+ def _publish_class_names(self):
+ """Send the interned label names to the sink, so a consumer can show a
+ name instead of an id. Re-sent when a new label is interned mid-frame."""
+ bound = self._bound
+ if bound.sink is None or bound.next_id == bound.published_names:
+ return
+ names = [""] * bound.next_id
+ for name, qid in bound.labels.items():
+ names[qid] = name
+ bound.sink.set_class_names(names)
+ bound.published_names = bound.next_id
+
+ def add_relation_meta(self, buf):
+ bound = self._bound
+ if bound.sink is None:
+ return None
+ if bound.meta is None:
+ bound.meta = _RelationMeta(bound.sink)
+ return bound.meta
+
+ def get_relation_meta(self, buf):
+ return self._bound.meta
+
+ def remove_relation_meta(self, buf):
+ bound = self._bound
+ had = bound.meta is not None
+ bound.meta = None
+ return had
+
+ def relation_length(self, meta):
+ return meta.count if meta else 0
+
+ def add_object(self, meta, label, x, y, w, h, score):
+ if meta is None:
+ return None
+ meta.count += 1
+ qid = self.quark(label)
+ self._publish_class_names()
+ return meta.sink.add_object(
+ qid, float(x), float(y), float(w), float(h), float(score)
+ )
+
+ def add_classification(self, meta, index, label):
+ if meta is None:
+ return None
+ # The flat sink's add_classification is (label, score); the gst `index`
+ # (stream id) has no place in it and is dropped.
+ meta.count += 1
+ qid = self.quark(label)
+ self._publish_class_names()
+ return meta.sink.add_classification(qid, 1.0)
+
+ def add_tracking(self, meta, track_id, timestamp=None):
+ # The host stamps its own arrival time, so the gst `timestamp` is dropped.
+ if meta is None:
+ return None
+ meta.count += 1
+ return meta.sink.add_tracking(int(track_id))
+
+ def relate(self, meta, src, dst):
+ if meta is None or src is None or dst is None:
+ return False
+ meta.sink.relate(int(src), int(dst))
+ return True
+
+ def read_objects(self, meta):
+ # The sink is write-only (staged straight into the host frame); the
+ # element cannot read its own staged detections back.
+ return []
+
+
+#: Defined here, like `frameio`, to avoid a circular import through `backend`.
+analytics = G2gAnalyticsBackend()
diff --git a/plugins/python/backend/g2g/frameio.py b/plugins/python/backend/g2g/frameio.py
new file mode 100644
index 0000000..495b708
--- /dev/null
+++ b/plugins/python/backend/g2g/frameio.py
@@ -0,0 +1,113 @@
+# G2gFrameIO (g2g backend)
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""Frame buffer I/O backed by the g2g host's `FrameBuffer`.
+
+The g2g host hands each frame to `g2g_process(buf, w, h, fmt, sink)` where `buf`
+is a `FrameBuffer`: a writable buffer-protocol view straight onto the frame's
+system memory (no copy in or out). This backend reads it as a numpy frame, writes
+a processed frame back in place, and routes opaque blobs to the host's `MetaSink`.
+
+Unlike the GStreamer backend there is no muxed/batched buffer: the g2g host
+delivers one source per `FrameBuffer` (batching is the aggregator's job, via
+`g2g_process_batch`), so `read_frames` always reports a single source.
+"""
+
+import threading
+
+import numpy as np
+
+from backend.frameio import FrameIO
+
+#: For each pixel format the host carries: how many channels a pixel takes, and
+#: which of them are R, G and B. `None` for a format with no colour to pick out.
+_FORMATS = {
+ "RGB": (3, (0, 1, 2)),
+ "BGR": (3, (2, 1, 0)),
+ "RGBA": (4, (0, 1, 2)),
+ "ARGB": (4, (1, 2, 3)),
+ "BGRA": (4, (2, 1, 0)),
+ "ABGR": (4, (3, 2, 1)),
+ "GRAY8": (1, None),
+}
+
+
+def as_rgb(frames, fmt):
+ """The RGB view of `frames` the ML elements infer over.
+
+ Nothing converts pixels ahead of a hosted element the way `videoconvert`
+ does in a gst pipeline, so a host format that is not already RGB is reduced
+ here. Read-only: the write-back target is still the original buffer.
+ """
+ channels, rgb = _FORMATS.get((fmt or "RGB").upper(), _FORMATS["RGB"])
+ already_rgb = channels == 3 and rgb == (0, 1, 2)
+ if rgb is None or already_rgb:
+ return frames
+ return frames[..., list(rgb)]
+
+
+class G2gFrameIO(FrameIO):
+ """`FrameIO` over the g2g `FrameBuffer` (pixels) and `MetaSink` (blobs)."""
+
+ def __init__(self):
+ # Per-thread because the host runs one thread per element: one shared
+ # binding would send an element's blobs to whichever element bound last.
+ self._bound = threading.local()
+
+ def bind(self, sink, fmt="RGB"):
+ """Bind the current frame's sink and pixel format (called per frame by
+ the g2g element bases before any read/write)."""
+ self._bound.sink = sink
+ self._bound.fmt = (fmt or "RGB").upper()
+
+ @property
+ def _sink(self):
+ return getattr(self._bound, "sink", None)
+
+ @property
+ def _fmt(self):
+ return getattr(self._bound, "fmt", "RGB")
+
+ def _channels(self, fmt=None):
+ channels, _ = _FORMATS.get((fmt or self._fmt).upper(), _FORMATS["RGB"])
+ return channels
+
+ def read_frame(self, target, source, width, height):
+ c = self._channels()
+ arr = np.frombuffer(target, dtype=np.uint8)
+ if arr.size < width * height * c:
+ return None
+ return arr[: width * height * c].reshape((height, width, c))
+
+ def read_frames(self, target, source, width, height, framerate=(30, 1)):
+ # One source per FrameBuffer; report (frame, num_sources=1, fmt).
+ frame = self.read_frame(target, source, width, height)
+ if frame is None:
+ return None, 0, self._fmt
+ return frame, 1, self._fmt
+
+ def write_frame(self, target, frame):
+ # The FrameBuffer is writable, so frombuffer yields a writable view we
+ # overwrite in place (no copy back to the host).
+ view = np.frombuffer(target, dtype=np.uint8)
+ flat = np.ascontiguousarray(frame, dtype=np.uint8).reshape(-1)
+ n = min(view.size, flat.size)
+ view[:n] = flat[:n]
+ return True
+
+ def append_blob(self, target, header, payload):
+ if self._sink is None:
+ return False
+ hdr = header if isinstance(header, str) else bytes(header).decode("latin-1")
+ self._sink.add_blob(hdr, bytes(payload))
+ return True
+
+
+#: Defined here rather than in the package __init__ so the element bases can
+#: import it without a circular import through `backend`.
+frameio = G2gFrameIO()
diff --git a/plugins/python/backend/g2g/shims.py b/plugins/python/backend/g2g/shims.py
new file mode 100644
index 0000000..f42b819
--- /dev/null
+++ b/plugins/python/backend/g2g/shims.py
@@ -0,0 +1,141 @@
+# Framework-primitive shims (g2g backend)
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""Stand-ins for the GStreamer framework primitives leaf elements reference.
+
+The g2g host drives plain-Python objects, not `GstBase` elements: there is no
+GObject type system and no `Gst.FlowReturn`. But leaf element code declares
+`@GObject.Property(...)` for its tunables and returns `FlowReturn.OK/ERROR` from
+its per-frame method. These shims let that same leaf code load and run unchanged
+under the g2g backend (`from backend import GObject, FlowReturn`).
+
+`GObject.Property` becomes a plain descriptor that stores values on the instance
+(or delegates to a getter/setter pair); gst-only metadata (nick, blurb, min/max)
+is accepted and ignored. `FlowReturn` is a plain enum whose `OK`/`ERROR` the host
+treats as success/failure of a frame.
+"""
+
+from enum import IntEnum, IntFlag
+
+#: How a pipeline line spells a bool, matching what the g2g host accepts.
+BOOL_SPELLINGS = {
+ "true": True,
+ "1": True,
+ "yes": True,
+ "false": False,
+ "0": False,
+ "no": False,
+}
+
+
+class ParamFlags(IntFlag):
+ """The `GObject.ParamFlags` leaves pass to `Property`, accepted and ignored."""
+
+ READABLE = 1
+ WRITABLE = 2
+ READWRITE = READABLE | WRITABLE
+
+
+class FlowReturn(IntEnum):
+ """The subset of `Gst.FlowReturn` leaf per-frame methods return."""
+
+ OK = 0
+ ERROR = -5
+ NOT_LINKED = -1
+ EOS = -3
+
+
+class Property:
+ """Descriptor mimicking `GObject.Property`, in both forms leaves use.
+
+ Decorator form (getter + optional setter)::
+
+ @GObject.Property(type=str)
+ def model_name(self): ...
+ @model_name.setter
+ def model_name(self, value): ...
+
+ Attribute form (plain storage)::
+
+ broker = GObject.Property(type=str, default=None, nick="...", blurb="...")
+
+ Values back a per-instance ``_g2gprop_`` attribute. The ``nick`` /
+ ``blurb`` / ``minimum`` / ``maximum`` keywords are accepted for source
+ compatibility and otherwise ignored (no GObject type system here), but
+ ``type`` is kept: the host cannot know it, so a value off a pipeline line
+ arrives as text and is converted here.
+ """
+
+ def __init__(self, fget=None, *, type=None, default=None, **_gst_meta):
+ self._fget = fget
+ self._fset = None
+ self._type = type
+ self._default = default
+ self._name = None
+
+ # `GObject.Property(type=str)` returns an instance that is then applied as a
+ # decorator to the getter; this captures it.
+ def __call__(self, fget):
+ self._fget = fget
+ return self
+
+ def setter(self, fset):
+ self._fset = fset
+ return self
+
+ def __set_name__(self, owner, name):
+ self._name = name
+
+ def _slot(self):
+ return f"_g2gprop_{self._name}"
+
+ def __get__(self, obj, owner=None):
+ if obj is None:
+ return self
+ if self._fget is not None:
+ return self._fget(obj)
+ return getattr(obj, self._slot(), self._default)
+
+ def __set__(self, obj, value):
+ value = self._converted(value)
+ if self._fset is not None:
+ self._fset(obj, value)
+ else:
+ setattr(obj, self._slot(), value)
+
+ def _converted(self, value):
+ """The value as the declared type, for one that arrived as text.
+
+ The host forwards a pipeline line's `key=value` verbatim, since only this
+ class knows what type the property is. Anything already of the right type
+ (a value set from Python) is left alone.
+ """
+ if self._type in (None, str) or not isinstance(value, str):
+ return value
+ if self._type is bool:
+ spelled = BOOL_SPELLINGS.get(value.strip().lower())
+ if spelled is None:
+ raise ValueError(f"{self._name}: {value!r} is not true or false")
+ return spelled
+ return self._type(value)
+
+
+class _GObjectShim:
+ """The ``GObject`` namespace leaves import from the backend."""
+
+ Property = Property
+ ParamFlags = ParamFlags
+
+ # Some leaves type-hint setters as `prop: GObject.GParamSpec`.
+ class GParamSpec: # noqa: N801
+ pass
+
+ ParamSpec = GParamSpec
+
+
+GObject = _GObjectShim()
diff --git a/plugins/python/backend/g2g/transform.py b/plugins/python/backend/g2g/transform.py
new file mode 100644
index 0000000..333a91d
--- /dev/null
+++ b/plugins/python/backend/g2g/transform.py
@@ -0,0 +1,72 @@
+# BaseTransform (g2g backend)
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""g2g backend for the `transform` element family.
+
+The gst counterpart subclasses `GstBase.BaseTransform`; the g2g element is a
+plain Python object driven by the host's `g2g_process(...)` call. All engine /
+model logic still comes from the portable `MLEngineMixin`, and the shared
+tunables come from `ml_property_namespace` declared with the `GObject` shim, so
+the same set exists here as on gst.
+
+The model is loaded lazily on the first frame (`_ensure_model`) rather than from
+a `do_start` framework virtual, since the g2g host has no start hook.
+
+The payload driver (`g2g_process_payload`, for a stream that is not raw video)
+also sits here rather than in the aggregator, so that the aggregator, which
+subclasses this, gets it too: a single-chain text or audio element is hosted
+1-in-1-out even though its gst counterpart is a `GstBase.Aggregator`.
+"""
+
+from backend.core import MLEngineMixin, PayloadProcessingMixin, ml_property_namespace
+from backend.g2g.shims import GObject
+
+
+class BaseTransform(MLEngineMixin, PayloadProcessingMixin):
+ """Base for g2g ML transform elements (same in/out format, e.g. detection)."""
+
+ locals().update(ml_property_namespace(GObject))
+
+ def __init__(self):
+ self._ml_init()
+
+ def g2g_properties(self):
+ """Every property this element declares, for the host to check a pipeline
+ line against before it runs.
+
+ Without this the host has no way to tell a knob this element has from a
+ typo, and would set the typo as an attribute nothing ever reads.
+ """
+ return sorted(
+ {
+ name
+ for klass in type(self).__mro__
+ for name, value in vars(klass).items()
+ if isinstance(value, GObject.Property)
+ }
+ )
+
+ def _ensure_model(self):
+ """Load the model on first use (the g2g host has no start hook).
+
+ Guard on the model, not the engine: setting the `device` property
+ eagerly creates the engine (with no model loaded), so an engine-only
+ check would skip the load and leave inference with a null model.
+ """
+ if self.mgr.engine_name and (self.engine is None or self.engine.model is None):
+ self.do_load_model()
+
+ def g2g_process_payload(self, buffers, caps, meta):
+ """Host driver for a stream that is not raw video: run the element's
+ `process_payload` over the input bytes and emit what it returns."""
+ self._ensure_model()
+ self._ensure_started()
+ # copied, not viewed: the host takes its buffer back when this returns,
+ # and an element that accumulates keeps what it was given
+ for payload in self.process_payload(bytes(memoryview(buffers[0]))):
+ meta.emit(payload, duration_ns=self.payload_duration_ns(len(payload)))
diff --git a/plugins/python/backend/g2g/video_transform.py b/plugins/python/backend/g2g/video_transform.py
new file mode 100644
index 0000000..1a9f360
--- /dev/null
+++ b/plugins/python/backend/g2g/video_transform.py
@@ -0,0 +1,47 @@
+# VideoTransform (g2g backend)
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""g2g backend for video transform elements.
+
+The host calls `g2g_process(buf, w, h, fmt, sink)` once per frame. This base
+binds the frame's sink onto the shared `frameio` / `analytics` (so leaf task code
+that calls them through `from backend import frameio, analytics` reaches this
+frame's buffer and sink), extracts the frame, then hands off to `process_frames`,
+the framework-agnostic per-frame hook the leaf element supplies (the same hook the
+gst backend drives from `do_transform_ip`).
+"""
+
+from backend.g2g.analytics import analytics
+from backend.g2g.frameio import as_rgb, frameio
+from backend.core import FrameProcessingMixin
+from backend.g2g.transform import BaseTransform
+
+
+class VideoTransform(BaseTransform, FrameProcessingMixin):
+ """Base for g2g video transform elements."""
+
+ def __init__(self):
+ super().__init__()
+ self.width = 0
+ self.height = 0
+
+ def g2g_process(self, buf, width, height, fmt, sink):
+ self.width = width
+ self.height = height
+ # Route this frame's pixels (buf) and metadata sink to the shared I/O the
+ # leaf task code uses.
+ frameio.bind(sink, fmt)
+ analytics.bind(sink)
+ self._ensure_model()
+ self._ensure_started()
+ frames, num_sources, fmt = frameio.read_frames(buf, None, width, height)
+ if frames is None:
+ return None
+ self.process_frames(as_rgb(frames, fmt), num_sources, fmt, buf)
+ # Blobs/detections are staged on the sink; no return payload needed.
+ return None
diff --git a/plugins/python/backend/gst/__init__.py b/plugins/python/backend/gst/__init__.py
new file mode 100644
index 0000000..e565b51
--- /dev/null
+++ b/plugins/python/backend/gst/__init__.py
@@ -0,0 +1,46 @@
+# GStreamer backend
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""GStreamer backend: element bases plus the analytics metadata implementation.
+
+Element registration metadata (`__gstmetadata__`, `__gsttemplates__`, and the
+`__gstelementfactory__` / `GObject.type_register` calls in each leaf plugin) is
+GStreamer-specific and stays in this backend.
+"""
+
+import gi
+
+gi.require_version("Gst", "1.0")
+from gi.repository import Gst, GObject # noqa: E402
+
+from backend.gst.transform import BaseTransform # noqa: E402
+from backend.gst.aggregator import BaseAggregator # noqa: E402
+from backend.gst.video_transform import VideoTransform # noqa: E402
+from backend.gst.analytics import GstAnalyticsBackend # noqa: E402
+from backend.gst.frameio import GstFrameIO # noqa: E402
+
+#: The analytics metadata implementation for this backend.
+analytics = GstAnalyticsBackend()
+
+#: The frame buffer I/O implementation for this backend.
+frameio = GstFrameIO()
+
+# Framework primitives exposed to leaf elements so their task code imports them
+# from the backend rather than touching `gi` directly. `FlowReturn` is the
+# process()/transform return type; `GObject` carries the property declarations.
+FlowReturn = Gst.FlowReturn
+
+__all__ = [
+ "BaseTransform",
+ "BaseAggregator",
+ "VideoTransform",
+ "analytics",
+ "frameio",
+ "FlowReturn",
+ "GObject",
+]
diff --git a/plugins/python/backend/gst/aggregator.py b/plugins/python/backend/gst/aggregator.py
new file mode 100644
index 0000000..02d2a8f
--- /dev/null
+++ b/plugins/python/backend/gst/aggregator.py
@@ -0,0 +1,163 @@
+# BaseAggregator (GStreamer backend)
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""GStreamer backend for the `aggregator` element family (input format differs
+from output format, e.g. audio in / text out).
+
+GStreamer half of the backend split: the element base (`GstBase.Aggregator`)
+and the framework virtuals (`do_change_state`, `do_aggregate`, segment
+handling). Engine/model logic lives in the portable `MLEngineMixin`, and the
+shared tunables come from `ml_property_namespace`.
+"""
+
+import gi
+
+gi.require_version("Gst", "1.0")
+gi.require_version("GstBase", "1.0")
+gi.require_version("GLib", "2.0")
+from gi.repository import Gst, GObject, GstBase # noqa: E402
+
+from backend.core import ( # noqa: E402
+ MLEngineMixin,
+ PayloadProcessingMixin,
+ ml_property_namespace,
+)
+
+
+class PayloadDriver:
+ """The GStreamer half of the payload seam.
+
+ Kept apart from the element base so it can be driven without standing up a
+ `GstBase.Aggregator`, the same way `FrameProcessingMixin` keeps the video
+ work apart from the element that hosts it.
+ """
+
+ #: Send output straight out of the element's own src pad instead of through
+ #: the aggregator. Two families have always done that.
+ PUSH_FROM_SRC_PAD = False
+
+ def do_process(self, buf):
+ """Read the input payload, run the element's `process_payload`, and send
+ each payload it returns as its own buffer. Elements supply
+ `process_payload`, not this."""
+ try:
+ success, map_info = buf.map(Gst.MapFlags.READ)
+ if not success:
+ self.logger.error("Failed to map input buffer")
+ return Gst.FlowReturn.ERROR
+
+ payload = bytes(map_info.data)
+ buf.unmap(map_info)
+
+ for output in self.process_payload(payload):
+ outbuf = Gst.Buffer.new_allocate(None, len(output), None)
+ outbuf.fill(0, output)
+ self.stamp_payload(outbuf, buf)
+ self.push_payload(outbuf)
+
+ return Gst.FlowReturn.OK
+
+ except Exception as e:
+ self.logger.error(f"Error processing buffer: {e}")
+ return Gst.FlowReturn.ERROR
+
+ def stamp_payload(self, outbuf, inbuf):
+ """Time one output buffer, the gst spelling of what `meta.emit` takes.
+
+ An element that generates media of its own length says so through
+ `payload_duration_ns`; that audio runs for as long as it runs and plays
+ wherever the pipeline reaches it, so the input's times say nothing about
+ it. Everything else covers the same stretch of stream as its input.
+ """
+ duration_ns = self.payload_duration_ns(outbuf.get_size())
+ if duration_ns is None:
+ outbuf.pts = inbuf.pts
+ outbuf.dts = inbuf.dts
+ outbuf.duration = inbuf.duration
+ return
+ outbuf.pts = Gst.CLOCK_TIME_NONE
+ outbuf.dts = Gst.CLOCK_TIME_NONE
+ outbuf.duration = duration_ns
+
+ def push_payload(self, outbuf):
+ """Send one output buffer downstream."""
+ if not self.PUSH_FROM_SRC_PAD:
+ self.finish_buffer(outbuf)
+ return
+ ret = self.srcpad.push(outbuf)
+ if ret != Gst.FlowReturn.OK:
+ raise RuntimeError(f"Error pushing payload to pipeline: {ret}")
+
+
+class BaseAggregator(
+ GstBase.Aggregator, MLEngineMixin, PayloadProcessingMixin, PayloadDriver
+):
+ """
+ Base class for GStreamer aggregator elements that perform inference
+ with a machine learning model. This class manages shared properties
+ and handles model loading and device management via MLEngine.
+ """
+
+ __gstmetadata__ = (
+ "BaseAggregator",
+ "Aggregator",
+ "Generic machine learning model aggregator element",
+ "Aaron Boxer ",
+ )
+
+ # unpacked here rather than inherited: pygobject installs a property only
+ # when it sits in the class's own dict
+ locals().update(ml_property_namespace(GObject))
+
+ def __init__(self):
+ super().__init__()
+ self._ml_init()
+ self.segment_pushed = False
+
+ # GStreamer framework virtual: load the model on NULL -> READY.
+ def do_change_state(self, transition):
+ if transition == Gst.StateChange.NULL_TO_READY:
+ self.do_load_model()
+ return Gst.Element.do_change_state(self, transition)
+
+ def push_segment_if_needed(self):
+ if not self.segment_pushed:
+ segment = Gst.Segment()
+ segment.init(Gst.Format.TIME)
+ segment.start = 0
+ segment.stop = Gst.CLOCK_TIME_NONE
+ segment.position = 0
+
+ self.srcpad.push_event(Gst.Event.new_segment(segment))
+ self.segment_pushed = True
+
+ # GStreamer framework virtual: pull buffers from sink pads and process.
+ def do_aggregate(self, timeout):
+ if all(pad.is_eos() for pad in self.sinkpads):
+ return Gst.FlowReturn.EOS
+ self.push_segment_if_needed()
+ self.process_all_sink_pads()
+ self.selected_samples(Gst.CLOCK_TIME_NONE, 0, 0, None)
+ return Gst.FlowReturn.OK
+
+ def process_all_sink_pads(self):
+ if len(self.sinkpads) == 0:
+ return
+ buf = self.sinkpads[0].pop_buffer()
+ if buf:
+ self.do_process(buf)
diff --git a/plugins/python/backend/gst/analytics.py b/plugins/python/backend/gst/analytics.py
new file mode 100644
index 0000000..9b2549d
--- /dev/null
+++ b/plugins/python/backend/gst/analytics.py
@@ -0,0 +1,102 @@
+# GstAnalyticsBackend
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""GStreamer implementation of the analytics metadata interface.
+
+Wraps `GstAnalytics` relation metadata. The opaque `meta` handle is a
+`GstAnalytics.RelationMeta`; the opaque detection handle is the `*_mtd` object
+returned by `add_*_mtd` (it carries the `.id` used by `relate`).
+"""
+
+import gi
+
+gi.require_version("Gst", "1.0")
+gi.require_version("GstAnalytics", "1.0")
+gi.require_version("GLib", "2.0")
+from gi.repository import Gst, GstAnalytics, GLib # noqa: E402
+
+from backend.analytics import AnalyticsBackend # noqa: E402
+
+
+class GstAnalyticsBackend(AnalyticsBackend):
+ def add_relation_meta(self, buf):
+ return GstAnalytics.buffer_add_analytics_relation_meta(buf)
+
+ def get_relation_meta(self, buf):
+ return GstAnalytics.buffer_get_analytics_relation_meta(buf)
+
+ def remove_relation_meta(self, buf):
+ meta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
+ if meta is None:
+ return False
+ # GstAnalytics relation meta is an add-once accumulator: detections are
+ # added into it as *_mtd entries. The Python bindings expose no path to
+ # detach it (the RelationMeta wrapper is not a Gst.Meta, so it cannot be
+ # passed to Gst.Buffer.remove_meta). Attempt it for bindings that do
+ # accept it; otherwise report that nothing was removed.
+ try:
+ return bool(buf.remove_meta(meta))
+ except (TypeError, AttributeError):
+ return False
+
+ def relation_length(self, meta):
+ return GstAnalytics.relation_get_length(meta)
+
+ def quark(self, label):
+ if isinstance(label, int):
+ return label
+ return GLib.quark_from_string(str(label))
+
+ def add_object(self, meta, label, x, y, w, h, score):
+ qk = self.quark(label)
+ ret, mtd = meta.add_od_mtd(qk, x, y, w, h, score)
+ return mtd if ret else None
+
+ def add_classification(self, meta, index, label):
+ qk = self.quark(label)
+ ret, mtd = meta.add_one_cls_mtd(index, qk)
+ return mtd if ret else None
+
+ def add_tracking(self, meta, track_id, timestamp=None):
+ if timestamp is None:
+ timestamp = Gst.util_get_timestamp()
+ ret, mtd = meta.add_tracking_mtd(track_id, timestamp)
+ return mtd if ret else None
+
+ def relate(self, meta, src, dst):
+ return bool(
+ GstAnalytics.RelationMeta.set_relation(
+ meta, GstAnalytics.RelTypes.RELATE_TO, src.id, dst.id
+ )
+ )
+
+ def read_objects(self, meta):
+ objects = []
+ count = GstAnalytics.relation_get_length(meta)
+ for index in range(count):
+ ret, od_mtd = meta.get_od_mtd(index)
+ if not ret or od_mtd is None:
+ continue
+ label = GLib.quark_to_string(od_mtd.get_obj_type())
+ presence, x, y, w, h, score = od_mtd.get_location()
+ if not presence:
+ continue
+ objects.append(
+ {"label": label, "x": x, "y": y, "w": w, "h": h, "score": score}
+ )
+ return objects
diff --git a/plugins/python/backend/gst/frameio.py b/plugins/python/backend/gst/frameio.py
new file mode 100644
index 0000000..4b9a56b
--- /dev/null
+++ b/plugins/python/backend/gst/frameio.py
@@ -0,0 +1,80 @@
+# GstFrameIO
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""GStreamer implementation of the frame buffer I/O interface.
+
+Reads use `MuxedBufferProcessor` (handles single and batched/muxed buffers);
+writes map the buffer for writing and memmove the frame bytes in place; blob
+append wraps the bytes in a new memory chunk on the buffer.
+"""
+
+import gi
+
+gi.require_version("Gst", "1.0")
+from gi.repository import Gst # noqa: E402
+
+from log.logger_factory import LoggerFactory # noqa: E402
+from utils.muxed_buffer_processor import MuxedBufferProcessor # noqa: E402
+from utils.format_converter import FormatConverter # noqa: E402
+from backend.frameio import FrameIO # noqa: E402
+
+
+class GstFrameIO(FrameIO):
+ def __init__(self):
+ self.logger = LoggerFactory.get(LoggerFactory.LOGGER_TYPE_GST)
+ self.format_converter = FormatConverter()
+
+ def read_frames(self, target, source, width, height, framerate=(30, 1)):
+ processor = MuxedBufferProcessor(
+ self.logger, width, height, framerate[0], framerate[1]
+ )
+ frames, _id_str, num_sources, fmt = processor.extract_frames(target, source)
+ return frames, num_sources, fmt
+
+ def read_frame(self, target, source, width, height):
+ success, map_info = target.map(Gst.MapFlags.READ)
+ if not success:
+ return None
+ try:
+ return self.format_converter.to_rgb(
+ map_info.data, width, height, target, source
+ )
+ finally:
+ target.unmap(map_info)
+
+ def write_frame(self, target, frame):
+ import ctypes
+ import numpy as np
+
+ success, map_info = target.map(Gst.MapFlags.WRITE)
+ if not success:
+ return False
+ try:
+ frame_bytes = np.ascontiguousarray(frame).tobytes()
+ dst = (ctypes.c_char * map_info.size).from_buffer(map_info.data)
+ ctypes.memmove(dst, frame_bytes, min(len(frame_bytes), map_info.size))
+ return True
+ finally:
+ target.unmap(map_info)
+
+ def append_blob(self, target, header, payload):
+ blob = bytes(header) + bytes(payload)
+ tmp = Gst.Buffer.new_allocate(None, len(blob), None)
+ tmp.fill(0, blob)
+ target.append_memory(tmp.get_memory(0))
+ return True
diff --git a/plugins/python/backend/gst/transform.py b/plugins/python/backend/gst/transform.py
new file mode 100644
index 0000000..62840ec
--- /dev/null
+++ b/plugins/python/backend/gst/transform.py
@@ -0,0 +1,67 @@
+# BaseTransform (GStreamer backend)
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""GStreamer backend for the `transform` element family (same input/output
+format, e.g. object detection).
+
+This file is the GStreamer half of the backend split: the element base
+(`GstBase.BaseTransform`) and the framework virtuals (`do_start`). All
+engine/model logic lives in the portable `MLEngineMixin`, and the shared
+tunables come from `ml_property_namespace`, so both backends declare the same
+set from one place.
+"""
+
+import gi
+
+gi.require_version("Gst", "1.0")
+gi.require_version("GstBase", "1.0")
+from gi.repository import GObject, GstBase # noqa: E402
+
+from backend.core import MLEngineMixin, ml_property_namespace # noqa: E402
+
+
+class BaseTransform(GstBase.BaseTransform, MLEngineMixin):
+ """
+ Base class for GStreamer transform elements that perform
+ inference with a machine learning model. This class manages shared properties
+ and handles model loading and device management via MLEngine.
+ """
+
+ __gstmetadata__ = (
+ "BaseTransform",
+ "Transform",
+ "Generic machine learning model transform element",
+ "Aaron Boxer ",
+ )
+
+ # unpacked here rather than inherited: pygobject installs a property only
+ # when it sits in the class's own dict
+ locals().update(ml_property_namespace(GObject))
+
+ def __init__(self):
+ super().__init__()
+ self._ml_init()
+
+ # GStreamer framework virtual: load the model when the element starts, then
+ # run whatever the element itself needs starting (the backend-neutral hook).
+ def do_start(self):
+ self.do_load_model()
+ on_start = getattr(self, "on_start", None)
+ if on_start:
+ on_start()
+ return True
diff --git a/plugins/python/backend/gst/video_transform.py b/plugins/python/backend/gst/video_transform.py
new file mode 100644
index 0000000..7185a54
--- /dev/null
+++ b/plugins/python/backend/gst/video_transform.py
@@ -0,0 +1,84 @@
+# VideoTransform (GStreamer backend)
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+import traceback
+
+import gi
+
+gi.require_version("Gst", "1.0")
+gi.require_version("GstBase", "1.0")
+gi.require_version("GstVideo", "1.0")
+from gi.repository import Gst # noqa: E402
+
+from backend.core import FrameProcessingMixin # noqa: E402
+from backend.gst.transform import BaseTransform # noqa: E402
+
+
+class VideoTransform(BaseTransform, FrameProcessingMixin):
+ """
+ GStreamer element for video transformation using a PyTorch model.
+ """
+
+ # Define VIDEO_CAPS to support multiple formats
+ VIDEO_CAPS = Gst.Caps.from_string(
+ "video/x-raw,format=(string){ RGB, RGBA, ARGB, BGRA, ABGR },"
+ "width=(int)[1,2147483647],height=(int)[1,2147483647]"
+ )
+ __gsttemplates__ = (
+ Gst.PadTemplate.new(
+ "src", Gst.PadDirection.SRC, Gst.PadPresence.ALWAYS, VIDEO_CAPS
+ ),
+ Gst.PadTemplate.new(
+ "sink", Gst.PadDirection.SINK, Gst.PadPresence.ALWAYS, VIDEO_CAPS
+ ),
+ )
+
+ def do_set_caps(self, incaps, outcaps):
+ struct = incaps.get_structure(0)
+ self.width = struct.get_int("width").value
+ self.height = struct.get_int("height").value
+
+ return True
+
+ def do_transform_ip(self, buf):
+ """GStreamer per-frame driver: extract the frame(s) through the backend
+ frame I/O, run the element's `process_frames`, and map the outcome to a
+ `Gst.FlowReturn`. Elements supply `process_frames`, not this."""
+ # Imported lazily: the frameio singleton lives in backend.gst, which is
+ # still being constructed when this module is imported.
+ from backend import frameio
+
+ try:
+ frames, num_sources, fmt = frameio.read_frames(
+ buf,
+ self.sinkpad,
+ self.width,
+ self.height,
+ (
+ getattr(self, "framerate_num", 30),
+ getattr(self, "framerate_denom", 1),
+ ),
+ )
+ if frames is None:
+ self.logger.error("Failed to extract frames")
+ return Gst.FlowReturn.ERROR
+ self.process_frames(frames, num_sources, fmt, buf)
+ return Gst.FlowReturn.OK
+ except Exception as e:
+ self.logger.error(f"Transform error: {e}\n{traceback.format_exc()}")
+ return Gst.FlowReturn.ERROR
diff --git a/plugins/python/base_aggregator.py b/plugins/python/base_aggregator.py
index f8bf863..d8f6bd0 100644
--- a/plugins/python/base_aggregator.py
+++ b/plugins/python/base_aggregator.py
@@ -16,208 +16,11 @@
# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
# Boston, MA 02110-1301, USA.
-from abc import abstractmethod
-import gi
+"""Compatibility shim. `BaseAggregator` now lives behind the pluggable backend
+in `backend/` (see `backend/__init__.py`). Leaf plugins keep importing it from
+here; the active backend is chosen by the PYML_BACKEND environment variable.
+"""
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-gi.require_version("GLib", "2.0")
-from gi.repository import Gst, GObject, GstBase # noqa: E402
+from backend import BaseAggregator
-from log.logger_factory import LoggerFactory
-from engine.engine_manager import EngineManager
-
-
-class BaseAggregator(GstBase.Aggregator):
- """
- Base class for GStreamer aggregator elements that perform inference
- with a machine learning model. This class manages shared properties
- and handles model loading and device management via MLEngine.
- """
-
- __gstmetadata__ = (
- "BaseAggregator",
- "Aggregator",
- "Generic machine learning model aggregator element",
- "Aaron Boxer ",
- )
-
- def __init__(self):
- super().__init__()
- self.logger = LoggerFactory.get(LoggerFactory.LOGGER_TYPE_GST)
- self.mgr = EngineManager(self.logger)
- self.kwargs = {}
- self.__batch_size = 1
- self.__frame_stride = 1
- self.__model_name = None
- self.__device_queue_id = 0
- self.__system_prompt = None
- self._prompt = None
- self.__compile = False
- self.segment_pushed = False
-
- @property
- def engine(self):
- return self.mgr.engine
-
- @GObject.Property(type=str)
- def device(self):
- "Device to run the inference on (cpu, cuda, cuda:0, cuda:1, etc.)"
- return self.mgr.device
-
- @device.setter
- def device(self, value):
- self.mgr.set_device(value)
- # todo why is this needed ?
- if self.engine_name:
- self.initialize_engine()
-
- @GObject.Property(type=int, default=1)
- def batch_size(self):
- "Number of items to process in a batch"
- return self.__batch_size
-
- @batch_size.setter
- def batch_size(self, value):
- self.__batch_size = value
- if self.engine:
- self.engine.batch_size = value
-
- @GObject.Property(type=int, default=1)
- def frame_stride(self):
- "How often to process a frame"
- return self.__frame_stride
-
- @frame_stride.setter
- def frame_stride(self, value):
- self.__frame_stride = value
- if self.engine:
- self.engine.frame_stride = value
-
- @GObject.Property(type=str)
- def model_name(self):
- "Name of the pre-trained model or local model path"
- return self.__model_name
-
- @model_name.setter
- def model_name(self, value):
- self.__model_name = value
-
- @GObject.Property(type=str)
- def engine_name(self):
- "Machine Learning Engine to use : pytorch, tflite, tensorflow, onnx or openvino, or custom engine name"
- return self.mgr.engine_name
-
- @engine_name.setter
- def engine_name(self, value):
- self.mgr.engine_name = value
-
- @GObject.Property(type=int, default=1)
- def device_queue_id(self):
- "ID of the DeviceQueue from the pool to use"
- return self.__device_queue_id
-
- @device_queue_id.setter
- def device_queue_id(self, value):
- self.__device_queue_id = value
- if self.engine:
- self.engine.device_queue_id = value
-
- @GObject.Property(type=bool, default=False, nick="compile")
- def compile(self):
- "Enable torch.compile optimization for the model"
- return self.__compile
-
- @compile.setter
- def compile(self, value):
- self.__compile = value
- if value:
- self.kwargs["compile"] = True
- else:
- self.kwargs.pop("compile", None)
-
- def do_change_state(self, transition):
- if transition == Gst.StateChange.NULL_TO_READY:
- self.do_load_model()
- return Gst.Element.do_change_state(self, transition)
-
- def initialize_engine(self):
- if not self.engine and self.mgr.engine_name:
- self.mgr.initialize_engine()
- self.engine.batch_size = self.__batch_size
- self.engine.frame_stride = self.__frame_stride
- if self.__device_queue_id:
- self.engine.device_queue_id = self.__device_queue_id
- if not self.engine:
- self.logger.error(f"Unsupported ML engine: {self.mgr.engine_name}")
-
- def do_load_model(self):
- self.initialize_engine()
- if self.engine is None:
- self.logger.error(
- f"Cannot load model {self.model_name}: engine not initialized"
- )
- return
- if self.model_name is None:
- self.logger.warning("Cannot load model as model name is not set")
- return
- self.mgr.do_load_model(self.model_name, **self.kwargs)
-
- def get_model(self):
- """Gets the model from the engine."""
- self.initialize_engine()
- if self.engine is None:
- self.logger.error(
- f"Cannot get model {self.model_name}: engine not initialized"
- )
- return None
- """Gets the model from the engine."""
- if self.engine:
- return self.engine.get_model()
- return None
-
- def set_model(self, model):
- """Sets the model in the engine."""
- self.initialize_engine()
- if self.engine is None:
- self.logger.error("Cannot load model: engine not initialized")
- return False
- self.engine.model = model
- self.logger.info("Model set successfully in the engine.")
-
- def get_tokenizer(self):
- self.initialize_engine()
- if self.engine is None:
- self.logger.error("Cannot get tokenizer: engine not initialized")
- return None
- return self.mgr.get_tokenizer()
-
- def push_segment_if_needed(self):
- if not self.segment_pushed:
- segment = Gst.Segment()
- segment.init(Gst.Format.TIME)
- segment.start = 0
- segment.stop = Gst.CLOCK_TIME_NONE
- segment.position = 0
-
- self.srcpad.push_event(Gst.Event.new_segment(segment))
- self.segment_pushed = True
-
- def do_aggregate(self, timeout):
- if all(pad.is_eos() for pad in self.sinkpads):
- return Gst.FlowReturn.EOS
- self.push_segment_if_needed()
- self.process_all_sink_pads()
- self.selected_samples(Gst.CLOCK_TIME_NONE, 0, 0, None)
- return Gst.FlowReturn.OK
-
- def process_all_sink_pads(self):
- if len(self.sinkpads) == 0:
- return
- buf = self.sinkpads[0].pop_buffer()
- if buf:
- self.do_process(buf)
-
- @abstractmethod
- def do_process(self, buf):
- pass
+__all__ = ["BaseAggregator"]
diff --git a/plugins/python/base_caption.py b/plugins/python/base_caption.py
index 2daefe3..fef65ef 100644
--- a/plugins/python/base_caption.py
+++ b/plugins/python/base_caption.py
@@ -17,25 +17,29 @@
# Boston, MA 02110-1301, USA.
-from utils.muxed_buffer_processor import MuxedBufferProcessor # Added import
+import backend
+from backend import GObject, analytics
+from video_transform import VideoTransform
-import gi
+# The text pad is a GStreamer request pad: a hosted element on g2g has one
+# source pad and stages its caption as metadata instead.
+if backend.BACKEND == "gst":
+ import gi
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-gi.require_version("GstVideo", "1.0")
-gi.require_version("GLib", "2.0")
-gi.require_version("GstAnalytics", "1.0")
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ from gi.repository import Gst, GstBase # noqa: E402
-from gi.repository import Gst, GObject, GstAnalytics, GLib, GstBase # noqa: E402
-from video_transform import VideoTransform
+ TEXT_CAPS = Gst.Caps.from_string("text/x-raw, format=utf8")
-TEXT_CAPS = Gst.Caps.from_string("text/x-raw, format=utf8")
+#: How long a caption stays on screen. Long enough that the subtitle is still up
+#: when the next frame's caption arrives.
+CAPTION_DURATION_SECONDS = 60
class BaseCaption(VideoTransform):
"""
- Base GStreamer element for captioning video frames.
+ Base element for captioning video frames.
"""
__gstmetadata__ = (
@@ -45,34 +49,35 @@ class BaseCaption(VideoTransform):
"Aaron Boxer ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new(
- "text_src", Gst.PadDirection.SRC, Gst.PadPresence.REQUEST, TEXT_CAPS
- ),
- )
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new(
+ "text_src", Gst.PadDirection.SRC, Gst.PadPresence.REQUEST, TEXT_CAPS
+ ),
+ )
def __init__(self):
super().__init__()
- self.__prompt = "What is shown in this image?"
+ self._prompt = "What is shown in this image?"
self.text_src_pad = None
@GObject.Property(type=str)
def system_prompt(self):
"Custom system prompt text"
- return self.__system_prompt
+ return self._system_prompt
@system_prompt.setter
def system_prompt(self, value):
- self.__system_prompt = value
+ self._system_prompt = value
@GObject.Property(type=str)
def prompt(self):
"Custom prompt text"
- return self.__prompt
+ return self._prompt
@prompt.setter
def prompt(self, value):
- self.__prompt = value
+ self._prompt = value
# make read only
@GObject.Property(type=str)
@@ -86,6 +91,75 @@ def engine_name(self, value):
"The 'engine_name' property cannot be set in this derived class."
)
+ def forward(self, frames):
+ return self.engine.do_forward(frames) if self.engine else None
+
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Caption each source, staging one classification per caption.
+
+ Captions carry no pixels and no blob, so this replaces the shared
+ infer-decode-write body rather than filling in `decode`.
+ """
+ result = self.forward(frames)
+ if result is None:
+ raise RuntimeError(f"{type(self).__name__}: captioning returned None")
+
+ captions = result if isinstance(result, list) else [result] * num_sources
+ if len(captions) != num_sources:
+ raise RuntimeError(f"expected {num_sources} captions, got {len(captions)}")
+
+ meta = analytics.add_relation_meta(target)
+ if meta is None:
+ self.logger.error("Failed to add analytics metadata to buffer")
+ return
+
+ for index, caption in enumerate(captions):
+ if not caption:
+ self.logger.warning(f"stream {index}: no caption generated")
+ continue
+ label = caption if num_sources == 1 else f"stream_{index}_{caption}"
+ if analytics.add_classification(meta, index, label) is None:
+ self.logger.error(f"stream {index}: failed to add the caption")
+ else:
+ self.logger.info(f"stream {index}: added caption {caption}")
+
+ self.push_captions(captions, target)
+
+ def push_captions(self, captions, buf):
+ """Send each caption out the text pad, one buffer per source.
+
+ The caption is staged as metadata either way, so the pad is an extra:
+ nothing to do unless something asked for it, which is also what makes
+ this a no-op on a backend that has no request pads.
+ """
+ if self.text_src_pad is None:
+ return
+
+ if buf.pts == Gst.CLOCK_TIME_NONE:
+ buf.pts = Gst.util_uint64_scale(
+ Gst.util_get_timestamp(),
+ 1, # framerate_denom
+ 30 * Gst.SECOND, # framerate_num
+ )
+ if buf.duration == Gst.CLOCK_TIME_NONE:
+ buf.duration = Gst.SECOND // 30 # framerate_num
+
+ share = buf.duration // len(captions)
+ for index, caption in enumerate(captions):
+ if caption:
+ self.push_text_buffer(caption, buf.pts + index * share, buf.dts)
+
+ def push_text_buffer(self, text, pts, dts):
+ """Push one caption to the `text_src` pad, timed with its video frame."""
+ text_buffer = Gst.Buffer.new_wrapped(text.encode("utf-8"))
+ text_buffer.pts = pts
+ text_buffer.dts = dts
+ text_buffer.duration = CAPTION_DURATION_SECONDS * Gst.SECOND
+
+ ret = self.text_src_pad.push(text_buffer)
+ if ret != Gst.FlowReturn.OK:
+ self.logger.warning(f"Failed to push text buffer: {ret}")
+
def do_request_new_pad(self, template, name, caps):
if self.text_src_pad:
self.logger.error("Element already has a text_src")
@@ -105,155 +179,6 @@ def do_release_pad(self, pad):
pad.set_active(False)
self.text_src_pad = None
- def push_text_buffer(self, text, buf_pts, buf_dts, buf_duration):
- """
- Pushes a text buffer to the `text_src` pad with proper timestamps.
-
- Args:
- text (str): The text to push as a buffer.
- buf_pts (int): The PTS of the associated video buffer.
- buf_duration (int): The duration of the associated video buffer.
- """
- text_buffer = Gst.Buffer.new_wrapped(text.encode("utf-8"))
-
- # Set the text buffer timestamps
- text_buffer.pts = buf_pts
- text_buffer.dts = buf_dts
- # Put a long duration so the subtitles are visible
- text_buffer.duration = 60 * Gst.SECOND
-
- # Push the buffer
- ret = self.text_src_pad.push(text_buffer)
- if ret != Gst.FlowReturn.OK:
- self.logger.warning(f"Failed to push text buffer: {ret}")
-
- def do_transform_ip(self, buf):
- """
- In-place transformation for captioning inference using MuxedBufferProcessor.
- """
- try:
- # Initialize MuxedBufferProcessor with default framerate
- muxed_processor = MuxedBufferProcessor(
- self.logger,
- self.width,
- self.height,
- framerate_num=30,
- framerate_denom=1,
- )
- frames, id_str, num_sources, format = muxed_processor.extract_frames(
- buf, self.sinkpad
- )
- if frames is None:
- self.logger.error("Failed to extract frames")
- return Gst.FlowReturn.ERROR
-
- # Set timestamps if none are set
- if buf.pts == Gst.CLOCK_TIME_NONE:
- buf.pts = Gst.util_uint64_scale(
- Gst.util_get_timestamp(),
- 1, # framerate_denom
- 30 * Gst.SECOND, # framerate_num
- )
- if buf.duration == Gst.CLOCK_TIME_NONE:
- buf.duration = Gst.SECOND // 30 # framerate_num
-
- # Process frames (single or batch)
- if num_sources == 1:
- # Single-frame case
- frame = frames
- if self.engine:
- result = self.engine.do_forward(frame)
- if result:
- self.caption = result
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
- if meta:
- qk = GLib.quark_from_string(f"{result}")
- ret, mtd = meta.add_one_cls_mtd(0, qk)
- if ret:
- self.logger.info(f"Successfully added caption {result}")
- else:
- self.logger.error(
- "Failed to add classification metadata"
- )
- else:
- self.logger.error(
- "Failed to add GstAnalytics metadata to buffer"
- )
-
- # Push text buffer if text_src pad is linked
- if self.text_src_pad:
- self.push_text_buffer(
- self.caption, buf.pts, buf.dts, buf.duration
- )
- else:
- self.logger.warning(
- "TextExtract: text_src pad is not linked, cannot push text buffer."
- )
- else:
- # Batch case
- self.logger.info(
- f"Processing batch with ID={id_str}, num_sources={num_sources}"
- )
- if self.engine:
- results = self.engine.do_forward(frames)
- if results is None:
- self.logger.error("Inference returned None")
- return Gst.FlowReturn.ERROR
-
- # Ensure results is a list for batch processing
- results_list = (
- results
- if isinstance(results, list)
- else [results] * num_sources
- )
- if len(results_list) != num_sources:
- self.logger.error(
- f"Expected {num_sources} results, got {len(results_list)}"
- )
- return Gst.FlowReturn.ERROR
-
- for idx, result in enumerate(results_list):
- if result:
- caption = result
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
- if meta:
- qk = GLib.quark_from_string(f"stream_{idx}_{result}")
- ret, mtd = meta.add_one_cls_mtd(idx, qk)
- if ret:
- self.logger.info(
- f"Stream {idx}: Successfully added caption {result}"
- )
- else:
- self.logger.error(
- f"Stream {idx}: Failed to add classification metadata"
- )
- else:
- self.logger.error(
- f"Stream {idx}: Failed to add GstAnalytics metadata"
- )
-
- # Push text buffer for each frame
- if self.text_src_pad:
- # Adjust PTS for each frame in the batch
- frame_pts = buf.pts + (
- idx * (buf.duration // num_sources)
- )
- self.push_text_buffer(
- caption, frame_pts, buf.duration // num_sources
- )
- else:
- self.logger.warning(
- f"Stream {idx}: TextExtract: text_src pad is not linked, cannot push text buffer."
- )
- else:
- self.logger.warning(f"Stream {idx}: No caption generated")
-
- return Gst.FlowReturn.OK
-
- except Exception as e:
- self.logger.error(f"Error during transformation: {e}")
- return Gst.FlowReturn.ERROR
-
def do_sink_event(self, event):
if self.text_src_pad:
text_event = (
diff --git a/plugins/python/base_classifier.py b/plugins/python/base_classifier.py
index 090976c..4a830b2 100644
--- a/plugins/python/base_classifier.py
+++ b/plugins/python/base_classifier.py
@@ -18,15 +18,9 @@
from utils.runtime_utils import runtime_check_gstreamer_version
-import gi
from video_transform import VideoTransform
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-gi.require_version("GstVideo", "1.0")
-gi.require_version("GstAnalytics", "1.0")
-gi.require_version("GLib", "2.0")
-from gi.repository import Gst, GstAnalytics, GLib # noqa: E402
+from backend import analytics
class BaseClassifier(VideoTransform):
@@ -48,36 +42,12 @@ def do_forward(self, frame):
self.logger.error("No model loaded in BaseClassifier.")
return None
- def do_transform_ip(self, buf):
- """
- Processes an image and attaches classification metadata.
- """
- import numpy as np
-
- try:
- with buf.map(Gst.MapFlags.READ | Gst.MapFlags.WRITE) as info:
- if info.data is None:
- self.logger.error("Buffer mapping returned None data.")
- return Gst.FlowReturn.ERROR
-
- frame = np.array(info.data, dtype=np.uint8).reshape(
- self.height, self.width, 3
- )
-
- # Perform classification
- results = self.do_forward(frame)
- if not results:
- self.logger.warning("Classification returned no results.")
- return Gst.FlowReturn.ERROR
-
- # Process results
- self.do_decode(buf, results)
-
- return Gst.FlowReturn.OK
-
- except Exception as e:
- self.logger.error(f"do_transform_ip: Unexpected error: {e}")
- return Gst.FlowReturn.ERROR
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Classify the frame and attach the label as metadata."""
+ results = self.do_forward(frames)
+ if not results:
+ raise RuntimeError("classification returned no results")
+ self.do_decode(target, results)
def do_decode(self, buf, output):
"""
@@ -111,8 +81,9 @@ def do_decode(self, buf, output):
self.logger.info(f"Classified as {label} with confidence score {score:.2f}")
# Attach classification metadata
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
+ meta = analytics.add_relation_meta(buf)
if meta:
- qk = GLib.quark_from_string(f"class_{label}")
- meta.add_od_mtd(qk, 0, 0, self.width, self.height, score)
+ analytics.add_object(
+ meta, f"class_{label}", 0, 0, self.width, self.height, score
+ )
self.logger.info(f"Classified as {label} with score {score:.2f}")
diff --git a/plugins/python/base_llm.py b/plugins/python/base_llm.py
index a4da09a..28cc494 100644
--- a/plugins/python/base_llm.py
+++ b/plugins/python/base_llm.py
@@ -17,138 +17,96 @@
# Boston, MA 02110-1301, USA.
-import gi
+import backend
+from backend import GObject
+from base_aggregator import BaseAggregator
-gi.require_version("Gst", "1.0")
-gi.require_version("GLib", "2.0")
-from gi.repository import Gst, GObject # noqa: E402
+if backend.BACKEND == "gst":
+ import gi
-from base_aggregator import BaseAggregator
+ gi.require_version("Gst", "1.0")
+ from gi.repository import Gst # noqa: E402
class BaseLlm(BaseAggregator):
"""
- GStreamer base element that performs language model inference
- with a PyTorch model.
+ Base element that performs language model inference with a PyTorch model.
"""
+ PUSH_FROM_SRC_PAD = True
+
@GObject.Property(type=str)
def system_prompt(self):
"Custom system prompt text"
- return self.__system_prompt
+ return self._system_prompt
@system_prompt.setter
def system_prompt(self, value):
- self.__system_prompt = value
+ self._system_prompt = value
@GObject.Property(type=str)
def prompt(self):
"Custom prompt text"
- return self.__prompt
+ return self._prompt
@prompt.setter
def prompt(self, value):
- self.__prompt = value
-
- __gsttemplates__ = (
- Gst.PadTemplate.new(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- Gst.Caps.from_string("text/x-raw,format=utf8"),
- ),
- Gst.PadTemplate.new(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.REQUEST,
- Gst.Caps.from_string("text/x-raw,format=utf8"),
- ),
- )
-
- def __init__(self):
- super().__init__()
-
- def do_process(self, buf):
- """
- Processes the input buffer with the language model
- and pushes the result downstream.
- """
- try:
- # Map buffer to read input text
- success, map_info = buf.map(Gst.MapFlags.READ)
- if not success:
- self.logger.error("Failed to map buffer")
- return Gst.FlowReturn.ERROR
-
- # Convert memoryview to bytes and decode to string
- input_text = bytes(map_info.data).decode("utf-8")
- self.logger.info(f"Received text for LLM processing: {input_text}")
-
- # Ensure engine is initialized
- if not self.engine:
- self.logger.info("Engine not initialized, initializing now")
- self.mgr.initialize_engine()
- self.mgr.do_load_model(self.model_name)
-
- # Retry model loading if tokenizer or model is missing
+ self._prompt = value
+
+ # the caps each pad negotiates, stated once for both backends
+ INPUT_CAPS = "text/x-raw,format=utf8"
+ OUTPUT_CAPS = "text/x-raw,format=utf8"
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string(OUTPUT_CAPS),
+ ),
+ Gst.PadTemplate.new(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.REQUEST,
+ Gst.Caps.from_string(INPUT_CAPS),
+ ),
+ )
+
+ def process_payload(self, payload: bytes) -> list[bytes]:
+ """Generates a reply to the input text with the language model."""
+ input_text = payload.decode("utf-8")
+ self.logger.info(f"Received text for LLM processing: {input_text}")
+
+ # Ensure engine is initialized
+ if not self.engine:
+ self.logger.info("Engine not initialized, initializing now")
+ self.mgr.initialize_engine()
+ self.mgr.do_load_model(self.model_name)
+
+ # Retry model loading if tokenizer or model is missing
+ tokenizer = self.get_tokenizer()
+ model = self.get_model()
+ self.logger.info(f"Tokenizer: {tokenizer}")
+ self.logger.info(f"Model: {model}")
+ if not tokenizer or not model:
+ self.logger.error(
+ f"Tokenizer initialized: {tokenizer is not None}, Model initialized: {model is not None}"
+ )
+ self.logger.warning("Attempting to reload model")
+ if not self.mgr.do_load_model(self.model_name):
+ self.logger.error("Model reload failed")
+ return []
tokenizer = self.get_tokenizer()
model = self.get_model()
- self.logger.info(f"Tokenizer: {tokenizer}")
- self.logger.info(f"Model: {model}")
if not tokenizer or not model:
- self.logger.error(
- f"Tokenizer initialized: {tokenizer is not None}, Model initialized: {model is not None}"
- )
- self.logger.warning("Attempting to reload model")
- if not self.mgr.do_load_model(self.model_name):
- self.logger.error("Model reload failed")
- buf.unmap(map_info)
- return Gst.FlowReturn.ERROR
- tokenizer = self.get_tokenizer()
- model = self.get_model()
- if not tokenizer or not model:
- self.logger.error("Model reload failed again")
- buf.unmap(map_info)
- return Gst.FlowReturn.ERROR
-
- # Generate text using the engine
- generated_text = self.engine.do_generate(
- input_text, system_prompt=self.system_prompt
- )
- self.logger.info(f"Generated text: {generated_text}")
-
- buf.unmap(map_info)
-
- # Push the generated text downstream
- return self.push_generated_text(buf, generated_text)
-
- except Exception as e:
- self.logger.error(f"Error in LLM processing: {e}")
- return Gst.FlowReturn.ERROR
-
- def push_generated_text(self, inbuf, generated_text):
- """
- Push the generated text downstream.
- """
- try:
- generated_bytes = generated_text.encode("utf-8")
- outbuf = Gst.Buffer.new_allocate(None, len(generated_bytes), None)
- success, map_info_out = outbuf.map(Gst.MapFlags.WRITE)
- if not success:
- self.logger.error("Failed to map output buffer for writing")
- return Gst.FlowReturn.ERROR
-
- map_info_out.data[: len(generated_bytes)] = generated_bytes
- outbuf.unmap(map_info_out)
- outbuf.pts = inbuf.pts
- outbuf.dts = inbuf.dts
- outbuf.duration = inbuf.duration
-
- # Push the buffer downstream
- self.logger.info("Pushed generated text downstream")
- ret = self.srcpad.push(outbuf)
-
- return ret
-
- except Exception as e:
- self.logger.error(f"Error pushing generated text: {e}")
+ self.logger.error("Model reload failed again")
+ return []
+
+ generated_text = self.engine.do_generate(
+ input_text, system_prompt=self.system_prompt
+ )
+ self.logger.info(f"Generated text: {generated_text}")
+
+ return [generated_text.encode("utf-8")]
diff --git a/plugins/python/base_objectdetector.py b/plugins/python/base_objectdetector.py
index 84f2f2d..335fbc6 100644
--- a/plugins/python/base_objectdetector.py
+++ b/plugins/python/base_objectdetector.py
@@ -16,25 +16,21 @@
# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
# Boston, MA 02110-1301, USA.
-import traceback
from utils.runtime_utils import runtime_check_gstreamer_version
-import gi
from video_transform import VideoTransform
from utils.format_converter import FormatConverter
-from utils.muxed_buffer_processor import MuxedBufferProcessor # Added import
+from backend import analytics, GObject
+from tasks.object_detector import ObjectDetectorTask
+from utils.metadata import Metadata
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-gi.require_version("GstAnalytics", "1.0")
-gi.require_version("GLib", "2.0")
-from gi.repository import Gst, GstAnalytics, GObject, GLib # noqa: E402
-from utils.metadata import Metadata # noqa: E402
-
-class BaseObjectDetector(VideoTransform):
+class BaseObjectDetector(VideoTransform, ObjectDetectorTask):
"""
- GStreamer element for object detection with batch processing support.
- Handles both single-frame buffers (no metadata) and batch buffers (metadata in last chunk).
+ GStreamer element shell for object detection with batch processing support.
+ Handles both single-frame buffers (no metadata) and batch buffers (metadata
+ in last chunk). The inference and metadata steps (do_forward / do_decode)
+ are inherited from the backend-agnostic ObjectDetectorTask; this class only
+ supplies the GStreamer per-buffer glue.
"""
def __init__(self):
@@ -46,6 +42,10 @@ def __init__(self):
self.metadata = Metadata("si")
self.logger.info("Initialized BaseObjectDetector")
self.__track = False
+ self.__interval = 1
+ self._det_counter = 0
+ self._cached_results = None
+ self._cached_num_sources = 1
@GObject.Property(type=bool, default=False)
def track(self):
@@ -60,130 +60,77 @@ def track(self, value):
if self.engine:
self.engine.track = value
- def do_forward(self, frames):
- self.logger.info(
- f"Forward called with frames shape: {frames.shape if frames is not None else 'None'}"
- )
- if self.engine:
- self.engine.track = self.track
- result = self.engine.do_forward(frames)
- self.logger.debug(f"Forward result: {result} (type: {type(result)})")
- return result
- return None
-
- def do_transform_ip(self, buf):
- """
- Transform the input buffer using MuxedBufferProcessor for frame extraction.
- """
- self.logger.info(f"Transforming buffer: {hex(id(buf))}")
- try:
- # Use MuxedBufferProcessor to extract frames and metadata
- muxed_processor = MuxedBufferProcessor(
- self.logger,
- self.width,
- self.height,
- self.framerate_num,
- self.framerate_denom,
- )
- frames, id_str, num_sources, format = muxed_processor.extract_frames(
- buf, self.sinkpad
- )
- if frames is None:
- self.logger.error("Failed to extract frames")
- return Gst.FlowReturn.ERROR
-
- # Process frames (single or batch)
- results = self.do_forward(frames)
- if results is None:
- self.logger.error("Inference returned None")
- return Gst.FlowReturn.ERROR
+ @GObject.Property(type=int, default=1, minimum=1, maximum=10000)
+ def interval(self):
+ "Run detection every Nth frame and re-attach the previous detections on"
+ "the frames in between (N=1 runs detection every frame). Lets downstream "
+ "tracking/overlay stay per-frame while detection runs at a lower rate."
+ return self.__interval
- # Handle single-frame case
- if num_sources == 1:
- self.do_decode(buf, results, stream_idx=0)
- # Handle batch case
- else:
- self.logger.info(
- f"Processing batch with ID={id_str}, num_sources={num_sources}"
- )
- results_list = results if isinstance(results, list) else [results]
- if len(results_list) != num_sources:
- self.logger.error(
- f"Expected {num_sources} results, got {len(results_list)}"
- )
- return Gst.FlowReturn.ERROR
+ @interval.setter
+ def interval(self, value):
+ self.__interval = max(1, int(value))
- for idx, result in enumerate(results_list):
- if result is None:
- self.logger.warning(f"Frame {idx} result is None")
- continue
- self.do_decode(buf, result, stream_idx=idx)
+ @GObject.Property(type=str, default="30/1")
+ def framerate(self):
+ "Source framerate as 'num/denom', used for muxed-stream frame timing"
+ return f"{self.framerate_num}/{self.framerate_denom}"
- attached_meta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
- if attached_meta:
- count = GstAnalytics.relation_get_length(attached_meta)
- self.logger.info(f"Total metadata relations attached: {count}")
- else:
- self.logger.debug("No detections on this buffer")
+ @framerate.setter
+ def framerate(self, value):
+ try:
+ num, denom = str(value).split("/")
+ self.framerate_num = int(num)
+ self.framerate_denom = int(denom)
+ except (ValueError, AttributeError):
+ self.logger.warning(f"Invalid framerate '{value}', expected 'num/denom'")
- return Gst.FlowReturn.OK
+ def process_frames(self, frames, num_sources, fmt, target):
+ """
+ Run detection on the extracted frame(s) and attach results to `target`.
- except Exception as e:
- self.logger.error(f"Transform error: {e}\n{traceback.format_exc()}")
- return Gst.FlowReturn.ERROR
+ Backend-agnostic per-frame hook: frame extraction and the FlowReturn
+ wrapping live in the backend driver (the gst `do_transform_ip` /
+ the g2g `g2g_process`), so the inference + metadata here is identical on
+ every backend. Raises on a hard failure; the driver maps that to its own
+ error return.
+ """
+ run_detect = (self._det_counter % self.__interval) == 0
+ self._det_counter += 1
- def do_decode(self, buf, output, stream_idx=0):
- self.logger.info(
- f"Decoding for stream {stream_idx}: {output} (type: {type(output)})"
- )
- if isinstance(output, dict):
- self.logger.info(f"Stream {stream_idx} - Processing dict")
- boxes = output["boxes"]
- labels = output["labels"]
- scores = output["scores"]
- elif hasattr(output, "boxes"): # Direct Results object (e.g., Ultralytics YOLO)
- self.logger.info(f"Stream {stream_idx} - Processing Ultralytics Results")
- boxes = output.boxes.xyxy.cpu().numpy() # [N, 4]
- scores = output.boxes.conf.cpu().numpy() # [N]
- labels = output.boxes.cls.cpu().numpy().astype(int) # [N]
- elif (
- isinstance(output, list) and len(output) >= 6
- ): # [x1, y1, x2, y2, score, label]
- self.logger.info(f"Stream {stream_idx} - Processing list of detections")
- boxes = [[det[0], det[1], det[2], det[3]] for det in output]
- scores = [det[4] for det in output]
- labels = [int(det[5]) for det in output]
+ if run_detect:
+ results = self.do_forward(frames)
+ if results is None:
+ raise RuntimeError("inference returned None")
+ self._cached_results = results
+ self._cached_num_sources = num_sources
+ self._decode_results(target, results, num_sources)
+ elif self._cached_results is not None:
+ # Skip inference on this frame and re-attach the previous
+ # detections so downstream tracking/overlay stay per-frame.
+ self._decode_results(target, self._cached_results, self._cached_num_sources)
+
+ attached_meta = analytics.get_relation_meta(target)
+ if attached_meta:
+ count = analytics.relation_length(attached_meta)
+ self.logger.info(f"Total metadata relations attached: {count}")
else:
- self.logger.error(
- f"Stream {stream_idx} - Unrecognized format: {output} (type: {type(output)})"
- )
- return
-
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
- if not meta:
- self.logger.error(
- f"Stream {stream_idx} - Failed to add analytics relation metadata"
- )
- return
+ self.logger.debug("No detections on this buffer")
- self.logger.info(f"Stream {stream_idx} - Adding {len(boxes)} detections")
- for i, (box, label, score) in enumerate(zip(boxes, labels, scores)):
- x1, y1, x2, y2 = box
- qk_string = f"stream_{stream_idx}_label_{label}"
- qk = GLib.quark_from_string(qk_string)
- ret, od_mtd = meta.add_od_mtd(qk, x1, y1, x2 - x1, y2 - y1, score)
- if not ret:
- self.logger.error(
- f"Stream {stream_idx} - Failed to add od_mtd for detection {i}"
+ def _decode_results(self, target, results, num_sources):
+ # Single-frame case
+ if num_sources == 1:
+ self.do_decode(target, results, stream_idx=0)
+ # Batch case
+ else:
+ self.logger.info(f"Processing batch with num_sources={num_sources}")
+ results_list = results if isinstance(results, list) else [results]
+ if len(results_list) != num_sources:
+ raise RuntimeError(
+ f"expected {num_sources} results, got {len(results_list)}"
)
- continue
- self.logger.info(
- f"Stream {stream_idx} - Added detection {i}: label={qk_string}, x1={x1}, y1={y1}, w={x2-x1}, h={y2-y1}, score={score}"
- )
-
- attached_meta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
- if attached_meta:
- count = GstAnalytics.relation_get_length(attached_meta)
- self.logger.info(
- f"Stream {stream_idx} - Metadata relations after adding: {count}"
- )
+ for idx, result in enumerate(results_list):
+ if result is None:
+ self.logger.warning(f"Frame {idx} result is None")
+ continue
+ self.do_decode(target, result, stream_idx=idx)
diff --git a/plugins/python/base_separate.py b/plugins/python/base_separate.py
index f793b5e..8164620 100644
--- a/plugins/python/base_separate.py
+++ b/plugins/python/base_separate.py
@@ -21,14 +21,17 @@
from abc import abstractmethod
import traceback
-import gi
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-gi.require_version("GObject", "2.0")
-from gi.repository import Gst, GObject, GstBase # noqa: E402
+import backend
+from backend import GObject
+from base_aggregator import BaseAggregator
-from base_aggregator import BaseAggregator # noqa: E402
+if backend.BACKEND == "gst":
+ import gi
+
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ from gi.repository import Gst, GstBase # noqa: E402
sys.stdout.reconfigure(encoding="utf-8")
sys.stderr.reconfigure(encoding="utf-8")
@@ -44,32 +47,29 @@ class BaseSeparate(BaseAggregator):
SAMPLE_RATE = 44100 # native sample rate of Demucs
- CAPS = Gst.Caps(
- Gst.Structure(
- "audio/x-raw",
- format="S16LE",
- layout="interleaved",
- rate=SAMPLE_RATE,
- channels=1,
+ # the caps each pad negotiates, stated once for both backends. The rate has
+ # to match SAMPLE_RATE, which the chunking works from.
+ INPUT_CAPS = "audio/x-raw,format=S16LE,layout=interleaved,rate=44100,channels=1"
+ OUTPUT_CAPS = "audio/x-raw,format=S16LE,layout=interleaved,rate=44100,channels=1"
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new_with_gtype(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.REQUEST,
+ Gst.Caps.from_string(INPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
+ Gst.PadTemplate.new_with_gtype(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string(OUTPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
)
- )
-
- __gsttemplates__ = (
- Gst.PadTemplate.new_with_gtype(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.REQUEST,
- CAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- Gst.PadTemplate.new_with_gtype(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- CAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- )
def __init__(self):
super().__init__()
@@ -102,84 +102,36 @@ def do_separate(self, audio_data):
def do_process_audio(self, audio_data):
# audio_data is np.int16 array
- encoded_size = audio_data.nbytes
- outbuf = Gst.Buffer.new_allocate(None, encoded_size, None)
- outbuf.fill(0, audio_data.tobytes())
- return outbuf
-
- def do_process(self, buf):
- import numpy as np
+ return audio_data.tobytes()
- self.push_segment_if_needed()
- """Process audio data from the input buffers using source separation."""
- audio_collected = False
+ def process_payload(self, payload: bytes) -> list[bytes]:
+ """Separates the requested stem, one payload per full chunk of audio.
- try:
- # Map the buffer to access the audio data
- success, map_info = buf.map(Gst.MapFlags.READ)
- if not success:
- self.logger.error("Failed to map input buffer")
- return Gst.FlowReturn.ERROR
-
- # Convert buffer to numpy array (int16)
- audio_data = np.frombuffer(map_info.data, dtype=np.int16)
- audio_collected = True
-
- self.clip_buffer.extend(audio_data)
-
- chunk_duration = 1.0 if self.streaming else 10.0
- chunk_size = int(self.SAMPLE_RATE * chunk_duration)
-
- while len(self.clip_buffer) >= chunk_size:
- chunk = np.fromiter(
- (self.clip_buffer.popleft() for _ in range(chunk_size)),
- dtype=np.int16,
- )
- separated = self._separate_audio(chunk)
- if separated is None:
- self.logger.warning("Empty separated audio")
- buf.unmap(map_info)
- return Gst.FlowReturn.ERROR
-
- self._process_and_send(separated, buf)
-
- # Handle remaining buffer on EOS
- if buf.flags & Gst.BufferFlags.LAST:
- if len(self.clip_buffer) > 0:
- chunk = np.fromiter(self.clip_buffer, dtype=np.int16)
- separated = self._separate_audio(chunk)
- if separated is None:
- self.logger.warning("Empty separated audio")
- buf.unmap(map_info)
- return Gst.FlowReturn.ERROR
-
- self._process_and_send(separated, buf)
- self.clip_buffer.clear()
-
- buf.unmap(map_info)
- except Exception as e:
- self.logger.error(f"Error during buffer processing: {e}")
- if audio_collected:
- buf.unmap(map_info)
- return Gst.FlowReturn.ERROR
+ Audio is accumulated until there is a whole chunk to separate, so a
+ buffer produces anywhere from zero to several payloads.
+ """
+ import numpy as np
- if not audio_collected:
- self.logger.warning("No audio data collected from sink pads.")
- return Gst.FlowReturn.ERROR
+ audio_data = np.frombuffer(payload, dtype=np.int16)
+ self.clip_buffer.extend(audio_data)
- return Gst.FlowReturn.OK
+ chunk_duration = 1.0 if self.streaming else 10.0
+ chunk_size = int(self.SAMPLE_RATE * chunk_duration)
- def _process_and_send(self, separated, inbuf):
- outbuf = self.do_process_audio(separated)
- if outbuf is None:
- return
+ payloads = []
+ while len(self.clip_buffer) >= chunk_size:
+ chunk = np.fromiter(
+ (self.clip_buffer.popleft() for _ in range(chunk_size)),
+ dtype=np.int16,
+ )
+ separated = self._separate_audio(chunk)
+ if separated is None:
+ self.logger.warning("Empty separated audio")
+ return payloads
- # Set PTS and duration from the input buffer
- outbuf.pts = inbuf.pts
- outbuf.duration = inbuf.duration
+ payloads.append(self.do_process_audio(separated))
- # Push the separated audio downstream
- self.finish_buffer(outbuf)
+ return payloads
def _separate_audio(self, chunk):
"""
diff --git a/plugins/python/base_transcribe.py b/plugins/python/base_transcribe.py
index b690a4c..79e9da4 100644
--- a/plugins/python/base_transcribe.py
+++ b/plugins/python/base_transcribe.py
@@ -20,32 +20,22 @@
import sys
from abc import abstractmethod
-import gi
+import backend
+from backend import GObject
+from base_aggregator import BaseAggregator
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-gi.require_version("GObject", "2.0")
-from gi.repository import Gst, GObject, GstBase # noqa: E402
+if backend.BACKEND == "gst":
+ import gi
-from base_aggregator import BaseAggregator # noqa: E402
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ from gi.repository import Gst, GstBase # noqa: E402
sys.stdout.reconfigure(encoding="utf-8")
sys.stderr.reconfigure(encoding="utf-8")
STT_SAMPLE_RATE = 16000 # Target sample rate for processing
-ICAPS = Gst.Caps(
- Gst.Structure(
- "audio/x-raw",
- format="S16LE",
- layout="interleaved",
- rate=STT_SAMPLE_RATE,
- channels=1,
- )
-)
-
-OCAPS = Gst.Caps(Gst.Structure("text/x-raw", format="utf8"))
-
class BaseTranscribe(BaseAggregator):
__gstmetadata__ = (
@@ -55,22 +45,29 @@ class BaseTranscribe(BaseAggregator):
"Aaron Boxer ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new_with_gtype(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.REQUEST,
- ICAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- Gst.PadTemplate.new_with_gtype(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- OCAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- )
+ # the caps each pad negotiates, stated once for both backends. The rate has
+ # to match STT_SAMPLE_RATE, which the VAD chunking works from.
+ INPUT_CAPS = "audio/x-raw,format=S16LE,layout=interleaved,rate=16000,channels=1"
+ OUTPUT_CAPS = "text/x-raw,format=utf8"
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new_with_gtype(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.REQUEST,
+ Gst.Caps.from_string(INPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
+ Gst.PadTemplate.new_with_gtype(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string(OUTPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
+ )
def __init__(self):
super().__init__()
@@ -132,101 +129,62 @@ def do_transcribe(self, audio_data, task):
pass
def do_process_text(self, transcript):
- # Encode the transcript as UTF-8
- text_bytes = transcript.encode("utf-8")
- encoded_size = len(text_bytes)
-
- # Create a new buffer for output and write the transcription
- outbuf = Gst.Buffer.new_allocate(None, encoded_size, None)
- outbuf.fill(0, text_bytes)
+ """The payload one transcript becomes. `None` sends nothing."""
+ return transcript.encode("utf-8")
- return outbuf
+ def process_payload(self, payload: bytes) -> list[bytes]:
+ """Runs VAD over the audio, transcribing each clip that ends.
- def do_process(self, buf):
+ Speech is accumulated across buffers, so most buffers produce nothing.
+ """
import numpy as np
- self.push_segment_if_needed()
- """Process audio data from the input buffers using VAD and Whisper."""
- audio_collected = False
+ audio_data = np.frombuffer(payload, dtype=np.int16)
- try:
- # Map the buffer to access the audio data
- success, map_info = buf.map(Gst.MapFlags.READ)
- if not success:
- self.logger.error("Failed to map input buffer")
- buf.unmap(map_info)
- return Gst.FlowReturn.OK
-
- # Convert buffer to numpy array (int16)
- audio_data = np.frombuffer(map_info.data, dtype=np.int16)
- audio_collected = True
-
- if len(audio_data) < self._vad_chunk_size:
- self.logger.warning("Insufficient audio data for processing")
- buf.unmap(map_info)
- return Gst.FlowReturn.OK
-
- # Process audio data with VAD (Voice Activity Detection)
- while len(audio_data) >= self._vad_chunk_size:
- vad_chunk = audio_data[: self._vad_chunk_size]
- audio_data = audio_data[self._vad_chunk_size :]
-
- vad_confidence = self._vad.process_chunk(vad_chunk.tobytes())
- if vad_confidence >= 0.7:
- if self.streaming:
- transcript = self._transcribe_audio(vad_chunk)
- if transcript is None:
- self.logger.warning("Empty transcript")
- buf.unmap(map_info)
- return Gst.FlowReturn.ERROR
-
- self._process_and_send(transcript, buf)
- else:
- # VAD detects voice activity, add to buffer
- self.active_clip = True
- self.silence_counter = 0
- self.clip_buffer.extend(vad_chunk)
- else:
- # Increment silence counter when no voice is detected
- self.silence_counter += 1
-
- # If silence is detected for too long, end the current segment
- if (
- self.active_clip
- and self.silence_counter > self.clip_silence_trigger_counter
- ):
- self.active_clip = False
- if not self.streaming:
- # Perform transcription in batch mode
- transcript = self._transcribe_audio(self.clip_buffer)
- if transcript is None:
- self.logger.warning("Empty transcript")
- buf.unmap(map_info)
- return Gst.FlowReturn.ERROR
-
- self._process_and_send(transcript, buf)
- self.clip_buffer.clear() # Clear the buffer for the next speech
- buf.unmap(map_info)
- except Exception as e:
- self.logger.error(f"Error during buffer processing: {e}")
+ if len(audio_data) < self._vad_chunk_size:
+ self.logger.warning("Insufficient audio data for processing")
+ return []
- if not audio_collected:
- self.logger.warning("No audio data collected from sink pads.")
- return Gst.FlowReturn.ERROR
+ payloads = []
+ while len(audio_data) >= self._vad_chunk_size:
+ vad_chunk = audio_data[: self._vad_chunk_size]
+ audio_data = audio_data[self._vad_chunk_size :]
- return Gst.FlowReturn.OK
-
- def _process_and_send(self, transcript, inbuf):
- outbuf = self.do_process_text(transcript)
- if outbuf is None:
+ vad_confidence = self._vad.process_chunk(vad_chunk.tobytes())
+ if vad_confidence >= 0.7:
+ if self.streaming:
+ self._collect_transcript(payloads, vad_chunk)
+ else:
+ # VAD detects voice activity, add to buffer
+ self.active_clip = True
+ self.silence_counter = 0
+ self.clip_buffer.extend(vad_chunk)
+ else:
+ # Increment silence counter when no voice is detected
+ self.silence_counter += 1
+
+ # If silence is detected for too long, end the current segment
+ if (
+ self.active_clip
+ and self.silence_counter > self.clip_silence_trigger_counter
+ ):
+ self.active_clip = False
+ if not self.streaming:
+ # Perform transcription in batch mode
+ self._collect_transcript(payloads, self.clip_buffer)
+ self.clip_buffer.clear() # Clear the buffer for the next speech
+
+ return payloads
+
+ def _collect_transcript(self, payloads, chunk):
+ transcript = self._transcribe_audio(chunk)
+ if transcript is None:
+ self.logger.warning("Empty transcript")
return
- # Set PTS and duration from the input buffer
- outbuf.pts = inbuf.pts
- outbuf.duration = inbuf.duration
-
- # Push the transcription downstream
- self.finish_buffer(outbuf)
+ payload = self.do_process_text(transcript)
+ if payload is not None:
+ payloads.append(payload)
def _transcribe_audio(self, chunk):
"""
diff --git a/plugins/python/base_transform.py b/plugins/python/base_transform.py
index 1c461e5..e554cca 100644
--- a/plugins/python/base_transform.py
+++ b/plugins/python/base_transform.py
@@ -16,191 +16,11 @@
# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
# Boston, MA 02110-1301, USA.
-import gi
-from engine.engine_manager import EngineManager
+"""Compatibility shim. `BaseTransform` now lives behind the pluggable backend
+in `backend/` (see `backend/__init__.py`). Leaf plugins keep importing it from
+here; the active backend is chosen by the PYML_BACKEND environment variable.
+"""
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-from gi.repository import GObject, GstBase # noqa: E402
+from backend import BaseTransform
-from log.logger_factory import LoggerFactory # noqa: E402
-
-
-class BaseTransform(GstBase.BaseTransform):
- """
- Base class for GStreamer transform elements that perform
- inference with a machine learning model. This class manages shared properties
- and handles model loading and device management via MLEngine.
- """
-
- __gstmetadata__ = (
- "BaseTransform",
- "Transform",
- "Generic machine learning model transform element",
- "Aaron Boxer ",
- )
-
- def __init__(self):
- super().__init__()
- self.logger = LoggerFactory.get(LoggerFactory.LOGGER_TYPE_GST)
- self.mgr = EngineManager(self.logger)
- self.kwargs = {}
- self.__batch_size = 1
- self.__frame_stride = 1
- self.__model_name = None
- self.__device_queue_id = 0
- self.__system_prompt = None
- self.__prompt = None
- self.__compile = False
-
- @property
- def engine(self):
- return self.mgr.engine
-
- @GObject.Property(type=str)
- def device(self):
- "Device to run the inference on (cpu, cuda, cuda:0, cuda:1, etc.)"
- return self.mgr.device
-
- @device.setter
- def device(self, value):
- self.mgr.set_device(value)
- # todo why is this needed, for example for yolo ?
- if self.engine_name:
- self.initialize_engine()
-
- @GObject.Property(type=int, default=1)
- def batch_size(self):
- "Number of items to process in a batch"
- return self.__batch_size
-
- @batch_size.setter
- def batch_size(self, value):
- self.__batch_size = value
- if self.engine:
- self.engine.batch_size = value
-
- @GObject.Property(type=int, default=1)
- def frame_stride(self):
- "How often to process a frame"
- return self.__frame_stride
-
- @frame_stride.setter
- def frame_stride(self, value):
- self.__frame_stride = value
- if self.engine:
- self.engine.frame_stride = value
-
- @GObject.Property(type=str)
- def model_name(self):
- "Name of the pre-trained model or local model path"
- return self.__model_name
-
- @model_name.setter
- def model_name(self, value):
- self.__model_name = value
-
- @GObject.Property(type=str)
- def engine_name(self):
- "Machine Learning Engine to use : pytorch, tflite, tensorflow, onnx, openvino, tvm, tinygrad, mlx, executorch, llamacpp, candle, jax, or custom engine name"
- return self.mgr.engine_name
-
- @engine_name.setter
- def engine_name(self, value):
- self.mgr.engine_name = value
-
- @GObject.Property(type=str, default="auto")
- def input_format(self):
- "Input tensor layout: auto, nhwc, or nchw"
- if self.engine:
- return self.engine.input_format
- return "auto"
-
- @input_format.setter
- def input_format(self, value):
- if self.engine:
- self.engine.input_format = value
-
- @GObject.Property(type=str, default="auto")
- def post_process(self):
- "Post-processing format for raw engine output (auto, none, or a key from detection_decoder)"
- if self.engine:
- return self.engine.post_process
- return "none"
-
- @post_process.setter
- def post_process(self, value):
- if self.engine:
- self.engine.post_process = value
-
- @GObject.Property(type=int, default=1)
- def device_queue_id(self):
- "ID of the DeviceQueue from the pool to use"
- return self.__device_queue_id
-
- @device_queue_id.setter
- def device_queue_id(self, value):
- self.__device_queue_id = value
- if self.engine:
- self.engine.device_queue_id = value
-
- @GObject.Property(type=bool, default=False, nick="compile")
- def compile(self):
- "Enable torch.compile optimization for the model"
- return self.__compile
-
- @compile.setter
- def compile(self, value):
- self.__compile = value
- if value:
- self.kwargs["compile"] = True
- else:
- self.kwargs.pop("compile", None)
-
- def do_start(self):
- self.do_load_model()
- return True
-
- def initialize_engine(self):
- if not self.engine and self.mgr.engine_name:
- self.mgr.initialize_engine()
- self.engine.batch_size = self.__batch_size
- self.engine.frame_stride = self.__frame_stride
- if self.__device_queue_id:
- self.engine.device_queue_id = self.__device_queue_id
- if not self.engine:
- self.logger.error(f"Unsupported ML engine: {self.mgr.engine_name}")
-
- def do_load_model(self):
- self.initialize_engine()
- if self.engine is None:
- self.logger.error(
- f"Cannot load model {self.model_name}: engine not initialized"
- )
- return
- if self.model_name is None:
- self.logger.warning("Cannot load model as model name is not set")
- return
- self.mgr.do_load_model(self.model_name, **self.kwargs)
-
- def get_model(self):
- """Gets the model from the engine."""
- self.initialize_engine()
- if self.engine is None:
- self.logger.error(
- f"Cannot get model {self.model_name}: engine not initialized"
- )
- return None
- """Gets the model from the engine."""
- if self.engine:
- return self.engine.get_model()
- return None
-
- def set_model(self, model):
- """Sets the model in the engine."""
- self.initialize_engine()
- if self.engine is None:
- self.logger.error("Cannot load model: engine not initialized")
- return False
- self.engine.model = model
- self.logger.info("Model set successfully in the engine.")
+__all__ = ["BaseTransform"]
diff --git a/plugins/python/base_translate.py b/plugins/python/base_translate.py
index 6dec868..8ebbce9 100644
--- a/plugins/python/base_translate.py
+++ b/plugins/python/base_translate.py
@@ -17,17 +17,17 @@
# Boston, MA 02110-1301, USA.
from abc import abstractmethod
-import gi
+import backend
+from backend import GObject
from base_aggregator import BaseAggregator
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-from gi.repository import Gst, GObject, GstBase # noqa: E402
+if backend.BACKEND == "gst":
+ import gi
-# Define input and output caps for text/x-raw format
-ICAPS = Gst.Caps(Gst.Structure("text/x-raw", format="utf8"))
-OCAPS = Gst.Caps(Gst.Structure("text/x-raw", format="utf8"))
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ from gi.repository import Gst, GstBase # noqa: E402
class BaseTranslate(BaseAggregator):
@@ -38,22 +38,28 @@ class BaseTranslate(BaseAggregator):
"Aaron Boxer ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new_with_gtype(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.REQUEST,
- ICAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- Gst.PadTemplate.new_with_gtype(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- OCAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- )
+ # the caps each pad negotiates, stated once for both backends
+ INPUT_CAPS = "text/x-raw,format=utf8"
+ OUTPUT_CAPS = "text/x-raw,format=utf8"
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new_with_gtype(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.REQUEST,
+ Gst.Caps.from_string(INPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
+ Gst.PadTemplate.new_with_gtype(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string(OUTPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
+ )
def __init__(self):
super().__init__()
@@ -82,64 +88,17 @@ def target(self, value):
def do_translate_text(self, text):
pass
- def do_process(self, buf):
- """
- Processes text data from the input buffers,
- translates it, and pushes it downstream.
- """
- try:
- success, map_info = buf.map(Gst.MapFlags.READ)
- if not success:
- self.logger.error("Failed to map input buffer")
- return Gst.FlowReturn.ERROR
-
- byte_data = bytes(map_info.data)
- buf.unmap(map_info)
-
- if not byte_data:
- return Gst.FlowReturn.OK
-
- try:
- text_data = byte_data.decode("utf-8", errors="replace")
- except Exception as e:
- self.logger.error(f"Error decoding text data: {e}")
- return Gst.FlowReturn.ERROR
-
- self.logger.info(f"Translating text: {text_data}")
-
- # Translate the text using the MarianMT model
- translated_text = self.do_translate_text(text_data)
-
- if translated_text:
- self.logger.info(f"Translated text: {translated_text}")
- # Convert the translated text to a GstBuffer and push it downstream
- outbuf = self.convert_text_to_buf(translated_text, buf)
- self.finish_buffer(outbuf)
-
- return Gst.FlowReturn.OK
-
- except Exception as e:
- self.logger.error(f"Error processing text buffer: {e}")
- return Gst.FlowReturn.ERROR
-
- def convert_text_to_buf(self, translated_text, inbuf):
- """
- Converts translated text to a GstBuffer.
- """
- try:
- # Encode the translated text as UTF-8
- text_bytes = translated_text.encode("utf-8")
- encoded_size = len(text_bytes)
-
- # Create a new buffer for output and write the translated text
- outbuf = Gst.Buffer.new_allocate(None, encoded_size, None)
- outbuf.fill(0, text_bytes)
-
- # Set PTS and duration from the input buffer
- outbuf.pts = inbuf.pts
- outbuf.duration = inbuf.duration
-
- return outbuf
- except Exception as e:
- self.logger.error(f"Error converting text to buffer: {e}")
- return None
+ def process_payload(self, payload: bytes) -> list[bytes]:
+ """Decodes the input text, translates it, and returns the result."""
+ if not payload:
+ return []
+
+ text_data = payload.decode("utf-8", errors="replace")
+ self.logger.info(f"Translating text: {text_data}")
+
+ translated_text = self.do_translate_text(text_data)
+ if not translated_text:
+ return []
+
+ self.logger.info(f"Translated text: {translated_text}")
+ return [translated_text.encode("utf-8")]
diff --git a/plugins/python/base_tts.py b/plugins/python/base_tts.py
index f3ffd77..9160d1c 100644
--- a/plugins/python/base_tts.py
+++ b/plugins/python/base_tts.py
@@ -19,16 +19,21 @@
from abc import abstractmethod
import io
import asyncio
-import gi
+import backend
+from backend import GObject
from base_aggregator import BaseAggregator
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-gi.require_version("GstAudio", "1.0")
-from gi.repository import Gst, GObject, GstBase, GstAudio # noqa: E402
+if backend.BACKEND == "gst":
+ import gi
-ICAPS = Gst.Caps(Gst.Structure("text/x-raw", format="utf8"))
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ gi.require_version("GstAudio", "1.0")
+ from gi.repository import Gst, GstBase, GstAudio # noqa: E402
+
+BYTES_PER_SAMPLE = 2 # S16LE, the format every subclass produces
+NANOSECONDS_PER_SECOND = 1_000_000_000
class BaseTts(BaseAggregator):
@@ -39,15 +44,23 @@ class BaseTts(BaseAggregator):
"Aaron Boxer ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new_with_gtype(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.REQUEST,
- ICAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- )
+ PUSH_FROM_SRC_PAD = True
+
+ # the sink pad caps, stated once for both backends; each subclass declares
+ # the audio it produces
+ INPUT_CAPS = "text/x-raw,format=utf8"
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new_with_gtype(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.REQUEST,
+ Gst.Caps.from_string(INPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
+ )
language = GObject.Property(
type=str,
@@ -55,12 +68,6 @@ class BaseTts(BaseAggregator):
nick="Language",
blurb="Two-character code for the language to be used by TTS model.",
)
- streaming = GObject.Property(
- type=bool,
- default=False,
- nick="Streaming",
- blurb="Enable streaming mode for real-time audio generation.",
- )
speaker = GObject.Property(
type=str,
default="Andrew Chipper",
@@ -72,7 +79,17 @@ def __init__(self):
super().__init__()
self.segment_pushed = False
self.device = "cpu"
- self.streaming_enabled = False
+ self.__streaming = False
+
+ @GObject.Property(type=bool, default=False)
+ def streaming(self):
+ "Enable streaming mode for real-time audio generation."
+ return self.__streaming
+
+ @streaming.setter
+ def streaming(self, value):
+ self.__streaming = value
+ self.logger.info(f"Streaming mode {'enabled' if value else 'disabled'}")
@abstractmethod
def do_load_model(self):
@@ -86,27 +103,6 @@ def do_generate_speech(self, transcript):
def do_get_sample_rate(self):
pass
- def set_property(self, property_name, value):
- if property_name == "speaker":
- self.speaker = value
- elif property_name == "language":
- self.language = value
- elif property_name == "streaming":
- self.streaming_enabled = value
- self.logger.info(f"Streaming mode {'enabled' if value else 'disabled'}")
- else:
- super().set_property(property_name, value)
-
- def get_property(self, property_name):
- if property_name == "speaker":
- return self.speaker
- elif property_name == "language":
- return self.language
- elif property_name == "streaming":
- return self.streaming_enabled
- else:
- return super().get_property(property_name)
-
def do_set_caps(self, in_caps, out_caps):
self.audio_info = GstAudio.AudioInfo()
self.audio_info.set_format(
@@ -114,37 +110,26 @@ def do_set_caps(self, in_caps, out_caps):
)
return True
- def do_process(self, buf):
- try:
- success, map_info = buf.map(Gst.MapFlags.READ)
- if not success:
- self.logger.error("Failed to map input buffer")
- return Gst.FlowReturn.ERROR
-
- byte_data = bytes(map_info.data)
- if not byte_data:
- buf.unmap(map_info)
- return Gst.FlowReturn.OK
+ def process_payload(self, payload: bytes) -> list[bytes]:
+ """Speaks the input text, one payload per stretch of generated audio."""
+ if not payload:
+ return []
- try:
- byte_data = byte_data.decode("utf-8", errors="replace")
- except Exception as e:
- self.logger.error(f"Error decoding text data: {e}")
- buf.unmap(map_info)
- return Gst.FlowReturn.ERROR
+ text = payload.decode("utf-8", errors="replace")
+ self.logger.info(f"TTS: received text: {text}")
- self.logger.info(f"TTS: received text: {byte_data}")
-
- if self.streaming_enabled:
- self.convert_text_to_audio_streaming_async(byte_data)
- else:
- self.convert_text_to_audio_async(byte_data)
+ chunks = self.split_text_into_chunks(text, 20) if self.streaming else [text]
+ payloads = []
+ for chunk in chunks:
+ audio = asyncio.run(self.process_transcript(chunk))
+ if audio is not None:
+ payloads.append(audio.tobytes())
- buf.unmap(map_info)
+ return payloads
- except Exception as e:
- self.logger.error(f"Error processing text buffer: {e}")
- return Gst.FlowReturn.ERROR
+ def payload_duration_ns(self, payload_size):
+ samples = payload_size // BYTES_PER_SAMPLE
+ return int(samples / self.do_get_sample_rate() * NANOSECONDS_PER_SECOND)
async def process_transcript(self, transcript):
import soundfile as sf
@@ -164,36 +149,11 @@ async def process_transcript(self, transcript):
if sr != self.do_get_sample_rate():
raise ValueError("Sample rate mismatch in audio processing")
- self.push_audio_to_pipeline(audio_bytes)
+ return audio_bytes
except Exception as e:
self.logger.error(f"Error processing TTS: {e}")
-
- def convert_text_to_audio_async(self, text):
- asyncio.run(self.process_transcript(text))
-
- def convert_text_to_audio_streaming_async(self, text):
- """Converts the text in smaller chunks for streaming."""
- chunks = self.split_text_into_chunks(text, 20)
- for chunk in chunks:
- asyncio.run(self.process_transcript(chunk))
+ return None
def split_text_into_chunks(self, text, max_length=50):
"""Splits text into smaller chunks for streaming."""
return [text[i : i + max_length] for i in range(0, len(text), max_length)]
-
- def push_audio_to_pipeline(self, audio_data):
- try:
- duration = len(audio_data) / self.do_get_sample_rate() * Gst.SECOND
- buffer = Gst.Buffer.new_wrapped(audio_data.tobytes())
-
- buffer.pts = Gst.CLOCK_TIME_NONE
- buffer.duration = duration
-
- ret = self.srcpad.push(buffer)
- if ret != Gst.FlowReturn.OK:
- raise RuntimeError(f"Error pushing audio to pipeline: {ret}")
-
- self.logger.info("TTS: audio generated and pushed downstream successfully.")
-
- except Exception as e:
- self.logger.error(f"Error pushing audio to pipeline: {e}")
diff --git a/plugins/python/caption_phi.py b/plugins/python/caption_phi.py
index 2572cdd..5b99f2a 100644
--- a/plugins/python/caption_phi.py
+++ b/plugins/python/caption_phi.py
@@ -17,18 +17,14 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
+
+from base_caption import BaseCaption
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
-
- from gi.repository import Gst, GObject # noqa: E402
-
- from engine.pytorch_vision_engine import PyTorchVisionEngine
+ from engine.caption_phi_engine import CaptionPhiEngine
from engine.engine_factory import EngineFactory
- from base_caption import BaseCaption
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -37,58 +33,6 @@
)
-class CaptionPhiEngine(PyTorchVisionEngine):
- def do_load_model(self, model_name, **kwargs):
- """Load a Phi-3-vision model from Hugging Face."""
- import torch
- from transformers import AutoModelForCausalLM, AutoProcessor, BitsAndBytesConfig
-
- try:
- quantization_config = BitsAndBytesConfig(load_in_4bit=True)
- self.model = AutoModelForCausalLM.from_pretrained(
- model_name,
- quantization_config=quantization_config,
- device_map="auto",
- torch_dtype=torch.float16,
- trust_remote_code=True,
- _attn_implementation="flash_attention_2",
- )
- self.processor = AutoProcessor.from_pretrained(
- model_name, trust_remote_code=True
- )
- self.logger.info("Phi-3.5-vision model and processor loaded successfully.")
- self.model.eval()
-
- # Skip .to() for 4-bit models
- if not (
- hasattr(self.model, "is_loaded_in_4bit")
- and self.model.is_loaded_in_4bit
- ):
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.logger.info(f"Model moved to {self.device}")
-
- return True
-
- except Exception as e:
- self.logger.error(f"Error loading model '{model_name}': {e}")
- self.tokenizer = None
- self.model = None
- return False
-
- def _prepare_messages(self, images):
- prompt_content = (
- "\n".join([f"<|image_{i+1}|>" for i in range(len(images))])
- + f"\n{self.prompt}"
- )
- return [{"role": "user", "content": prompt_content}]
-
- def _process_inputs(self, prompt_text, images):
- return self.processor(prompt_text, images, return_tensors="pt").to(self.device)
-
- def _trim_generated_ids(self, inputs, generate_ids):
- return generate_ids[:, inputs["input_ids"].shape[1] :]
-
-
class CaptionPhi(BaseCaption):
"""
GStreamer element for captioning video frames using Phi Vision.
@@ -108,10 +52,11 @@ def __init__(self):
EngineFactory.register(self.engine_name, CaptionPhiEngine)
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(CaptionPhi, "pyml_caption_phi")
- __gstelementfactory__ = ("pyml_caption_phi", Gst.Rank.NONE, CaptionPhi)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_caption_phi", CaptionPhi, "pyml_caption_phi"
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_caption_phi' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/caption_qwen.py b/plugins/python/caption_qwen.py
index 0e528f3..5aee95b 100644
--- a/plugins/python/caption_qwen.py
+++ b/plugins/python/caption_qwen.py
@@ -17,18 +17,14 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
+
+from base_caption import BaseCaption
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
-
- from gi.repository import Gst, GObject # noqa: E402
-
- from engine.pytorch_vision_engine import PyTorchVisionEngine
+ from engine.caption_qwen_engine import CaptionQwenEngine
from engine.engine_factory import EngineFactory
- from base_caption import BaseCaption
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -37,63 +33,6 @@
)
-class CaptionQwenEngine(PyTorchVisionEngine):
- def do_load_model(self, model_name, **kwargs):
- """Load a Qwen2.5-VL model from Hugging Face."""
- import torch
- from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
-
- try:
- self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
- model_name,
- torch_dtype="auto",
- dtype=torch.float16,
- device_map="auto",
- )
- self.processor = AutoProcessor.from_pretrained(model_name)
-
- self.logger.info(f"{model_name} model and processor loaded successfully.")
- self.model.eval()
-
- # Skip .to() for quantized models
- if not (hasattr(self.model, "is_quantized") and self.model.is_quantized):
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.logger.info(f"Model moved to {self.device}")
-
- return True
-
- except Exception as e:
- self.logger.error(f"Error loading model '{model_name}': {e}")
- self.processor = None
- self.model = None
- return False
-
- def _prepare_messages(self, images):
- content = [{"type": "image", "image": img} for img in images]
- content.append({"type": "text", "text": self.prompt})
- return [{"role": "user", "content": content}]
-
- def _process_inputs(self, prompt_text, images):
- from qwen_vl_utils import process_vision_info
-
- image_inputs, video_inputs = process_vision_info(
- self._prepare_messages(images)
- ) # Note: Uses messages directly
- return self.processor(
- text=[prompt_text],
- images=image_inputs,
- videos=video_inputs,
- padding=True,
- return_tensors="pt",
- ).to(self.device)
-
- def _trim_generated_ids(self, inputs, generate_ids):
- return [
- out_ids[len(in_ids) :]
- for in_ids, out_ids in zip(inputs.input_ids, generate_ids)
- ]
-
-
class CaptionQwen(BaseCaption):
"""
GStreamer element for captioning video frames using Qwen Vision.
@@ -113,10 +52,11 @@ def __init__(self):
EngineFactory.register(self.engine_name, CaptionQwenEngine)
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(CaptionQwen, "pyml_caption_qwen")
- __gstelementfactory__ = ("pyml_caption_qwen", Gst.Rank.NONE, CaptionQwen)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_caption_qwen", CaptionQwen, "pyml_caption_qwen"
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_caption_qwen' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/clap.py b/plugins/python/clap.py
index 6edbfbd..2b2bcbe 100644
--- a/plugins/python/clap.py
+++ b/plugins/python/clap.py
@@ -17,6 +17,7 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
@@ -26,10 +27,11 @@
gi.require_version("Gst", "1.0")
gi.require_version("GstBase", "1.0")
- from gi.repository import Gst, GObject, GstBase
+ from gi.repository import Gst, GstBase
+ from backend import GObject
from log.logger_factory import LoggerFactory
- from engine.pytorch_engine import PyTorchEngine
+ from engine.clap_engine import ClapEngine
from engine.engine_factory import EngineFactory
except ImportError as e:
@@ -39,92 +41,6 @@
# Header prefix for CLAP classification metadata
CLAP_META_HEADER = b"GST-CLAP:"
-CLAP_SAMPLE_RATE = 48000
-
-
-class ClapEngine(PyTorchEngine):
- """
- PyTorch engine for CLAP audio-text contrastive inference.
-
- Uses the HuggingFace transformers ClapModel + ClapProcessor to encode
- audio waveforms and compare them against precomputed text label embeddings.
- """
-
- def __init__(self):
- super().__init__()
- self.processor = None
- self.text_embeddings = None
- self._labels = []
-
- def do_load_model(self, model_name, **kwargs):
- try:
- from transformers import ClapModel, ClapProcessor
-
- self.processor = ClapProcessor.from_pretrained(model_name)
- self.model = ClapModel.from_pretrained(model_name)
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.model.eval()
- self.logger.info(f"CLAP model '{model_name}' loaded on {self.device}")
- labels = kwargs.get("labels", [])
- if labels:
- self._precompute_text_embeddings(labels)
- except Exception as e:
- raise ValueError(f"Failed to load CLAP model '{model_name}': {e}")
-
- def _precompute_text_embeddings(self, labels):
- """Precompute and cache normalized text embeddings for the label list."""
- import torch
-
- self._labels = list(labels)
- if not self._labels or self.processor is None:
- self.text_embeddings = None
- return
- inputs = self.processor(text=self._labels, return_tensors="pt", padding=True)
- inputs = {k: v.to(self.device) for k, v in inputs.items()}
- with torch.no_grad():
- text_emb = self.model.get_text_features(**inputs)
- self.text_embeddings = text_emb / text_emb.norm(dim=-1, keepdim=True)
- self.logger.info(f"Precomputed text embeddings for {len(self._labels)} labels")
-
- def do_forward(self, audio_waveform):
- """
- Encode an audio waveform and compute similarity against text labels.
-
- Args:
- audio_waveform: numpy float32 array of audio samples (mono).
-
- Returns:
- List of (label, score) tuples sorted by descending score,
- or None on failure.
- """
- import torch
-
- if self.text_embeddings is None or len(self._labels) == 0:
- self.logger.warning("No text labels configured for CLAP inference")
- return None
-
- try:
- inputs = self.processor(
- audios=audio_waveform,
- sampling_rate=CLAP_SAMPLE_RATE,
- return_tensors="pt",
- )
- inputs = {k: v.to(self.device) for k, v in inputs.items()}
- with torch.no_grad():
- audio_emb = self.model.get_audio_features(**inputs)
- audio_emb = audio_emb / audio_emb.norm(dim=-1, keepdim=True)
- similarities = (audio_emb @ self.text_embeddings.T).squeeze(0)
- scores = similarities.cpu().numpy()
-
- results = [
- (label, float(score)) for label, score in zip(self._labels, scores)
- ]
- results.sort(key=lambda x: x[1], reverse=True)
- return results
- except Exception as e:
- self.logger.error(f"CLAP inference error: {e}")
- return None
-
class ClapTransform(GstBase.BaseTransform):
"""
@@ -154,17 +70,19 @@ class ClapTransform(GstBase.BaseTransform):
"Aaron Boxer ",
)
- AUDIO_CAPS = Gst.Caps.from_string(
- "audio/x-raw,format=F32LE,layout=interleaved,rate=48000,channels=1"
- )
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ AUDIO_CAPS = Gst.Caps.from_string(
+ "audio/x-raw,format=F32LE,layout=interleaved,rate=48000,channels=1"
+ )
- sink_template = Gst.PadTemplate.new(
- "sink", Gst.PadDirection.SINK, Gst.PadPresence.ALWAYS, AUDIO_CAPS
- )
- src_template = Gst.PadTemplate.new(
- "src", Gst.PadDirection.SRC, Gst.PadPresence.ALWAYS, AUDIO_CAPS
- )
- __gsttemplates__ = (sink_template, src_template)
+ sink_template = Gst.PadTemplate.new(
+ "sink", Gst.PadDirection.SINK, Gst.PadPresence.ALWAYS, AUDIO_CAPS
+ )
+ src_template = Gst.PadTemplate.new(
+ "src", Gst.PadDirection.SRC, Gst.PadPresence.ALWAYS, AUDIO_CAPS
+ )
+ __gsttemplates__ = (sink_template, src_template)
model_name = GObject.Property(
type=str,
@@ -279,10 +197,9 @@ def do_transform_ip(self, buf):
return Gst.FlowReturn.OK
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(ClapTransform)
- __gstelementfactory__ = ("pyml_clap", Gst.Rank.NONE, ClapTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_clap", ClapTransform)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_clap' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/classifier.py b/plugins/python/classifier.py
index d9e2538..4af23ce 100644
--- a/plugins/python/classifier.py
+++ b/plugins/python/classifier.py
@@ -17,15 +17,10 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject # noqa: E402
from base_classifier import BaseClassifier
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -50,10 +45,9 @@ def __init__(self):
super().__init__()
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(Classifier)
- __gstelementfactory__ = ("pyml_classifier", Gst.Rank.NONE, Classifier)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_classifier", Classifier)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_classifier' element will not be registered because a module is missing."
)
diff --git a/plugins/python/clip.py b/plugins/python/clip.py
index 46241be..6e8d4df 100644
--- a/plugins/python/clip.py
+++ b/plugins/python/clip.py
@@ -17,109 +17,24 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
import threading
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- gi.require_version("GstAnalytics", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject, GstAnalytics, GLib
-
from video_transform import VideoTransform
- from utils.muxed_buffer_processor import MuxedBufferProcessor
- from engine.pytorch_engine import PyTorchEngine
+ from engine.clip_engine import ClipEngine
from engine.engine_factory import EngineFactory
+ from backend import GObject
+ from tasks.clip import ClipTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
GlobalLogger().warning(f"The 'clip' element will not be available. Error {e}")
-class ClipEngine(PyTorchEngine):
- """
- PyTorch engine for CLIP and SigLIP zero-shot image classification.
-
- Works with any HuggingFace CLIP-compatible model:
- openai/clip-vit-base-patch32
- openai/clip-vit-large-patch14
- google/siglip-base-patch16-224
- google/siglip-large-patch16-384
- """
-
- def __init__(self):
- super().__init__()
- self._labels = []
-
- @property
- def clip_labels(self):
- return self._labels
-
- @clip_labels.setter
- def clip_labels(self, value):
- self._labels = value
-
- def do_load_model(self, model_name, **kwargs):
- try:
- from transformers import AutoProcessor, AutoModel
-
- self.image_processor = AutoProcessor.from_pretrained(model_name)
- self.model = AutoModel.from_pretrained(model_name)
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.model.eval()
- self.logger.info(f"CLIP model '{model_name}' loaded on {self.device}")
- except Exception as e:
- raise ValueError(f"Failed to load CLIP model '{model_name}': {e}")
-
- def do_forward(self, frame):
- """
- Run zero-shot classification.
-
- Args:
- frame: RGB numpy array [H, W, 3]
-
- Returns:
- List of (label, probability) tuples sorted by probability descending,
- or None if no labels are set.
- """
- import numpy as np
- import torch
- from PIL import Image
-
- if not self._labels:
- self.logger.warning("No labels set — set the 'labels' property")
- return None
-
- try:
- pil_img = Image.fromarray(frame.astype(np.uint8))
- inputs = self.image_processor(
- text=self._labels,
- images=pil_img,
- return_tensors="pt",
- padding=True,
- )
- inputs = {k: v.to(self.device) for k, v in inputs.items()}
-
- with torch.no_grad():
- outputs = self.model(**inputs)
-
- # logits_per_image: [1, num_labels]
- probs = outputs.logits_per_image.softmax(dim=1)[0]
- results = [(label, prob.item()) for label, prob in zip(self._labels, probs)]
- results.sort(key=lambda x: x[1], reverse=True)
- return results
-
- except Exception as e:
- self.logger.error(f"CLIP inference error: {e}")
- return None
-
-
-class CLIPTransform(VideoTransform):
+class CLIPTransform(VideoTransform, ClipTask):
"""
GStreamer element for zero-shot image classification using CLIP or SigLIP.
@@ -134,7 +49,7 @@ class CLIPTransform(VideoTransform):
google/siglip-large-patch16-384 (SigLIP large)
Example pipeline:
- gst-launch-1.0 filesrc location=data/people.mp4 ! decodebin \\
+ python pyml-launch.py filesrc location=data/people.mp4 ! decodebin \\
! videoconvert ! videoscale ! video/x-raw,width=640,height=480 \\
! pyml_clip model-name=openai/clip-vit-base-patch32 device=cuda \\
labels="person, bicycle, car, dog, cat" top-k=3 \\
@@ -212,8 +127,7 @@ def engine_name(self):
def engine_name(self, value):
raise ValueError("'engine_name' is read-only for pyml_clip")
- def do_start(self):
- result = super().do_start()
+ def on_start(self):
# Push labels into the engine after it has been initialised
if self.engine and self._labels_list:
self.engine.clip_labels = self._labels_list
@@ -223,7 +137,6 @@ def do_start(self):
target=self._inference_worker, daemon=True
)
self._infer_thread.start()
- return result
def do_stop(self):
self._running = False
@@ -248,64 +161,31 @@ def _inference_worker(self):
with self._infer_lock:
self._last_results = results
- def do_transform_ip(self, buf):
- try:
- processor = MuxedBufferProcessor(
- self.logger, self.width, self.height, 30, 1
- )
- frames, _, num_sources, _ = processor.extract_frames(buf, self.sinkpad)
- if frames is None:
- return Gst.FlowReturn.ERROR
-
- # Use first frame for classification (batch not typical for CLIP)
- frame = frames[0] if num_sources > 1 else frames
-
- if self.engine:
- self.engine.clip_labels = self._labels_list
-
- # Post frame to background thread; never block the streaming thread
- with self._infer_lock:
- self._pending_frame = frame.copy()
- self._infer_event.set()
-
- with self._infer_lock:
- results = self._last_results
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Hand the frame to the inference thread and decode the latest result."""
+ # Use first frame for classification (batch not typical for CLIP)
+ frame = frames[0] if num_sources > 1 else frames
- if results is None:
- return Gst.FlowReturn.OK
+ if self.engine:
+ self.engine.clip_labels = self._labels_list
- self._attach_metadata(buf, results)
- return Gst.FlowReturn.OK
+ # Post frame to background thread; never block the streaming thread
+ with self._infer_lock:
+ self._pending_frame = frame.copy()
+ self._infer_event.set()
- except Exception as e:
- self.logger.error(f"CLIP transform error: {e}")
- return Gst.FlowReturn.ERROR
+ with self._infer_lock:
+ results = self._last_results
- def _attach_metadata(self, buf, results):
- """Attach top-k classification results above threshold as GstAnalytics metadata."""
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
- if not meta:
- self.logger.error("Failed to add analytics relation metadata")
+ if results is None:
return
- attached = 0
- for label, prob in results:
- if attached >= self.top_k:
- break
- if prob < self.threshold:
- break
-
- qk = GLib.quark_from_string(f"clip_{label.replace(' ', '_')}")
- ret, _ = meta.add_od_mtd(qk, 0, 0, self.width, self.height, prob)
- if ret:
- attached += 1
- self.logger.info(f"CLIP: {label} = {prob:.3f}")
+ self.decode(target, results)
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(CLIPTransform)
- __gstelementfactory__ = ("pyml_clip", Gst.Rank.NONE, CLIPTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_clip", CLIPTransform)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_clip' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/coalescehistory.py b/plugins/python/coalescehistory.py
index 8d77b52..7d6245e 100644
--- a/plugins/python/coalescehistory.py
+++ b/plugins/python/coalescehistory.py
@@ -17,9 +17,10 @@
# Boston, MA 02110-1301, USA.
+import backend
+
CAN_REGISTER_ELEMENT = True
try:
-
import collections
import gi
@@ -42,20 +43,22 @@ class CoalesceHistory(Gst.Element):
"Olivier Crête ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- Gst.Caps.from_string("text/x-raw,format=utf8"),
- ),
- Gst.PadTemplate.new(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.ALWAYS,
- Gst.Caps.from_string("text/x-raw,format=utf8"),
- ),
- )
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string("text/x-raw,format=utf8"),
+ ),
+ Gst.PadTemplate.new(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string("text/x-raw,format=utf8"),
+ ),
+ )
@GObject.Property(type=GObject.TYPE_UINT)
def history_length(self):
@@ -120,6 +123,7 @@ def do_change_state(self, state_change):
return ret
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(CoalesceHistory)
- __gstelementfactory__ = ("coalescehistory", Gst.Rank.NONE, CoalesceHistory)
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "coalescehistory", CoalesceHistory
+ )
diff --git a/plugins/python/coquitts.py b/plugins/python/coquitts.py
index ef4d3bc..c504b92 100644
--- a/plugins/python/coquitts.py
+++ b/plugins/python/coquitts.py
@@ -17,14 +17,10 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- from gi.repository import Gst, GObject, GstBase # noqa: E402
from base_tts import BaseTts
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -32,17 +28,14 @@
f"The 'pyml_coquitts' element will not be available. Error: {e}"
)
-TTS_SAMPLE_RATE = 22050
+if backend.BACKEND == "gst":
+ import gi
-OCAPS = Gst.Caps(
- Gst.Structure(
- "audio/x-raw",
- format="S16LE",
- layout="interleaved",
- rate=TTS_SAMPLE_RATE,
- channels=1,
- )
-)
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ from gi.repository import Gst, GstBase # noqa: E402
+
+TTS_SAMPLE_RATE = 22050
class CoquiTTS(BaseTts):
@@ -53,15 +46,20 @@ class CoquiTTS(BaseTts):
"Aaron Boxer ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new_with_gtype(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- OCAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- )
+ # the rate has to match TTS_SAMPLE_RATE, which the element reports downstream
+ OUTPUT_CAPS = "audio/x-raw,format=S16LE,layout=interleaved,rate=22050,channels=1"
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new_with_gtype(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string(OUTPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
+ )
def do_load_model(self):
from TTS.api import TTS
@@ -85,10 +83,9 @@ def do_get_sample_rate(self):
return TTS_SAMPLE_RATE
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(CoquiTTS)
- __gstelementfactory__ = ("pyml_coquitts", Gst.Rank.NONE, CoquiTTS)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_coquitts", CoquiTTS)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_coquitts' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/demo_soccer.py b/plugins/python/demo_soccer.py
index 2fb2997..1b133ef 100644
--- a/plugins/python/demo_soccer.py
+++ b/plugins/python/demo_soccer.py
@@ -17,6 +17,7 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
@@ -25,45 +26,19 @@
gi.require_version("Gst", "1.0")
gi.require_version("GstBase", "1.0")
gi.require_version("GstVideo", "1.0")
- gi.require_version("GstAnalytics", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject, GstAnalytics, GLib # noqa: E402
+ from gi.repository import Gst # noqa: E402
+ from backend import analytics, GObject # noqa: E402
from base_objectdetector import BaseObjectDetector
import os
from collections import deque
- from engine.pytorch_engine import PyTorchEngine
from engine.engine_factory import EngineFactory
-
- BOT_OK = BYTE_OK = CFG_OK = True
- BOTSORT = BYTETracker = get_cfg = Boxes = None
-
- def _init_ultralytics():
- global BOT_OK, BYTE_OK, CFG_OK, BOTSORT, BYTETracker, get_cfg, Boxes
- try:
- from ultralytics.trackers.bot_sort import BOTSORT as _BS
-
- BOTSORT = _BS
- except Exception:
- BOT_OK = False
- try:
- from ultralytics.trackers.byte_tracker import BYTETracker as _BT
-
- BYTETracker = _BT
- except Exception:
- BYTE_OK = False
- try:
- from ultralytics.cfg import get_cfg as _gcfg
-
- get_cfg = _gcfg
- except Exception:
- CFG_OK = False
- try:
- from ultralytics.engine.results import Boxes as _Boxes
-
- Boxes = _Boxes
- except Exception:
- pass
+ from engine.yolo_advanced_engine import (
+ YoloAdvancedEngine,
+ BoTSORTWrapper,
+ ByteTrackWrapper,
+ tlbr_of,
+ )
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -108,1014 +83,6 @@ def _init_ultralytics():
}
-def eye3():
- import numpy as np
-
- return np.eye(3, dtype=np.float32)
-
-
-def estimate_global_motion(
- prev_gray,
- gray,
- gmc_mode,
- gmc_scale,
- gft_max_corners,
- gft_quality,
- gft_min_dist,
- lk_win,
- lk_levels,
- ransac_thresh,
- frame_idx,
- verbose=False,
-):
- import cv2
- import numpy as np
-
- if gmc_mode == "off":
- if verbose:
- print(f"[frame {frame_idx}] GMC OFF → I")
- return eye3()
-
- def down(img):
- if gmc_scale == 1.0:
- return img
- w = max(2, int(img.shape[1] * gmc_scale))
- h = max(2, int(img.shape[0] * gmc_scale))
- return cv2.resize(img, (w, h), interpolation=cv2.INTER_AREA)
-
- pg, cg = down(prev_gray), down(gray)
- pts_prev = cv2.goodFeaturesToTrack(
- pg,
- maxCorners=gft_max_corners,
- qualityLevel=gft_quality,
- minDistance=gft_min_dist,
- )
- if pts_prev is None or len(pts_prev) < 6:
- if verbose:
- print(f"[frame {frame_idx}] GMC: insufficient corners → I")
- return eye3()
-
- pts_curr, st, _ = cv2.calcOpticalFlowPyrLK(
- pg,
- cg,
- pts_prev,
- None,
- winSize=(lk_win, lk_win),
- maxLevel=lk_levels,
- criteria=(cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 30, 0.01),
- )
- if pts_curr is None or st is None:
- if verbose:
- print(f"[frame {frame_idx}] GMC: LK failed → I")
- return eye3()
-
- m = st.reshape(-1).astype(bool)
- if m.sum() < (4 if gmc_mode == "homography" else 3):
- if verbose:
- print(f"[frame {frame_idx}] GMC: not enough inliers → I")
- return eye3()
-
- src = pts_prev[m]
- dst = pts_curr[m]
- if gmc_scale != 1.0:
- s = 1.0 / gmc_scale
- src *= s
- dst *= s
-
- if gmc_mode == "homography":
- H, _ = cv2.findHomography(
- src,
- dst,
- cv2.RANSAC,
- ransacReprojThreshold=ransac_thresh,
- maxIters=1000,
- )
- H = H.astype(np.float32) if H is not None else eye3()
- else:
- A, _ = cv2.estimateAffine2D(
- src, dst, ransacReprojThreshold=ransac_thresh, maxIters=1000
- )
- H = np.vstack([A, [0, 0, 1]]).astype(np.float32) if A is not None else eye3()
-
- avg = float(np.mean(np.linalg.norm(dst - src, axis=1))) if len(src) > 0 else 0.0
- if verbose:
- print(
- f"[frame {frame_idx}] GMC {gmc_mode} inliers={int(m.sum())} avg_motion={avg:.2f}px"
- )
- return H
-
-
-def warp_points(points_xy, M):
- import numpy as np
-
- if not points_xy:
- return []
- P = np.c_[
- np.array(points_xy, dtype=np.float32), np.ones((len(points_xy), 1), np.float32)
- ]
- Q = (M @ P.T).T
- Q = Q[:, :2] / np.clip(Q[:, 2:3], 1e-6, None)
- return [tuple(q) for q in Q]
-
-
-def classwise_keep(result, person_thr, ball_thr):
- import numpy as np
-
- if result is None or result.boxes is None or len(result.boxes) == 0:
- return np.zeros((0, 6), np.float32), np.zeros((0, 6), np.float32)
-
- b = result.boxes
- xyxy = b.xyxy.cpu().numpy()
- conf = (
- b.conf.cpu().numpy() if b.conf is not None else np.ones((len(b),), np.float32)
- )
- cls = (
- b.cls.cpu().numpy().astype(int)
- if b.cls is not None
- else np.zeros((len(b),), np.int32)
- )
-
- keep_p = (cls == 0) & (conf >= person_thr)
- keep_b = (cls == 32) & (conf >= ball_thr)
-
- dets_p = np.c_[xyxy[keep_p], conf[keep_p], cls[keep_p]].astype(np.float32)
- dets_b = np.c_[xyxy[keep_b], conf[keep_b], cls[keep_b]].astype(np.float32)
- return dets_p, dets_b
-
-
-def dets_to_boxes(dets_xyxy_conf_cls, frame_shape):
- _init_ultralytics()
- if dets_xyxy_conf_cls is None or dets_xyxy_conf_cls.size == 0:
- import torch as _torch
-
- data = _torch.zeros((0, 6), dtype=_torch.float32)
- return Boxes(data, frame_shape)
- import torch as _torch
-
- data = _torch.from_numpy(dets_xyxy_conf_cls).to(_torch.float32)
- return Boxes(data, frame_shape)
-
-
-def clamp_imgsz_for_device(imgsz, device_str):
- if device_str in ("cpu", "auto"):
- return min(imgsz, 1280)
- return imgsz
-
-
-def expand_roi(xyxy, scale, W, H, min_side=256, max_side=1920):
- x1, y1, x2, y2 = map(float, xyxy)
- cx, cy = (x1 + x2) * 0.5, (y1 + y2) * 0.5
- w, h = (x2 - x1), (y2 - y1)
- side = max(w, h) * float(scale)
- side = max(min_side, min(side, max_side))
- x1n = max(0, int(cx - side * 0.5))
- y1n = max(0, int(cy - side * 0.5))
- x2n = min(W - 1, int(cx + side * 0.5))
- y2n = min(H - 1, int(cy + side * 0.5))
- return x1n, y1n, x2n, y2n
-
-
-def _normalize_tracker_args(args, kind="byte"):
- def has(a):
- return hasattr(args, a) and getattr(args, a) is not None
-
- def setif(a, v):
- setattr(args, a, v)
-
- def copy_if_missing(target, *sources, default=None):
- if not has(target):
- for s in sources:
- if has(s):
- setif(target, getattr(args, s))
- return
- if default is not None:
- setif(target, default)
-
- copy_if_missing("track_high_thresh", "track_thresh", default=0.5)
- copy_if_missing("track_thresh", "track_high_thresh", default=0.5)
- copy_if_missing("track_low_thresh", default=0.1)
- copy_if_missing("new_track_thresh", default=0.4)
- copy_if_missing("match_thresh", default=0.8)
- copy_if_missing("asso_thresh", "match_thresh", default=0.8)
- copy_if_missing("track_buffer", default=30)
- copy_if_missing("frame_rate", default=30)
- copy_if_missing("fuse_score", default=True)
- copy_if_missing("fuse_score_coef", default=1.0)
- copy_if_missing("mot20", default=False)
-
- if kind == "botsort":
- copy_if_missing("with_reid", default=True)
- copy_if_missing("proximity_thresh", default=0.5)
- copy_if_missing("appearance_thresh", default=0.25)
- if has("cmc_method") and not has("gmc_method"):
- setif("gmc_method", getattr(args, "cmc_method"))
- copy_if_missing("gmc_method", default="sparseOptFlow")
-
-
-class ByteTrackWrapper:
- """Ultralytics BYTETracker; load params from YAML using get_cfg, accept Boxes."""
-
- def __init__(self, yaml_path, frame_rate):
- if not BYTE_OK:
- raise RuntimeError("Ultralytics BYTETracker not available")
- if not CFG_OK:
- raise RuntimeError(
- "Ultralytics get_cfg not available; update ultralytics package."
- )
- args = get_cfg(yaml_path)
- args.frame_rate = int(frame_rate)
- _normalize_tracker_args(args, kind="byte")
- self.tracker = BYTETracker(args, frame_rate=int(frame_rate))
-
- def update(self, boxes: Boxes, frame):
- return self.tracker.update(boxes, frame)
-
-
-class BoTSORTWrapper:
- """Ultralytics BOTSORT; load params via get_cfg, accept Boxes, safe ReID encoder."""
-
- def __init__(self, yaml_path, frame_rate, enable_reid=True):
- if not BOT_OK:
- raise RuntimeError("Ultralytics BOTSORT not available")
- if not CFG_OK:
- raise RuntimeError(
- "Ultralytics get_cfg not available; update ultralytics package."
- )
- args = get_cfg(yaml_path)
- args.frame_rate = int(frame_rate)
- args.with_reid = bool(enable_reid)
- _normalize_tracker_args(args, kind="botsort")
- self.tracker = BOTSORT(args, frame_rate=int(frame_rate))
- self._install_safe_encoder()
-
- def _install_safe_encoder(self):
- import cv2
- import numpy as np
- import torch as _torch
-
- def _safe_hsv_hist(img_bgr, bboxes):
- feats = []
- if img_bgr is None or bboxes is None:
- return feats
-
- if hasattr(bboxes, "detach"):
- bb = bboxes.detach().cpu().numpy()
- elif isinstance(bboxes, np.ndarray):
- bb = bboxes
- else:
- try:
- bb = np.asarray(bboxes, dtype=np.float32)
- except Exception:
- bb = None
-
- H, W = img_bgr.shape[:2]
-
- def process_one(b):
- b = np.asarray(b, dtype=np.float32).reshape(-1)
- x1, y1, x2, y2 = map(float, b[:4])
- x1i = max(0, min(int(x1), W - 1))
- y1i = max(0, min(int(y1), H - 1))
- x2i = max(0, min(int(x2), W - 1))
- y2i = max(0, min(int(y2), H - 1))
- if x2i <= x1i or y2i <= y1i:
- return _torch.zeros(512, dtype=_torch.float32)
- crop = img_bgr[y1i:y2i, x1i:x2i]
- if crop.size == 0:
- return _torch.zeros(512, dtype=_torch.float32)
- hsv = cv2.cvtColor(crop, cv2.COLOR_BGR2HSV)
- hist = cv2.calcHist(
- [hsv], [0, 1, 2], None, [8, 8, 8], [0, 180, 0, 256, 0, 256]
- ).flatten()
- norm = np.linalg.norm(hist) + 1e-6
- hist = (hist / norm).astype(np.float32)
- return _torch.from_numpy(hist)
-
- if bb is not None:
- for b in bb:
- feats.append(process_one(b))
- else:
- for b in bboxes:
- feats.append(process_one(b))
- return feats
-
- self.tracker.encoder = lambda img, tlbrs: _safe_hsv_hist(img, tlbrs)
-
- def update(self, boxes: Boxes, frame):
- return self.tracker.update(boxes, frame)
-
-
-def _callable_or_attr(obj, name):
- v = getattr(obj, name, None)
- if v is None:
- return None
- return v() if callable(v) else v
-
-
-def tlbr_of(tr):
- a = _callable_or_attr(tr, "tlbr")
- if a is not None:
- a = a.tolist() if hasattr(a, "tolist") else a
- if len(a) == 4:
- return a
- a = _callable_or_attr(tr, "tlwh")
- if a is not None:
- a = a.tolist() if hasattr(a, "tolist") else a
- if len(a) == 4:
- x, y, w, h = a
- return [x, y, x + w, y + h]
- if hasattr(tr, "bbox"):
- b = tr.bbox
- return b.tolist() if hasattr(b, "tolist") else list(b)
- return None
-
-
-class BallState:
- def __init__(self):
- self.cx = None
- self.cy = None
- self.vx = 0.0
- self.vy = 0.0
- self.frame = -1
-
- def predict(self, frame_idx, decay=0.85):
- if self.cx is None:
- return None
- return (self.cx + decay * self.vx, self.cy + decay * self.vy)
-
- def update_from_xyxy(self, xyxy, frame_idx):
- x1, y1, x2, y2 = map(float, xyxy)
- cx, cy = 0.5 * (x1 + x2), 0.5 * (y1 + y2)
- if self.cx is not None and self.frame >= 0:
- dt = max(1, frame_idx - self.frame)
- self.vx = (cx - self.cx) / dt
- self.vy = (cy - self.cy) / dt
- self.cx, self.cy, self.frame = cx, cy, frame_idx
-
- def update_from_center(self, cx, cy, frame_idx):
- if self.cx is not None and self.frame >= 0:
- dt = max(1, frame_idx - self.frame)
- self.vx = (cx - self.cx) / dt
- self.vy = (cy - self.cy) / dt
- self.cx, self.cy, self.frame = float(cx), float(cy), int(frame_idx)
-
-
-def _aspect_round_penalty(w, h):
- import numpy as np
-
- ar = w / max(h, 1e-6)
- roundness = np.exp(-((ar - 1.0) ** 2) / 0.15)
- return 1.0 - float(roundness)
-
-
-def _size_penalty(w, h, H, W):
- import numpy as np
-
- s = max(w, h)
- tgt = 0.03 * min(H, W)
- return float(np.clip(abs(s - tgt) / (tgt + 1e-6), 0.0, 2.0)) * 0.5
-
-
-def select_best_ball(
- dets_b,
- frame_shape,
- ball_state,
- frame_idx,
- w_conf=1.0,
- w_dist=0.015,
- w_size=0.5,
- w_round=0.4,
-):
- import numpy as np
-
- if dets_b is None or len(dets_b) == 0:
- return None
- H, W = frame_shape
- pred = ball_state.predict(frame_idx)
- scores = []
- for d in dets_b:
- x1, y1, x2, y2, conf, _ = d
- cx, cy = 0.5 * (x1 + x2), 0.5 * (y1 + y2)
- w, h = (x2 - x1), (y2 - y1)
- s_conf = float(conf)
- if pred is None:
- dist_pen = 0.0
- else:
- px, py = pred
- dist = np.hypot(cx - px, cy - py)
- dist_pen = float(dist / (0.5 * (H + W)))
- size_pen = _size_penalty(w, h, H, W)
- round_pen = _aspect_round_penalty(w, h)
- s = (
- w_conf * s_conf
- - w_dist * dist_pen
- - w_size * size_pen
- - w_round * round_pen
- )
- scores.append(s)
- if not scores:
- return None
- return int(np.argmax(scores))
-
-
-def nms_class(dets, iou_thr=0.5):
- import numpy as np
-
- if dets is None or len(dets) == 0:
- return dets
- boxes = dets[:, :4].copy()
- scores = dets[:, 4].copy()
- order = scores.argsort()[::-1]
- keep = []
-
- def iou(a, b):
- xx1 = np.maximum(a[0], b[0])
- yy1 = np.maximum(a[1], b[1])
- xx2 = np.minimum(a[2], b[2])
- yy2 = np.minimum(a[3], b[3])
- w = np.maximum(0.0, xx2 - xx1)
- h = np.maximum(0.0, yy2 - yy1)
- inter = w * h
- area_a = (a[2] - a[0]) * (a[3] - a[1])
- area_b = (b[2] - b[0]) * (b[3] - b[1])
- return inter / (area_a + area_b - inter + 1e-6)
-
- while order.size > 0:
- i = order[0]
- keep.append(i)
- if order.size == 1:
- break
- ious = np.array([iou(boxes[i], boxes[j]) for j in order[1:]])
- remain = np.where(ious <= iou_thr)[0]
- order = order[remain + 1]
- return dets[keep]
-
-
-def add_trail_point(seq: deque, x: int, y: int, k: int, densify=True, max_gap=5):
- if len(seq) > 0 and densify:
- _, _, k_prev = seq[-1]
- gap = k - k_prev
- if 1 < gap <= max_gap:
- x_prev, y_prev, _ = seq[-1]
- for t in range(1, gap):
- alpha = t / gap
- xi = int(round((1 - alpha) * x_prev + alpha * x))
- yi = int(round((1 - alpha) * y_prev + alpha * y))
- seq.append((xi, yi, k_prev + t))
- seq.append((int(x), int(y), int(k)))
-
-
-def iou_xyxy(a, b):
- if a is None or b is None:
- return 0.0
- ax1, ay1, ax2, ay2 = a
- bx1, by1, bx2, by2 = b
- xx1 = max(ax1, bx1)
- yy1 = max(ay1, by1)
- xx2 = min(ax2, bx2)
- yy2 = min(ay2, by2)
- w = max(0.0, xx2 - xx1)
- h = max(0.0, yy2 - yy1)
- inter = w * h
- if inter <= 0:
- return 0.0
- area_a = max(0.0, (ax2 - ax1)) * max(0.0, (ay2 - ay1))
- area_b = max(0.0, (bx2 - bx1)) * max(0.0, (by2 - by1))
- denom = area_a + area_b - inter + 1e-6
- return float(inter / denom)
-
-
-def lerp(a, b, t):
- return a * (1.0 - t) + b * t
-
-
-def gate_accept(
- center,
- cand_box,
- trail_seq,
- last_shown_box,
- recent_speed,
- frame_shape,
- args,
- pred=None,
- from_track=True,
-):
- import numpy as np
-
- if center is None:
- return False
-
- H, W = frame_shape
- hard_cap = float(args.ball_max_jump_rel) * float(min(H, W))
-
- if len(trail_seq) == 0:
- dist_prev = 0.0
- else:
- x_prev, y_prev, _ = trail_seq[-1]
- dist_prev = float(np.hypot(center[0] - x_prev, center[1] - y_prev))
-
- if dist_prev > hard_cap:
- return False
-
- base_gate = max(
- float(args.ball_gate_min), float(args.ball_gate_rel) * float(min(H, W))
- )
- gate_px = base_gate if from_track else base_gate * 1.25
- pass_prev = (len(trail_seq) == 0) or (dist_prev <= gate_px)
-
- pred_ok = False
- if getattr(args, "ball_gate_use_pred", False) and pred is not None:
- d_pred = float(np.hypot(center[0] - pred[0], center[1] - pred[1]))
- pred_ok = d_pred <= gate_px * 1.25
-
- if not from_track:
- return pass_prev or pred_ok
-
- iou_ok = (last_shown_box is None) or (
- iou_xyxy(cand_box, last_shown_box) >= float(args.ball_min_iou)
- )
-
- speed_ok = True
- if recent_speed is not None and recent_speed > 0:
- speed_ok = dist_prev <= float(args.ball_speed_mult) * float(recent_speed + 1e-6)
-
- return (pass_prev and iou_ok and speed_ok) or pred_ok
-
-
-def safe_int_pair(wx, wy, W, H):
- import numpy as np
-
- if wx is None or wy is None:
- return None
- if not (np.isfinite(wx) and np.isfinite(wy)):
- return None
- try:
- xi = int(round(float(wx)))
- yi = int(round(float(wy)))
- except Exception:
- return None
- if abs(xi) > 10 * W or abs(yi) > 10 * H:
- return None
- return xi, yi
-
-
-class YoloAdvancedEngine(PyTorchEngine):
- def __init__(self, device=None, **kwargs):
- super().__init__(device=device)
- # Then set self.device_str = device if device else 'auto'
- self.device_str = device if device else "auto"
- self.det_model = None
- self.fb_model = None
- self.people_tracker = None
- self.ball_tracker = None
- self.ball_state = BallState()
- self.single_ball_trail = deque(maxlen=kwargs.get("trail", 200))
- self.cum_H_history = [eye3()]
- self.cum_H = eye3()
- self.prev_gray = None
- self.last_ball_xyxy = None
- self.last_shown_box = None
- self.recent_speed = None
- self.ema_cxcy = None
- self.coast_streak = 0
- self.det_reject_streak = 0
- self.dropped_by_gate = 0
- self.coast_used = 0
- self.frame_idx = 0
- self.frame_rate = kwargs.get("frame_rate", 30.0)
- # Set all params from kwargs
- self.device_str = kwargs.get("device", "auto")
- self.imgsz = kwargs.get("imgsz", 1280)
- self.conf = kwargs.get("conf", 0.25)
- self.iou = kwargs.get("iou", 0.45)
- self.classes = kwargs.get("classes", [0, 32])
- self.person_conf_keep = kwargs.get("person_conf_keep", 0.25)
- self.ball_conf_keep = kwargs.get("ball_conf_keep", 0.04)
- self.ball_mode = kwargs.get("ball_mode", True)
- self.hires_fallback = kwargs.get("hires_fallback", True)
- self.hires_imgsz = kwargs.get("hires_imgsz", 1536)
- self.fallback_every = kwargs.get("fallback_every", 6)
- self.fallback_tiles = kwargs.get("fallback_tiles", False)
- self.tile_size = kwargs.get("tile_size", 1280)
- self.tile_overlap = kwargs.get("tile_overlap", 256)
- self.fallback_budget_ms = kwargs.get("fallback_budget_ms", 300)
- self.ball_roi_boost = kwargs.get("ball_roi_boost", False)
- self.roi_scale = kwargs.get("roi_scale", 2.5)
- self.roi_min = kwargs.get("roi_min", 256)
- self.roi_max = kwargs.get("roi_max", 1920)
- self.tracker_people = kwargs.get("tracker_people", "botsort_people_reid.yaml")
- self.tracker_ball = kwargs.get("tracker_ball", "bytetrack_ball.yaml")
- self.people_reid = kwargs.get("people_reid", True)
- self.trail = kwargs.get("trail", 200)
- self.gmc = kwargs.get("gmc", "affine")
- self.gmc_scale = kwargs.get("gmc_scale", 0.5)
- self.gft_max_corners = kwargs.get("gft_max_corners", 400)
- self.gft_quality = kwargs.get("gft_quality", 0.01)
- self.gft_min_dist = kwargs.get("gft_min_dist", 8)
- self.lk_win = kwargs.get("lk_win", 21)
- self.lk_levels = kwargs.get("lk_levels", 3)
- self.ransac_thresh = kwargs.get("ransac_thresh", 3.0)
- self.ball_gate_rel = kwargs.get("ball_gate_rel", 0.06)
- self.ball_gate_min = kwargs.get("ball_gate_min", 12)
- self.ball_gate_use_pred = kwargs.get("ball_gate_use_pred", False)
- self.ball_min_iou = kwargs.get("ball_min_iou", 0.20)
- self.ball_max_jump_rel = kwargs.get("ball_max_jump_rel", 0.12)
- self.ball_speed_mult = kwargs.get("ball_speed_mult", 3.0)
- self.ball_smooth_ema = kwargs.get("ball_smooth_ema", 0.0)
- self.det_override_conf = kwargs.get("det_override_conf", 0.28)
- self.det_override_after = kwargs.get("det_override_after", 2)
- self.reacquire_frames = kwargs.get("reacquire_frames", 6)
- self.ball_coast = kwargs.get("ball_coast", False)
- self.coast_max = kwargs.get("coast_max", 6)
- self.coast_decay = kwargs.get("coast_decay", 0.90)
- self.verbose = kwargs.get("verbose", False)
-
- def do_load_model(self, model_name, **kwargs):
- _init_ultralytics()
- try:
- from ultralytics import YOLO
-
- # YOLO load unchanged...
- self.det_model = YOLO(f"{model_name}.pt")
- self.execute_with_stream(lambda: self.det_model.to(self.device))
- self.logger.info(
- f"YOLO primary model '{model_name}' loaded on {self.device}"
- )
-
- if self.hires_fallback:
- self.fb_model = YOLO(f"{model_name}.pt")
- self.execute_with_stream(lambda: self.fb_model.to(self.device))
-
- self.model = self.det_model # Alias for base compat
-
- # Trackers with fallback
- if self.tracker_people:
- try:
- self.people_tracker = BoTSORTWrapper(
- self.tracker_people, self.frame_rate, self.people_reid
- )
- self.logger.info(
- f"People tracker loaded from {self.tracker_people}"
- )
- except Exception as te:
- self.logger.warning(f"People tracker failed ({te}); disabling.")
- self.people_tracker = None
-
- if self.tracker_ball:
- try:
- self.ball_tracker = ByteTrackWrapper(
- self.tracker_ball, self.frame_rate
- )
- self.logger.info(f"Ball tracker loaded from {self.tracker_ball}")
- except Exception as te:
- self.logger.warning(f"Ball tracker failed ({te}); disabling.")
- self.ball_tracker = None
-
- # ... kwargs update unchanged ...
- return self.tracker_people and self.tracker_ball
-
- except Exception as e:
- self.logger.error(f"Core model load failed: {e}")
- return False # No raise—let base handle
-
- def do_forward(self, frames):
- import cv2
- import numpy as np
-
- is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
- if is_batch:
- frame_bgr = frames[0] # Assume single for stateful; extend if needed
- else:
- frame_bgr = np.array(frames, copy=True)
- gray = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2GRAY)
- Hh, Ww = frame_bgr.shape[:2]
-
- # GMC
- if self.prev_gray is not None:
- H = estimate_global_motion(
- self.prev_gray,
- gray,
- self.gmc,
- self.gmc_scale,
- self.gft_max_corners,
- self.gft_quality,
- self.gft_min_dist,
- self.lk_win,
- self.lk_levels,
- self.ransac_thresh,
- self.frame_idx,
- self.verbose,
- )
- self.cum_H = H @ self.cum_H
- self.cum_H_history.append(self.cum_H.copy())
- self.prev_gray = gray
-
- # Detection
- det_res = self.execute_with_stream(
- lambda: self.det_model.predict(
- frame_bgr,
- imgsz=self.imgsz,
- conf=self.conf,
- iou=self.iou,
- classes=self.classes,
- device=self.device_str if self.device_str != "auto" else None,
- verbose=False,
- )[0]
- )
-
- dets_p, dets_b = classwise_keep(
- det_res, self.person_conf_keep, self.ball_conf_keep
- )
-
- dets_b = nms_class(dets_b, iou_thr=0.35)
- best_idx = select_best_ball(
- dets_b, frame_bgr.shape[:2], self.ball_state, self.frame_idx
- )
- if best_idx is not None:
- dets_b = dets_b[[best_idx]]
- self.ball_state.update_from_xyxy(dets_b[0, :4], self.frame_idx)
- self.last_ball_xyxy = dets_b[0, :4].astype(int).tolist()
- else:
- dets_b = dets_b[:0]
-
- # Fallback logic
- if (
- self.ball_mode
- and self.hires_fallback
- and dets_b.shape[0] == 0
- and self.frame_idx % max(1, self.fallback_every) == 0
- ):
- clamp_imgsz = clamp_imgsz_for_device(self.hires_imgsz, self.device_str)
- collected = []
-
- if (
- self.ball_roi_boost
- and self.last_ball_xyxy is not None
- and self.fb_model is not None
- ):
- x1r, y1r, x2r, y2r = expand_roi(
- self.last_ball_xyxy,
- self.roi_scale,
- Ww,
- Hh,
- min_side=self.roi_min,
- max_side=self.roi_max,
- )
- crop = frame_bgr[y1r:y2r, x1r:x2r]
- pred = self.execute_with_stream(
- lambda: self.fb_model.predict(
- crop,
- imgsz=min(max(x2r - x1r, y2r - y1r), clamp_imgsz),
- conf=max(self.ball_conf_keep, 0.02),
- iou=max(self.iou, 0.50),
- classes=[32],
- device=self.device_str if self.device_str != "auto" else None,
- verbose=False,
- )[0]
- )
- if pred.boxes is not None and len(pred.boxes) > 0:
- b = pred.boxes
- xyxy = b.xyxy.cpu().numpy()
- conf = (
- b.conf.cpu().numpy()
- if b.conf is not None
- else np.ones((len(b),), np.float32)
- )
- cls = np.full((len(b),), 32, dtype=np.float32)
- xyxy[:, [0, 2]] += x1r
- xyxy[:, [1, 3]] += y1r
- collected.append(np.c_[xyxy, conf, cls])
-
- if self.fb_model is not None and not collected:
- pred = self.execute_with_stream(
- lambda: self.fb_model.predict(
- frame_bgr,
- imgsz=clamp_imgsz,
- conf=max(self.ball_conf_keep, 0.02),
- iou=max(self.iou, 0.50),
- classes=[32],
- device=self.device_str if self.device_str != "auto" else None,
- verbose=False,
- )[0]
- )
- if pred.boxes is not None and len(pred.boxes) > 0:
- b = pred.boxes
- xyxy = b.xyxy.cpu().numpy()
- conf = (
- b.conf.cpu().numpy()
- if b.conf is not None
- else np.ones((len(b),), np.float32)
- )
- cls = np.full((len(b),), 32, dtype=np.float32)
- collected.append(np.c_[xyxy, conf, cls])
-
- if collected:
- dets_b = np.vstack(collected).astype(np.float32)
- dets_b = nms_class(dets_b, iou_thr=0.35)
- best_idx = select_best_ball(
- dets_b, frame_bgr.shape[:2], self.ball_state, self.frame_idx
- )
- if best_idx is not None:
- dets_b = dets_b[[best_idx]]
- self.ball_state.update_from_xyxy(dets_b[0, :4], self.frame_idx)
- self.last_ball_xyxy = dets_b[0, :4].astype(int).tolist()
- else:
- dets_b = dets_b[:0]
-
- frame_shape = frame_bgr.shape[:2]
- boxes_p = dets_to_boxes(dets_p, frame_shape)
- boxes_b = dets_to_boxes(dets_b, frame_shape)
- tracks_p = (
- self.people_tracker.update(boxes_p, frame_bgr)
- if self.people_tracker
- else []
- )
- tracks_b = (
- self.ball_tracker.update(boxes_b, frame_bgr) if self.ball_tracker else []
- )
-
- # Ball candidate selection and gating
- cand_box = None
- ball_center_candidate = None
- cand_conf = None
- from_track = False
-
- if len(tracks_b) >= 1:
- tr = tracks_b[0]
- tb = tlbr_of(tr)
- if tb is not None:
- x1, y1, x2, y2 = map(int, tb)
- cand_box = [x1, y1, x2, y2]
- ball_center_candidate = ((x1 + x2) // 2, (y1 + y2) // 2)
- from_track = True
- cand_conf = None
- elif dets_b.shape[0] == 1:
- x1, y1, x2, y2 = map(int, dets_b[0, :4])
- cand_box = [x1, y1, x2, y2]
- ball_center_candidate = ((x1 + x2) // 2, (y1 + y2) // 2)
- cand_conf = float(dets_b[0, 4])
- from_track = False
-
- pred_pos = (
- self.ball_state.predict(self.frame_idx) if self.ball_gate_use_pred else None
- )
- accept = gate_accept(
- ball_center_candidate,
- cand_box,
- self.single_ball_trail,
- self.last_shown_box,
- self.recent_speed,
- frame_bgr.shape[:2],
- self, # Use self as args
- pred=pred_pos,
- from_track=from_track,
- )
-
- # Det override logic
- if (not accept) and (ball_center_candidate is not None) and (not from_track):
- self.det_reject_streak += 1
- gap_frames = (
- (self.frame_idx - self.single_ball_trail[-1][2])
- if len(self.single_ball_trail)
- else 9999
- )
- Hmin = float(min(Hh, Ww))
- base_gate = max(float(self.ball_gate_min), float(self.ball_gate_rel) * Hmin)
-
- if cand_conf is not None and cand_conf >= float(self.det_override_conf):
- x_prev, y_prev = (
- (self.single_ball_trail[-1][0], self.single_ball_trail[-1][1])
- if self.single_ball_trail
- else (ball_center_candidate[0], ball_center_candidate[1])
- )
- dist_prev = float(
- np.hypot(
- ball_center_candidate[0] - x_prev,
- ball_center_candidate[1] - y_prev,
- )
- )
- if (
- (self.det_reject_streak >= int(self.det_override_after))
- or (gap_frames >= int(self.reacquire_frames))
- or (dist_prev <= 2.5 * base_gate)
- ):
- accept = True # force accept the detection
- else:
- self.det_reject_streak = 0
-
- if accept and ball_center_candidate is not None:
- cx_raw, cy_raw = ball_center_candidate
- alpha = float(self.ball_smooth_ema)
- if 0.0 < alpha <= 1.0:
- if self.ema_cxcy is None:
- self.ema_cxcy = (float(cx_raw), float(cy_raw))
- else:
- self.ema_cxcy = (
- lerp(self.ema_cxcy[0], float(cx_raw), alpha),
- lerp(self.ema_cxcy[1], float(cy_raw), alpha),
- )
- cx, cy = int(round(self.ema_cxcy[0])), int(round(self.ema_cxcy[1]))
- else:
- cx, cy = cx_raw, cy_raw
-
- if from_track:
- x1, y1, x2, y2 = cand_box
- # No drawing here, as it's engine
-
- # Update trail
- if (
- len(self.single_ball_trail) >= 1
- and self.single_ball_trail[-1][2] <= self.frame_idx - 2
- ):
- x_prev, y_prev, k_prev = self.single_ball_trail[-1]
- if self.frame_idx - k_prev == 2:
- add_trail_point(
- self.single_ball_trail,
- (x_prev + cx) // 2,
- (y_prev + cy) // 2,
- k_prev + 1,
- densify=False,
- )
-
- add_trail_point(
- self.single_ball_trail, cx, cy, self.frame_idx, densify=True, max_gap=5
- )
-
- if from_track:
- self.ball_state.update_from_xyxy(cand_box, self.frame_idx)
- else:
- if cand_box is not None:
- self.ball_state.update_from_center(cx, cy, self.frame_idx)
-
- self.last_shown_box = (
- cand_box if cand_box is not None else self.last_shown_box
- )
-
- if len(self.single_ball_trail) >= 2:
- x0, y0, _ = self.single_ball_trail[-2]
- step = float(np.hypot(cx - x0, cy - y0))
- self.recent_speed = (
- step
- if self.recent_speed is None
- else 0.8 * self.recent_speed + 0.2 * step
- )
-
- self.coast_streak = 0
- self.det_reject_streak = 0
-
- else:
- if ball_center_candidate is not None:
- self.dropped_by_gate += 1
-
- if (
- self.ball_coast
- and len(self.single_ball_trail) > 0
- and self.coast_streak < int(self.coast_max)
- ):
- pred = self.ball_state.predict(
- self.frame_idx, decay=float(self.coast_decay)
- )
- if pred is not None and np.all(np.isfinite(pred)):
- px, py = int(round(pred[0])), int(round(pred[1]))
- x_prev, y_prev, _ = self.single_ball_trail[-1]
- hard_cap = float(self.ball_max_jump_rel) * float(min(Hh, Ww))
- if (
- 0 <= px < Ww
- and 0 <= py < Hh
- and float(np.hypot(px - x_prev, py - y_prev)) <= 1.25 * hard_cap
- ):
- add_trail_point(
- self.single_ball_trail,
- px,
- py,
- self.frame_idx,
- densify=False,
- )
- self.ball_state.update_from_center(px, py, self.frame_idx)
- if len(self.single_ball_trail) >= 2:
- step = float(np.hypot(px - x_prev, py - y_prev))
- self.recent_speed = (
- step
- if self.recent_speed is None
- else 0.8 * self.recent_speed + 0.2 * step
- )
- self.coast_streak += 1
- self.coast_used += 1
- else:
- self.coast_streak = 0
-
- self.frame_idx += 1
-
- # Return result for decode
- class AdvancedResult:
- def __init__(self, tracks_p, tracks_b, ball_trail, boxes):
- self.tracks_p = tracks_p
- self.tracks_b = tracks_b
- self.ball_trail = list(ball_trail)
- self.boxes = boxes
-
- return AdvancedResult(tracks_p, tracks_b, self.single_ball_trail, det_res.boxes)
-
-
class DemoSoccer(BaseObjectDetector):
"""
GStreamer element for advanced YOLO inference focused on person and ball tracking with fallback and gating.
@@ -1753,7 +720,7 @@ def do_decode(self, buf, result, stream_idx=0):
ball_trail = result.ball_trail
boxes = result.boxes
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
+ meta = analytics.add_relation_meta(buf)
if not meta:
self.logger.error(
f"Stream {stream_idx} - Failed to add analytics relation metadata"
@@ -1773,42 +740,38 @@ def do_decode(self, buf, result, stream_idx=0):
score = 1.0 # Track confidence
track_id = getattr(tr, "track_id", 0)
qk_string = f"stream_{stream_idx}_person_id_{track_id}"
- qk = GLib.quark_from_string(qk_string)
- ret, od_mtd = meta.add_od_mtd(
- qk,
+ od_mtd = analytics.add_object(
+ meta,
+ qk_string,
x1,
y1,
x2 - x1,
y2 - y1,
score,
)
- if not ret:
+ if od_mtd is None:
self.logger.error(
f"Stream {stream_idx} - Failed to add person detection metadata"
)
continue
self.logger.debug(
- f"Stream {stream_idx} - Added person od_mtd: id={track_id}, x1={x1}, y1={y1}, w={x2-x1}, h={y2-y1}, score={score}"
+ f"Stream {stream_idx} - Added person od_mtd: id={track_id}, x1={x1}, y1={y1}, w={x2 - x1}, h={y2 - y1}, score={score}"
)
- ret, tracking_mtd = meta.add_tracking_mtd(
- track_id, Gst.util_get_timestamp()
- )
- if not ret:
+ tracking_mtd = analytics.add_tracking(meta, track_id)
+ if tracking_mtd is None:
self.logger.error(
f"Stream {stream_idx} - Failed to add person tracking metadata"
)
continue
- ret = GstAnalytics.RelationMeta.set_relation(
- meta, GstAnalytics.RelTypes.RELATE_TO, od_mtd.id, tracking_mtd.id
- )
+ ret = analytics.relate(meta, od_mtd, tracking_mtd)
if not ret:
self.logger.error(
f"Stream {stream_idx} - Failed to relate person od and tracking metadata"
)
else:
self.logger.debug(
- f"Stream {stream_idx} - Linked person od_mtd {od_mtd.id} to tracking_mtd {tracking_mtd.id}"
+ f"Stream {stream_idx} - Linked person od_mtd {od_mtd} to tracking_mtd {tracking_mtd}"
)
# Ball tracks (unchanged)
@@ -1820,42 +783,38 @@ def do_decode(self, buf, result, stream_idx=0):
score = 1.0
track_id = getattr(tr, "track_id", 0)
qk_string = f"stream_{stream_idx}_ball_id_{track_id}"
- qk = GLib.quark_from_string(qk_string)
- ret, od_mtd = meta.add_od_mtd(
- qk,
+ od_mtd = analytics.add_object(
+ meta,
+ qk_string,
x1,
y1,
x2 - x1,
y2 - y1,
score,
)
- if not ret:
+ if od_mtd is None:
self.logger.error(
f"Stream {stream_idx} - Failed to add ball detection metadata"
)
continue
self.logger.debug(
- f"Stream {stream_idx} - Added ball od_mtd: id={track_id}, x1={x1}, y1={y1}, w={x2-x1}, h={y2-y1}, score={score}"
+ f"Stream {stream_idx} - Added ball od_mtd: id={track_id}, x1={x1}, y1={y1}, w={x2 - x1}, h={y2 - y1}, score={score}"
)
- ret, tracking_mtd = meta.add_tracking_mtd(
- track_id, Gst.util_get_timestamp()
- )
- if not ret:
+ tracking_mtd = analytics.add_tracking(meta, track_id)
+ if tracking_mtd is None:
self.logger.error(
f"Stream {stream_idx} - Failed to add ball tracking metadata"
)
continue
- ret = GstAnalytics.RelationMeta.set_relation(
- meta, GstAnalytics.RelTypes.RELATE_TO, od_mtd.id, tracking_mtd.id
- )
+ ret = analytics.relate(meta, od_mtd, tracking_mtd)
if not ret:
self.logger.error(
f"Stream {stream_idx} - Failed to relate ball od and tracking metadata"
)
else:
self.logger.debug(
- f"Stream {stream_idx} - Linked ball od_mtd {od_mtd.id} to tracking_mtd {tracking_mtd.id}"
+ f"Stream {stream_idx} - Linked ball od_mtd {od_mtd} to tracking_mtd {tracking_mtd}"
)
# Ball trail - attach as custom GstStructure meta (fixed API, uncommented)
@@ -1896,9 +855,9 @@ def do_decode(self, buf, result, stream_idx=0):
# Add any non-person/ball or raw detections if desired
pass # Optional
- attached_meta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
+ attached_meta = analytics.get_relation_meta(buf)
if attached_meta:
- count = GstAnalytics.relation_get_length(attached_meta)
+ count = analytics.relation_length(attached_meta)
self.logger.info(
f"Stream {stream_idx} - Advanced metadata attached to buffer {hex(id(buf))}: {count} relations, ball trail: {len(ball_trail)}"
)
@@ -1908,10 +867,9 @@ def do_decode(self, buf, result, stream_idx=0):
)
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(DemoSoccer)
- __gstelementfactory__ = ("demo_soccer", Gst.Rank.NONE, DemoSoccer)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("demo_soccer", DemoSoccer)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'demo_soccer' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/demucs.py b/plugins/python/demucs.py
index 7e905cd..37f119d 100644
--- a/plugins/python/demucs.py
+++ b/plugins/python/demucs.py
@@ -17,18 +17,14 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GObject", "2.0")
- from gi.repository import Gst, GObject # noqa: E402
+ from backend import GObject
from base_separate import BaseSeparate
- from engine.pytorch_engine import PyTorchEngine
+ from engine.demucs_engine import DemucsEngine
from engine.engine_factory import EngineFactory
except ImportError as e:
@@ -38,62 +34,6 @@
)
-class DemucsEngine(PyTorchEngine):
- def __init__(self):
- super().__init__()
- self.sample_rate = 0
-
- def do_load_model(self, model_name, **kwargs):
- from torchaudio.pipelines import HDEMUCS_HIGH_MUSDB_PLUS
-
- if not model_name:
- return
- self.logger.info(f"Loading Demucs model on device: {self.device}")
- bundle = (
- HDEMUCS_HIGH_MUSDB_PLUS # You can choose other bundles like DEMUCS_MUSDB
- )
- self.model = bundle.get_model()
- if hasattr(self.model, "to") and callable(getattr(self.model, "to")):
- self.model = self.model.to(self.device)
- self.sample_rate = bundle.sample_rate # 44100 Hz
- self.sources = self.model.sources # ['drums', 'bass', 'other', 'vocals']
-
- def separate_sources(
- self,
- mix,
- segment=10.0,
- overlap=0.1,
- ):
- import torch
- from torchaudio.transforms import Fade
-
- device = mix.device
- batch, channels, length = mix.shape
- chunk_len = int(self.sample_rate * segment * (1 + overlap))
- start = 0
- end = chunk_len
- overlap_frames = int(overlap * self.sample_rate)
- fade = Fade(fade_in_len=0, fade_out_len=overlap_frames, fade_shape="linear")
-
- final = torch.zeros(batch, len(self.sources), channels, length, device=device)
-
- while start < length - overlap_frames:
- chunk = mix[:, :, start:end]
- with torch.no_grad():
- out = self.model.forward(chunk)
- out = fade(out)
- final[:, :, :, start:end] += out
- if start == 0:
- fade.fade_in_len = overlap_frames
- start += int(chunk_len - overlap_frames)
- else:
- start += chunk_len
- end += chunk_len
- if end >= length:
- fade.fade_out_len = 0
- return final
-
-
class Demucs(BaseSeparate):
__gstmetadata__ = (
"Demucs",
@@ -173,10 +113,9 @@ def do_separate(self, audio_data):
return selected_resampled.cpu().numpy() # float32
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(Demucs)
- __gstelementfactory__ = ("pyml_demucs", Gst.Rank.NONE, Demucs)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_demucs", Demucs)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_demucs' element will not be registered because base_separate module is missing."
)
diff --git a/plugins/python/depth.py b/plugins/python/depth.py
index fc6a3ec..a0b93ec 100644
--- a/plugins/python/depth.py
+++ b/plugins/python/depth.py
@@ -17,23 +17,16 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import ctypes
-
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- from gi.repository import Gst, GObject
-
from video_transform import VideoTransform
from utils.format_converter import FormatConverter
- from utils.muxed_buffer_processor import MuxedBufferProcessor
- from engine.pytorch_engine import PyTorchEngine
+ from engine.depth_anything_engine import DepthAnythingEngine
from engine.engine_factory import EngineFactory
+ from backend import GObject
+ from tasks.depth import DepthTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -42,75 +35,8 @@
# Header prefix for depth map buffer metadata
DEPTH_META_HEADER = b"GST-DEPTH:"
-# cv2 colormap IDs for depth visualization
-COLORMAP_IDS = {
- "inferno": 9,
- "jet": 2,
- "viridis": 16,
- "plasma": 18,
- "magma": 13,
-}
-
-class DepthAnythingEngine(PyTorchEngine):
- """
- PyTorch engine for DepthAnything V2 monocular depth estimation.
-
- Supports HuggingFace model IDs:
- depth-anything/Depth-Anything-V2-Small-hf (fastest)
- depth-anything/Depth-Anything-V2-Base-hf
- depth-anything/Depth-Anything-V2-Large-hf (most accurate)
- """
-
- def do_load_model(self, model_name, **kwargs):
- try:
- from transformers import AutoImageProcessor, AutoModelForDepthEstimation
-
- self.image_processor = AutoImageProcessor.from_pretrained(model_name)
- self.model = AutoModelForDepthEstimation.from_pretrained(model_name)
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.model.eval()
- self.logger.info(
- f"DepthAnything model '{model_name}' loaded on {self.device}"
- )
- except Exception as e:
- raise ValueError(f"Failed to load depth model '{model_name}': {e}")
-
- def do_forward(self, frames):
- import numpy as np
- import torch
- import torch.nn.functional as F
- from PIL import Image
-
- is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
- if not is_batch:
- frames = frames[np.newaxis]
-
- results = []
- for frame in frames:
- try:
- pil_img = Image.fromarray(frame.astype(np.uint8))
- H, W = frame.shape[:2]
- inputs = self.image_processor(images=pil_img, return_tensors="pt")
- inputs = {k: v.to(self.device) for k, v in inputs.items()}
- with torch.no_grad():
- outputs = self.model(**inputs)
- # outputs.predicted_depth: [1, H', W']
- depth_up = F.interpolate(
- outputs.predicted_depth.unsqueeze(0),
- size=(H, W),
- mode="bicubic",
- align_corners=False,
- ).squeeze()
- results.append(depth_up.cpu().numpy())
- except Exception as e:
- self.logger.error(f"Depth inference error on frame: {e}")
- results.append(None)
-
- return results[0] if not is_batch else results
-
-
-class DepthTransform(VideoTransform):
+class DepthTransform(VideoTransform, DepthTask):
"""
GStreamer element for monocular depth estimation using DepthAnything V2.
@@ -132,6 +58,8 @@ class DepthTransform(VideoTransform):
pyml_depth model-name=depth-anything/Depth-Anything-V2-Small-hf frame-stride=2
"""
+ META_HEADER = DEPTH_META_HEADER
+
__gstmetadata__ = (
"Depth",
"Transform",
@@ -170,103 +98,10 @@ def engine_name(self):
def engine_name(self, value):
raise ValueError("'engine_name' is read-only for pyml_depth")
- def do_transform_ip(self, buf):
- try:
- processor = MuxedBufferProcessor(
- self.logger, self.width, self.height, 30, 1
- )
- frames, _, num_sources, fmt = processor.extract_frames(buf, self.sinkpad)
- if frames is None:
- return Gst.FlowReturn.ERROR
-
- if num_sources == 1:
- depth = self._do_forward(frames)
- if depth is None:
- return Gst.FlowReturn.ERROR
- self._apply_depth(buf, depth, fmt)
- else:
- depths = self._do_forward(frames)
- if depths:
- # For batch: apply only the first depth map (primary frame)
- self._apply_depth(buf, depths[0], fmt)
-
- return Gst.FlowReturn.OK
-
- except Exception as e:
- self.logger.error(f"Depth transform error: {e}")
- return Gst.FlowReturn.ERROR
-
- def _do_forward(self, frames):
- if self.engine:
- return self.engine.do_forward(frames)
- return None
-
- def _apply_depth(self, buf, depth_map, fmt):
- """Normalize depth, optionally visualize, then append as metadata."""
- import cv2
- import numpy as np
-
- d_min, d_max = depth_map.min(), depth_map.max()
- if d_max > d_min:
- depth_norm = ((depth_map - d_min) / (d_max - d_min) * 255).astype(np.uint8)
- else:
- depth_norm = np.zeros_like(depth_map, dtype=np.uint8)
-
- # Visualize first, before appending any read-only metadata memory.
- # (A READONLY chunk on the buffer would prevent buf.map(WRITE) from succeeding.)
- if self.visualize:
- cmap_id = COLORMAP_IDS.get(self.colormap, COLORMAP_IDS["inferno"])
- depth_bgr = cv2.applyColorMap(depth_norm, cmap_id)
- output = self._convert_bgr_to_format(depth_bgr, fmt)
- if output is not None:
- success, map_info = buf.map(Gst.MapFlags.WRITE)
- if success:
- try:
- frame_bytes = np.ascontiguousarray(output).tobytes()
- dst = (ctypes.c_char * map_info.size).from_buffer(map_info.data)
- ctypes.memmove(
- dst, frame_bytes, min(len(frame_bytes), map_info.size)
- )
- finally:
- buf.unmap(map_info)
-
- # Append uint8 depth map as a custom buffer memory chunk.
- # Use new_allocate+fill: PyGI hides the maxsize arg in new_wrapped
- # (it derives it from data length), so passing it explicitly shifts
- # all subsequent args and causes a GI assertion crash.
- depth_bytes = DEPTH_META_HEADER + depth_norm.tobytes()
- tmp = Gst.Buffer.new_allocate(None, len(depth_bytes), None)
- tmp.fill(0, depth_bytes)
- buf.append_memory(tmp.get_memory(0))
-
- @staticmethod
- def _convert_bgr_to_format(bgr, fmt):
- """Convert a BGR numpy array to the target GStreamer video format."""
- import cv2
- import numpy as np
-
- if fmt == "RGB":
- return cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
- elif fmt == "BGR":
- return bgr
- elif fmt == "RGBA":
- return cv2.cvtColor(bgr, cv2.COLOR_BGR2RGBA)
- elif fmt == "BGRA":
- return cv2.cvtColor(bgr, cv2.COLOR_BGR2BGRA)
- elif fmt == "ARGB":
- rgba = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGBA)
- return np.roll(rgba, 1, axis=-1) # RGBA -> ARGB
- elif fmt == "ABGR":
- bgra = cv2.cvtColor(bgr, cv2.COLOR_BGR2BGRA)
- return np.roll(bgra, 1, axis=-1) # BGRA -> ABGR
- else:
- return cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(DepthTransform)
- __gstelementfactory__ = ("pyml_depth", Gst.Rank.NONE, DepthTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_depth", DepthTransform)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_depth' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/embedding.py b/plugins/python/embedding.py
index 6dfabee..5acf900 100644
--- a/plugins/python/embedding.py
+++ b/plugins/python/embedding.py
@@ -17,23 +17,15 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import json
- import struct
-
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- from gi.repository import Gst, GObject
-
from video_transform import VideoTransform
- from utils.format_converter import FormatConverter
- from engine.pytorch_engine import PyTorchEngine
+ from engine.embedding_engine import EmbeddingEngine
from engine.engine_factory import EngineFactory
+ from backend import frameio, GObject
+ from tasks.embedding import EmbeddingTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -43,132 +35,7 @@
EMBEDDING_META_HEADER = b"GST-EMBEDDING:"
-class EmbeddingEngine(PyTorchEngine):
- """
- PyTorch engine for image/text embedding extraction.
-
- Supports CLIP and DINOv2 models via HuggingFace transformers:
- openai/clip-vit-large-patch14 (CLIP — image + text)
- facebook/dinov2-base (DINOv2 — image only)
- """
-
- def __init__(self):
- super().__init__()
- self.processor = None
- self.tokenizer = None
- self.output_dim = 0
- self._is_clip = False
-
- def do_load_model(self, model_name, **kwargs):
- try:
-
- if "clip" in model_name.lower():
- self._load_clip(model_name)
- elif "dino" in model_name.lower():
- self._load_dinov2(model_name)
- else:
- # Default to CLIP-style loading
- self._load_clip(model_name)
-
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.model.eval()
- self.logger.info(
- f"Embedding model '{model_name}' loaded on {self.device} "
- f"(dim={self.output_dim})"
- )
- except Exception as e:
- raise ValueError(f"Failed to load embedding model '{model_name}': {e}")
-
- def _load_clip(self, model_name):
- from transformers import CLIPModel, CLIPProcessor
-
- self.model = CLIPModel.from_pretrained(model_name)
- self.processor = CLIPProcessor.from_pretrained(model_name)
- self._is_clip = True
- # Determine output dim from config
- self.output_dim = self.model.config.projection_dim
-
- def _load_dinov2(self, model_name):
- from transformers import AutoModel, AutoImageProcessor
-
- self.model = AutoModel.from_pretrained(model_name)
- self.processor = AutoImageProcessor.from_pretrained(model_name)
- self._is_clip = False
- self.output_dim = self.model.config.hidden_size
-
- def do_forward(self, frame, normalize=True):
- """
- Extract an embedding vector from a video frame.
-
- Args:
- frame: numpy RGB array (H, W, 3).
- normalize: if True, L2-normalize the embedding.
-
- Returns:
- numpy float32 array of shape (output_dim,), or None on failure.
- """
- import numpy as np
- import torch
- from PIL import Image
-
- try:
- pil_img = Image.fromarray(frame.astype(np.uint8))
- inputs = self.processor(images=pil_img, return_tensors="pt")
- inputs = {k: v.to(self.device) for k, v in inputs.items()}
-
- with torch.no_grad():
- if self._is_clip:
- emb = self.model.get_image_features(**inputs)
- else:
- outputs = self.model(**inputs)
- # Use CLS token embedding
- emb = outputs.last_hidden_state[:, 0]
-
- emb = emb.squeeze(0).cpu().numpy().astype(np.float32)
- if normalize:
- norm = np.linalg.norm(emb)
- if norm > 0:
- emb = emb / norm
- return emb
- except Exception as e:
- self.logger.error(f"Embedding inference error: {e}")
- return None
-
- def do_text_embedding(self, text, normalize=True):
- """
- Extract a text embedding (CLIP only).
-
- Args:
- text: input string.
- normalize: if True, L2-normalize the embedding.
-
- Returns:
- numpy float32 array of shape (output_dim,), or None.
- """
- import numpy as np
- import torch
-
- if not self._is_clip:
- self.logger.warning("Text embeddings only supported for CLIP models")
- return None
-
- try:
- inputs = self.processor(text=[text], return_tensors="pt", padding=True)
- inputs = {k: v.to(self.device) for k, v in inputs.items()}
- with torch.no_grad():
- emb = self.model.get_text_features(**inputs)
- emb = emb.squeeze(0).cpu().numpy().astype(np.float32)
- if normalize:
- norm = np.linalg.norm(emb)
- if norm > 0:
- emb = emb / norm
- return emb
- except Exception as e:
- self.logger.error(f"Text embedding error: {e}")
- return None
-
-
-class EmbeddingTransform(VideoTransform):
+class EmbeddingTransform(VideoTransform, EmbeddingTask):
"""
GStreamer element for extracting frame embeddings for similarity search,
clustering, or RAG.
@@ -221,7 +88,6 @@ def __init__(self):
super().__init__()
self.mgr.engine_name = "pyml_embedding_engine"
EngineFactory.register(self.mgr.engine_name, EmbeddingEngine)
- self.format_converter = FormatConverter()
self._frame_count = 0
self._text_embedding = None
self._cached_text = None
@@ -259,80 +125,39 @@ def _update_text_embedding(self):
self._text_embedding = None
self._cached_text = None
- def do_transform_ip(self, buf):
- import numpy as np
-
- try:
- self._frame_count += 1
- if self.frame_stride > 1 and (self._frame_count % self.frame_stride) != 1:
- return Gst.FlowReturn.OK
-
- if self.engine is None:
- return Gst.FlowReturn.OK
-
- success, map_info = buf.map(Gst.MapFlags.READ)
- if not success:
- self.logger.error("Failed to map video buffer for reading")
- return Gst.FlowReturn.ERROR
-
- try:
- frame = self.format_converter.to_rgb(
- map_info.data, self.width, self.height, buf, self.sinkpad
- )
- finally:
- buf.unmap(map_info)
-
- if frame is None:
- return Gst.FlowReturn.ERROR
-
- emb = self.engine.do_forward(frame, normalize=self.normalize)
- if emb is None:
- return Gst.FlowReturn.OK
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Embed the frame and append the payload, skipping strided-out frames."""
+ self._frame_count += 1
+ if self.frame_stride > 1 and (self._frame_count % self.frame_stride) != 1:
+ return
- # Update text embedding if needed
- self._update_text_embedding()
+ if self.engine is None:
+ return
- # Build JSON header with dimension info and optional similarity score
- header = {"dim": int(emb.shape[0]), "dtype": "float32"}
- if self._text_embedding is not None:
- similarity = float(np.dot(emb, self._text_embedding))
- header["text"] = self.text
- header["similarity"] = round(similarity, 6)
+ frame = frames[0] if num_sources > 1 else frames
- header_bytes = json.dumps(header).encode("utf-8")
- header_len = struct.pack("= threshold
+
+ results.append(
+ {
+ "score": anomaly_score,
+ "is_anomaly": is_anomaly,
+ "heatmap": heatmap,
+ }
+ )
+ except Exception as e:
+ self.logger.error(f"Anomaly inference error on frame: {e}")
+ results.append(
+ {
+ "score": 0.0,
+ "is_anomaly": False,
+ "heatmap": None,
+ }
+ )
+
+ return results[0] if not is_batch else results
diff --git a/plugins/python/engine/caption_phi_engine.py b/plugins/python/engine/caption_phi_engine.py
new file mode 100644
index 0000000..819fb0e
--- /dev/null
+++ b/plugins/python/engine/caption_phi_engine.py
@@ -0,0 +1,71 @@
+# CaptionPhiEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public License
+# as published by the Free Software Foundation; either version 2
+# of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_vision_engine import PyTorchVisionEngine
+
+
+class CaptionPhiEngine(PyTorchVisionEngine):
+ def do_load_model(self, model_name, **kwargs):
+ """Load a Phi-3-vision model from Hugging Face."""
+ import torch
+ from transformers import AutoModelForCausalLM, AutoProcessor, BitsAndBytesConfig
+
+ try:
+ quantization_config = BitsAndBytesConfig(load_in_4bit=True)
+ self.model = AutoModelForCausalLM.from_pretrained(
+ model_name,
+ quantization_config=quantization_config,
+ device_map="auto",
+ torch_dtype=torch.float16,
+ trust_remote_code=True,
+ _attn_implementation="flash_attention_2",
+ )
+ self.processor = AutoProcessor.from_pretrained(
+ model_name, trust_remote_code=True
+ )
+ self.logger.info("Phi-3.5-vision model and processor loaded successfully.")
+ self.model.eval()
+
+ # Skip .to() for 4-bit models
+ if not (
+ hasattr(self.model, "is_loaded_in_4bit")
+ and self.model.is_loaded_in_4bit
+ ):
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.logger.info(f"Model moved to {self.device}")
+
+ return True
+
+ except Exception as e:
+ self.logger.error(f"Error loading model '{model_name}': {e}")
+ self.tokenizer = None
+ self.model = None
+ return False
+
+ def _prepare_messages(self, images):
+ prompt_content = (
+ "\n".join([f"<|image_{i+1}|>" for i in range(len(images))])
+ + f"\n{self.prompt}"
+ )
+ return [{"role": "user", "content": prompt_content}]
+
+ def _process_inputs(self, prompt_text, images):
+ return self.processor(prompt_text, images, return_tensors="pt").to(self.device)
+
+ def _trim_generated_ids(self, inputs, generate_ids):
+ return generate_ids[:, inputs["input_ids"].shape[1] :]
diff --git a/plugins/python/engine/caption_qwen_engine.py b/plugins/python/engine/caption_qwen_engine.py
new file mode 100644
index 0000000..7e64253
--- /dev/null
+++ b/plugins/python/engine/caption_qwen_engine.py
@@ -0,0 +1,76 @@
+# CaptionQwenEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public License
+# as published by the Free Software Foundation; either version 2
+# of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_vision_engine import PyTorchVisionEngine
+
+
+class CaptionQwenEngine(PyTorchVisionEngine):
+ def do_load_model(self, model_name, **kwargs):
+ """Load a Qwen2.5-VL model from Hugging Face."""
+ import torch
+ from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
+
+ try:
+ self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+ model_name,
+ torch_dtype="auto",
+ dtype=torch.float16,
+ device_map="auto",
+ )
+ self.processor = AutoProcessor.from_pretrained(model_name)
+
+ self.logger.info(f"{model_name} model and processor loaded successfully.")
+ self.model.eval()
+
+ # Skip .to() for quantized models
+ if not (hasattr(self.model, "is_quantized") and self.model.is_quantized):
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.logger.info(f"Model moved to {self.device}")
+
+ return True
+
+ except Exception as e:
+ self.logger.error(f"Error loading model '{model_name}': {e}")
+ self.processor = None
+ self.model = None
+ return False
+
+ def _prepare_messages(self, images):
+ content = [{"type": "image", "image": img} for img in images]
+ content.append({"type": "text", "text": self.prompt})
+ return [{"role": "user", "content": content}]
+
+ def _process_inputs(self, prompt_text, images):
+ from qwen_vl_utils import process_vision_info
+
+ image_inputs, video_inputs = process_vision_info(
+ self._prepare_messages(images)
+ ) # Note: Uses messages directly
+ return self.processor(
+ text=[prompt_text],
+ images=image_inputs,
+ videos=video_inputs,
+ padding=True,
+ return_tensors="pt",
+ ).to(self.device)
+
+ def _trim_generated_ids(self, inputs, generate_ids):
+ return [
+ out_ids[len(in_ids) :]
+ for in_ids, out_ids in zip(inputs.input_ids, generate_ids)
+ ]
diff --git a/plugins/python/engine/clap_engine.py b/plugins/python/engine/clap_engine.py
new file mode 100644
index 0000000..07204c3
--- /dev/null
+++ b/plugins/python/engine/clap_engine.py
@@ -0,0 +1,105 @@
+# ClapEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+CLAP_SAMPLE_RATE = 48000
+
+
+class ClapEngine(PyTorchEngine):
+ """
+ PyTorch engine for CLAP audio-text contrastive inference.
+
+ Uses the HuggingFace transformers ClapModel + ClapProcessor to encode
+ audio waveforms and compare them against precomputed text label embeddings.
+ """
+
+ def __init__(self):
+ super().__init__()
+ self.processor = None
+ self.text_embeddings = None
+ self._labels = []
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from transformers import ClapModel, ClapProcessor
+
+ self.processor = ClapProcessor.from_pretrained(model_name)
+ self.model = ClapModel.from_pretrained(model_name)
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.model.eval()
+ self.logger.info(f"CLAP model '{model_name}' loaded on {self.device}")
+ labels = kwargs.get("labels", [])
+ if labels:
+ self._precompute_text_embeddings(labels)
+ except Exception as e:
+ raise ValueError(f"Failed to load CLAP model '{model_name}': {e}")
+
+ def _precompute_text_embeddings(self, labels):
+ """Precompute and cache normalized text embeddings for the label list."""
+ import torch
+
+ self._labels = list(labels)
+ if not self._labels or self.processor is None:
+ self.text_embeddings = None
+ return
+ inputs = self.processor(text=self._labels, return_tensors="pt", padding=True)
+ inputs = {k: v.to(self.device) for k, v in inputs.items()}
+ with torch.no_grad():
+ text_emb = self.model.get_text_features(**inputs)
+ self.text_embeddings = text_emb / text_emb.norm(dim=-1, keepdim=True)
+ self.logger.info(f"Precomputed text embeddings for {len(self._labels)} labels")
+
+ def do_forward(self, audio_waveform):
+ """
+ Encode an audio waveform and compute similarity against text labels.
+
+ Args:
+ audio_waveform: numpy float32 array of audio samples (mono).
+
+ Returns:
+ List of (label, score) tuples sorted by descending score,
+ or None on failure.
+ """
+ import torch
+
+ if self.text_embeddings is None or len(self._labels) == 0:
+ self.logger.warning("No text labels configured for CLAP inference")
+ return None
+
+ try:
+ inputs = self.processor(
+ audios=audio_waveform,
+ sampling_rate=CLAP_SAMPLE_RATE,
+ return_tensors="pt",
+ )
+ inputs = {k: v.to(self.device) for k, v in inputs.items()}
+ with torch.no_grad():
+ audio_emb = self.model.get_audio_features(**inputs)
+ audio_emb = audio_emb / audio_emb.norm(dim=-1, keepdim=True)
+ similarities = (audio_emb @ self.text_embeddings.T).squeeze(0)
+ scores = similarities.cpu().numpy()
+
+ results = [
+ (label, float(score)) for label, score in zip(self._labels, scores)
+ ]
+ results.sort(key=lambda x: x[1], reverse=True)
+ return results
+ except Exception as e:
+ self.logger.error(f"CLAP inference error: {e}")
+ return None
diff --git a/plugins/python/engine/clip_engine.py b/plugins/python/engine/clip_engine.py
new file mode 100644
index 0000000..9081ef4
--- /dev/null
+++ b/plugins/python/engine/clip_engine.py
@@ -0,0 +1,97 @@
+# ClipEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class ClipEngine(PyTorchEngine):
+ """
+ PyTorch engine for CLIP and SigLIP zero-shot image classification.
+
+ Works with any HuggingFace CLIP-compatible model:
+ openai/clip-vit-base-patch32
+ openai/clip-vit-large-patch14
+ google/siglip-base-patch16-224
+ google/siglip-large-patch16-384
+ """
+
+ def __init__(self):
+ super().__init__()
+ self._labels = []
+
+ @property
+ def clip_labels(self):
+ return self._labels
+
+ @clip_labels.setter
+ def clip_labels(self, value):
+ self._labels = value
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from transformers import AutoProcessor, AutoModel
+
+ self.image_processor = AutoProcessor.from_pretrained(model_name)
+ self.model = AutoModel.from_pretrained(model_name)
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.model.eval()
+ self.logger.info(f"CLIP model '{model_name}' loaded on {self.device}")
+ except Exception as e:
+ raise ValueError(f"Failed to load CLIP model '{model_name}': {e}")
+
+ def do_forward(self, frame):
+ """
+ Run zero-shot classification.
+
+ Args:
+ frame: RGB numpy array [H, W, 3]
+
+ Returns:
+ List of (label, probability) tuples sorted by probability descending,
+ or None if no labels are set.
+ """
+ import numpy as np
+ import torch
+ from PIL import Image
+
+ if not self._labels:
+ self.logger.warning("No labels set — set the 'labels' property")
+ return None
+
+ try:
+ pil_img = Image.fromarray(frame.astype(np.uint8))
+ inputs = self.image_processor(
+ text=self._labels,
+ images=pil_img,
+ return_tensors="pt",
+ padding=True,
+ )
+ inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+ with torch.no_grad():
+ outputs = self.model(**inputs)
+
+ # logits_per_image: [1, num_labels]
+ probs = outputs.logits_per_image.softmax(dim=1)[0]
+ results = [(label, prob.item()) for label, prob in zip(self._labels, probs)]
+ results.sort(key=lambda x: x[1], reverse=True)
+ return results
+
+ except Exception as e:
+ self.logger.error(f"CLIP inference error: {e}")
+ return None
diff --git a/plugins/python/engine/demucs_engine.py b/plugins/python/engine/demucs_engine.py
new file mode 100644
index 0000000..ec543fc
--- /dev/null
+++ b/plugins/python/engine/demucs_engine.py
@@ -0,0 +1,75 @@
+# DemucsEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class DemucsEngine(PyTorchEngine):
+ def __init__(self):
+ super().__init__()
+ self.sample_rate = 0
+
+ def do_load_model(self, model_name, **kwargs):
+ from torchaudio.pipelines import HDEMUCS_HIGH_MUSDB_PLUS
+
+ if not model_name:
+ return
+ self.logger.info(f"Loading Demucs model on device: {self.device}")
+ bundle = (
+ HDEMUCS_HIGH_MUSDB_PLUS # You can choose other bundles like DEMUCS_MUSDB
+ )
+ self.model = bundle.get_model()
+ if hasattr(self.model, "to") and callable(getattr(self.model, "to")):
+ self.model = self.model.to(self.device)
+ self.sample_rate = bundle.sample_rate # 44100 Hz
+ self.sources = self.model.sources # ['drums', 'bass', 'other', 'vocals']
+
+ def separate_sources(
+ self,
+ mix,
+ segment=10.0,
+ overlap=0.1,
+ ):
+ import torch
+ from torchaudio.transforms import Fade
+
+ device = mix.device
+ batch, channels, length = mix.shape
+ chunk_len = int(self.sample_rate * segment * (1 + overlap))
+ start = 0
+ end = chunk_len
+ overlap_frames = int(overlap * self.sample_rate)
+ fade = Fade(fade_in_len=0, fade_out_len=overlap_frames, fade_shape="linear")
+
+ final = torch.zeros(batch, len(self.sources), channels, length, device=device)
+
+ while start < length - overlap_frames:
+ chunk = mix[:, :, start:end]
+ with torch.no_grad():
+ out = self.model.forward(chunk)
+ out = fade(out)
+ final[:, :, :, start:end] += out
+ if start == 0:
+ fade.fade_in_len = overlap_frames
+ start += int(chunk_len - overlap_frames)
+ else:
+ start += chunk_len
+ end += chunk_len
+ if end >= length:
+ fade.fade_out_len = 0
+ return final
diff --git a/plugins/python/engine/depth_anything_engine.py b/plugins/python/engine/depth_anything_engine.py
new file mode 100644
index 0000000..9de31c0
--- /dev/null
+++ b/plugins/python/engine/depth_anything_engine.py
@@ -0,0 +1,77 @@
+# DepthAnythingEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class DepthAnythingEngine(PyTorchEngine):
+ """
+ PyTorch engine for DepthAnything V2 monocular depth estimation.
+
+ Supports HuggingFace model IDs:
+ depth-anything/Depth-Anything-V2-Small-hf (fastest)
+ depth-anything/Depth-Anything-V2-Base-hf
+ depth-anything/Depth-Anything-V2-Large-hf (most accurate)
+ """
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from transformers import AutoImageProcessor, AutoModelForDepthEstimation
+
+ self.image_processor = AutoImageProcessor.from_pretrained(model_name)
+ self.model = AutoModelForDepthEstimation.from_pretrained(model_name)
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.model.eval()
+ self.logger.info(
+ f"DepthAnything model '{model_name}' loaded on {self.device}"
+ )
+ except Exception as e:
+ raise ValueError(f"Failed to load depth model '{model_name}': {e}")
+
+ def do_forward(self, frames):
+ import numpy as np
+ import torch
+ import torch.nn.functional as F
+ from PIL import Image
+
+ is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
+ if not is_batch:
+ frames = frames[np.newaxis]
+
+ results = []
+ for frame in frames:
+ try:
+ pil_img = Image.fromarray(frame.astype(np.uint8))
+ H, W = frame.shape[:2]
+ inputs = self.image_processor(images=pil_img, return_tensors="pt")
+ inputs = {k: v.to(self.device) for k, v in inputs.items()}
+ with torch.no_grad():
+ outputs = self.model(**inputs)
+ # outputs.predicted_depth: [1, H', W']
+ depth_up = F.interpolate(
+ outputs.predicted_depth.unsqueeze(0),
+ size=(H, W),
+ mode="bicubic",
+ align_corners=False,
+ ).squeeze()
+ results.append(depth_up.cpu().numpy())
+ except Exception as e:
+ self.logger.error(f"Depth inference error on frame: {e}")
+ results.append(None)
+
+ return results[0] if not is_batch else results
diff --git a/plugins/python/engine/drpai_engine.py b/plugins/python/engine/drpai_engine.py
new file mode 100644
index 0000000..485b1a1
--- /dev/null
+++ b/plugins/python/engine/drpai_engine.py
@@ -0,0 +1,145 @@
+# DRPAIEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+import os
+import numpy as np
+
+from .ml_engine import MLEngine
+
+
+def _anchor_count(imgsz):
+ """Total anchors a YOLO model emits for a square input at strides 8/16/32."""
+ return sum((imgsz // s) ** 2 for s in (8, 16, 32))
+
+
+class DRPAIEngine(MLEngine):
+ """DRP-AI TVM runtime engine for Renesas RZ/V boards (RZ/V2H).
+
+ Runs a model compiled with the Renesas DRP-AI TVM compiler on the DRP-AI
+ NPU. `model_name` is the path to the compiled deploy directory containing
+ ``deploy.so`` / ``deploy.json`` / ``deploy.params``.
+
+ Inference goes through the ``drpai_runtime`` pybind11 module (built from
+ ``extern/rzv2h/`` against the board's DRP-AI TVM runtime.
+ """
+
+ def __init__(self):
+ super().__init__()
+ self.runtime = None
+ self.model_name = None
+ self.kwargs = None
+ self.imgsz = 640
+
+ self.input_format = "nchw"
+ self.post_process = "anchor_free"
+
+ def do_load_model(self, model_name, **kwargs):
+ self.model_name = model_name
+ self.kwargs = kwargs
+ imgsz = kwargs.get("imgsz")
+ if imgsz:
+ try:
+ self.imgsz = int(imgsz)
+ except (TypeError, ValueError):
+ pass
+
+ try:
+ import drpai_runtime
+ except ImportError as e:
+ self.logger.error(
+ "drpai_runtime module not found. Build the pybind11 binding in "
+ "extern/rzv2h/ inside the RZ/V2H DRP-AI TVM SDK and put it on PYTHONPATH "
+ f"(see extern/rzv2h/README.md). Import error: {e}"
+ )
+ return False
+
+ if not os.path.isdir(model_name):
+ self.logger.error(
+ f"DRP-AI model directory not found: {model_name!r} "
+ "(expected a folder with deploy.so/json/params)"
+ )
+ return False
+
+ try:
+ self.runtime = drpai_runtime.Runtime()
+ if not self.runtime.load(model_name):
+ self.logger.error(f"DRP-AI failed to load model from {model_name}")
+ self.runtime = None
+ return False
+ self.logger.info(
+ f"DRP-AI model loaded from {model_name} (imgsz={self.imgsz})"
+ )
+ return True
+ except Exception as e:
+ self.logger.error(f"DRP-AI load error: {e}")
+ self.runtime = None
+ return False
+
+ def do_set_device(self, device):
+ self.device = device
+ self.logger.info(f"DRP-AI engine device set to {device}")
+
+ def do_generate(self, input_text, max_length=1000, system_prompt=None):
+ raise NotImplementedError(
+ "DRP-AI engine is a vision-inference engine; text generation is not "
+ "supported."
+ )
+
+ def _preprocess(self, frame_hwc):
+ """HWC uint8 RGB(A) frame -> contiguous (1, 3, H, W) float32 in [0, 1]."""
+ x = np.asarray(frame_hwc, dtype=np.float32)
+ if x.shape[-1] > 3:
+ x = x[..., :3]
+ x = x / 255.0
+ x = np.transpose(x, (2, 0, 1))
+ x = np.expand_dims(x, 0)
+ return np.ascontiguousarray(x, dtype=np.float32)
+
+ def _gather_output(self):
+ """Read output 0 and reshape the flat buffer to (1, 4+nc, anchors)."""
+ out = np.asarray(self.runtime.get_output(0), dtype=np.float32).reshape(-1)
+ anchors = _anchor_count(self.imgsz)
+ if anchors and out.size % anchors == 0:
+ channels = out.size // anchors
+ return out.reshape(1, channels, anchors)
+ self.logger.warning(
+ f"DRP-AI output size {out.size} not divisible by {anchors} anchors; "
+ "passing raw to post-process"
+ )
+ return out
+
+ def do_forward(self, frames):
+ if self.runtime is None:
+ self.logger.error("DRP-AI runtime not loaded")
+ return None
+
+ is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
+ batch = frames if is_batch else frames[np.newaxis, ...]
+
+ results = []
+ for img in batch:
+ try:
+ self.runtime.set_input(0, self._preprocess(img))
+ self.runtime.run()
+ raw = self._gather_output()
+ results.append(self._apply_post_process(raw, is_batch=False))
+ except Exception as e:
+ self.logger.error(f"DRP-AI inference error: {e}")
+ results.append(None)
+
+ return results if is_batch else results[0]
diff --git a/plugins/python/engine/embedding_engine.py b/plugins/python/engine/embedding_engine.py
new file mode 100644
index 0000000..1f45889
--- /dev/null
+++ b/plugins/python/engine/embedding_engine.py
@@ -0,0 +1,144 @@
+# EmbeddingEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class EmbeddingEngine(PyTorchEngine):
+ """
+ PyTorch engine for image/text embedding extraction.
+
+ Supports CLIP and DINOv2 models via HuggingFace transformers:
+ openai/clip-vit-large-patch14 (CLIP — image + text)
+ facebook/dinov2-base (DINOv2 — image only)
+ """
+
+ def __init__(self):
+ super().__init__()
+ self.processor = None
+ self.tokenizer = None
+ self.output_dim = 0
+ self._is_clip = False
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+
+ if "clip" in model_name.lower():
+ self._load_clip(model_name)
+ elif "dino" in model_name.lower():
+ self._load_dinov2(model_name)
+ else:
+ # Default to CLIP-style loading
+ self._load_clip(model_name)
+
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.model.eval()
+ self.logger.info(
+ f"Embedding model '{model_name}' loaded on {self.device} "
+ f"(dim={self.output_dim})"
+ )
+ except Exception as e:
+ raise ValueError(f"Failed to load embedding model '{model_name}': {e}")
+
+ def _load_clip(self, model_name):
+ from transformers import CLIPModel, CLIPProcessor
+
+ self.model = CLIPModel.from_pretrained(model_name)
+ self.processor = CLIPProcessor.from_pretrained(model_name)
+ self._is_clip = True
+ # Determine output dim from config
+ self.output_dim = self.model.config.projection_dim
+
+ def _load_dinov2(self, model_name):
+ from transformers import AutoModel, AutoImageProcessor
+
+ self.model = AutoModel.from_pretrained(model_name)
+ self.processor = AutoImageProcessor.from_pretrained(model_name)
+ self._is_clip = False
+ self.output_dim = self.model.config.hidden_size
+
+ def do_forward(self, frame, normalize=True):
+ """
+ Extract an embedding vector from a video frame.
+
+ Args:
+ frame: numpy RGB array (H, W, 3).
+ normalize: if True, L2-normalize the embedding.
+
+ Returns:
+ numpy float32 array of shape (output_dim,), or None on failure.
+ """
+ import numpy as np
+ import torch
+ from PIL import Image
+
+ try:
+ pil_img = Image.fromarray(frame.astype(np.uint8))
+ inputs = self.processor(images=pil_img, return_tensors="pt")
+ inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+ with torch.no_grad():
+ if self._is_clip:
+ emb = self.model.get_image_features(**inputs)
+ else:
+ outputs = self.model(**inputs)
+ # Use CLS token embedding
+ emb = outputs.last_hidden_state[:, 0]
+
+ emb = emb.squeeze(0).cpu().numpy().astype(np.float32)
+ if normalize:
+ norm = np.linalg.norm(emb)
+ if norm > 0:
+ emb = emb / norm
+ return emb
+ except Exception as e:
+ self.logger.error(f"Embedding inference error: {e}")
+ return None
+
+ def do_text_embedding(self, text, normalize=True):
+ """
+ Extract a text embedding (CLIP only).
+
+ Args:
+ text: input string.
+ normalize: if True, L2-normalize the embedding.
+
+ Returns:
+ numpy float32 array of shape (output_dim,), or None.
+ """
+ import numpy as np
+ import torch
+
+ if not self._is_clip:
+ self.logger.warning("Text embeddings only supported for CLIP models")
+ return None
+
+ try:
+ inputs = self.processor(text=[text], return_tensors="pt", padding=True)
+ inputs = {k: v.to(self.device) for k, v in inputs.items()}
+ with torch.no_grad():
+ emb = self.model.get_text_features(**inputs)
+ emb = emb.squeeze(0).cpu().numpy().astype(np.float32)
+ if normalize:
+ norm = np.linalg.norm(emb)
+ if norm > 0:
+ emb = emb / norm
+ return emb
+ except Exception as e:
+ self.logger.error(f"Text embedding error: {e}")
+ return None
diff --git a/plugins/python/engine/engine_factory.py b/plugins/python/engine/engine_factory.py
index 2a0e5fb..d361bdf 100644
--- a/plugins/python/engine/engine_factory.py
+++ b/plugins/python/engine/engine_factory.py
@@ -44,6 +44,7 @@ class EngineFactory:
MIGRAPHX_ENGINE = "migraphx"
IREE_ENGINE = "iree"
NCNN_ENGINE = "ncnn"
+ DRPAI_ENGINE = "drpai"
_builtins_registered: bool = False # Class-level flag for singleton-like lazy init
@@ -154,6 +155,13 @@ def _register_builtins(cls) -> None:
except ImportError:
pass
+ try:
+ from .drpai_engine import DRPAIEngine
+
+ _try_register(cls.DRPAI_ENGINE, DRPAIEngine)
+ except ImportError:
+ pass
+
@staticmethod
def register(engine_type: str, engine_class: Type) -> None:
_engine_registry[engine_type] = engine_class
diff --git a/plugins/python/engine/face_engine.py b/plugins/python/engine/face_engine.py
new file mode 100644
index 0000000..6679038
--- /dev/null
+++ b/plugins/python/engine/face_engine.py
@@ -0,0 +1,118 @@
+# FaceEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+import os
+
+from .pytorch_engine import PyTorchEngine
+
+
+class FaceEngine(PyTorchEngine):
+ """
+ PyTorch engine for face detection and recognition using InsightFace.
+
+ Supports InsightFace model packs:
+ buffalo_l (large, most accurate)
+ buffalo_s (small, fastest)
+ buffalo_sc (small with recognition)
+ """
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from insightface.app import FaceAnalysis
+
+ self.app = FaceAnalysis(
+ name=model_name,
+ providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+ )
+ self.app.prepare(ctx_id=0, det_size=(640, 640))
+ self.gallery = {}
+ self.logger.info(f"InsightFace model '{model_name}' loaded")
+ except Exception as e:
+ raise ValueError(f"Failed to load InsightFace model '{model_name}': {e}")
+
+ def load_gallery(self, gallery_path):
+ """Load known face embeddings from a directory of images."""
+ import numpy as np
+ from PIL import Image
+
+ if not gallery_path or not os.path.isdir(gallery_path):
+ return
+
+ self.gallery = {}
+ for fname in os.listdir(gallery_path):
+ fpath = os.path.join(gallery_path, fname)
+ if not os.path.isfile(fpath):
+ continue
+ try:
+ img = np.array(Image.open(fpath).convert("RGB"))
+ faces = self.app.get(img)
+ if faces:
+ name = os.path.splitext(fname)[0]
+ self.gallery[name] = faces[0].embedding
+ self.logger.info(f"Loaded gallery face: {name}")
+ except Exception as e:
+ self.logger.warning(f"Failed to load gallery image '{fname}': {e}")
+
+ def do_forward(self, frames, threshold=0.5):
+ import numpy as np
+
+ is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
+ if not is_batch:
+ frames = frames[np.newaxis]
+
+ results = []
+ for frame in frames:
+ try:
+ faces = self.app.get(frame.astype(np.uint8))
+ detections = []
+ for face in faces:
+ bbox = face.bbox.astype(float).tolist()
+ score = float(face.det_score)
+ embedding = face.embedding
+
+ identity = "unknown"
+ best_sim = 0.0
+ if self.gallery and embedding is not None:
+ for name, gallery_emb in self.gallery.items():
+ sim = float(
+ np.dot(embedding, gallery_emb)
+ / (
+ np.linalg.norm(embedding)
+ * np.linalg.norm(gallery_emb)
+ + 1e-8
+ )
+ )
+ if sim > best_sim:
+ best_sim = sim
+ if sim >= threshold:
+ identity = name
+
+ detections.append(
+ {
+ "bbox": bbox,
+ "score": score,
+ "identity": identity,
+ "similarity": best_sim,
+ }
+ )
+ results.append(detections)
+ except Exception as e:
+ self.logger.error(f"Face inference error on frame: {e}")
+ results.append([])
+
+ return results[0] if not is_batch else results
diff --git a/plugins/python/engine/ml_engine.py b/plugins/python/engine/ml_engine.py
index 3d5a74c..58f2e9d 100644
--- a/plugins/python/engine/ml_engine.py
+++ b/plugins/python/engine/ml_engine.py
@@ -96,7 +96,12 @@ def _apply_post_process(self, raw, is_batch):
if pp != "none" and not isinstance(raw, list):
from utils.detection_decoder import decode
- results = decode(raw, pp)
+ results = decode(
+ raw,
+ pp,
+ conf_threshold=getattr(self, "conf", 0.25),
+ iou_threshold=getattr(self, "iou", 0.45),
+ )
return results[0] if not is_batch else results
return raw
diff --git a/plugins/python/engine/ocr_engine.py b/plugins/python/engine/ocr_engine.py
new file mode 100644
index 0000000..432c8dc
--- /dev/null
+++ b/plugins/python/engine/ocr_engine.py
@@ -0,0 +1,93 @@
+# OcrEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class OcrEngine(PyTorchEngine):
+ """
+ PyTorch engine for TrOCR text recognition.
+
+ Supports HuggingFace model IDs:
+ microsoft/trocr-base-printed
+ microsoft/trocr-large-printed
+ microsoft/trocr-base-handwritten
+ """
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from transformers import TrOCRProcessor, VisionEncoderDecoderModel
+
+ self.processor = TrOCRProcessor.from_pretrained(model_name)
+ self.model = VisionEncoderDecoderModel.from_pretrained(model_name)
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.model.eval()
+ self.logger.info(f"TrOCR model '{model_name}' loaded on {self.device}")
+ except Exception as e:
+ raise ValueError(f"Failed to load TrOCR model '{model_name}': {e}")
+
+ def do_forward(self, frames):
+ import numpy as np
+ import torch
+ from PIL import Image
+
+ is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
+ if not is_batch:
+ frames = frames[np.newaxis]
+
+ results = []
+ for frame in frames:
+ try:
+ pil_img = Image.fromarray(frame.astype(np.uint8))
+ H, W = frame.shape[:2]
+
+ # Split frame into horizontal strips for text region detection
+ strip_height = max(H // 4, 32)
+ texts = []
+ regions = []
+ for y_start in range(0, H, strip_height):
+ y_end = min(y_start + strip_height, H)
+ strip = pil_img.crop((0, y_start, W, y_end))
+ pixel_values = self.processor(
+ images=strip, return_tensors="pt"
+ ).pixel_values.to(self.device)
+
+ with torch.no_grad():
+ generated_ids = self.model.generate(pixel_values)
+
+ text = self.processor.batch_decode(
+ generated_ids, skip_special_tokens=True
+ )[0].strip()
+ if text:
+ texts.append(text)
+ regions.append(
+ {
+ "x": 0,
+ "y": y_start,
+ "w": W,
+ "h": y_end - y_start,
+ "text": text,
+ }
+ )
+
+ results.append({"texts": texts, "regions": regions})
+ except Exception as e:
+ self.logger.error(f"OCR inference error on frame: {e}")
+ results.append({"texts": [], "regions": []})
+
+ return results[0] if not is_batch else results
diff --git a/plugins/python/engine/onnx_engine.py b/plugins/python/engine/onnx_engine.py
index ba85e84..ad3dce7 100644
--- a/plugins/python/engine/onnx_engine.py
+++ b/plugins/python/engine/onnx_engine.py
@@ -49,6 +49,62 @@ def _input_is_nchw(self):
shape = self.session.get_inputs()[0].shape
return len(shape) == 4 and shape[1] in (1, 3, 4)
+ def _model_input_hw(self):
+ """(H, W) the model's input expects, or None if dynamic/unknown."""
+ if self.session is None:
+ return None
+ shape = self.session.get_inputs()[0].shape
+ if len(shape) != 4:
+ return None
+ h, w = shape[2], shape[3]
+ if isinstance(h, int) and isinstance(w, int) and h > 0 and w > 0:
+ return (h, w)
+ return None
+
+ def _letterbox(self, frames, is_batch):
+ """Resize frame(s) to the model input size, preserving aspect ratio with
+ grey padding (YOLO-style). Returns (processed, transform); transform =
+ (ratio, pad_x, pad_y, orig_w, orig_h) maps model coords back to the
+ original frame. Returns (frames, None) when no resize is needed (already
+ model-sized, or dynamic input) -- so pre-sized callers are unaffected."""
+ import numpy as np
+ import cv2
+
+ mhw = self._model_input_hw()
+ if mhw is None:
+ return frames, None
+ mh, mw = mhw
+ imgs = frames if is_batch else frames[None]
+ h, w = int(imgs.shape[1]), int(imgs.shape[2])
+ if (h, w) == (mh, mw):
+ return frames, None
+ r = min(mh / h, mw / w)
+ nh, nw = int(round(h * r)), int(round(w * r))
+ pad_x, pad_y = (mw - nw) // 2, (mh - nh) // 2
+ out = np.full((imgs.shape[0], mh, mw, imgs.shape[3]), 114, dtype=imgs.dtype)
+ for i in range(imgs.shape[0]):
+ out[i, pad_y : pad_y + nh, pad_x : pad_x + nw] = cv2.resize(
+ imgs[i], (nw, nh), interpolation=cv2.INTER_LINEAR
+ )
+ proc = out if is_batch else out[0]
+ return proc, (r, float(pad_x), float(pad_y), w, h)
+
+ def _unletterbox(self, results, transform):
+ """Map detection boxes from model coords back to original-frame coords."""
+ import numpy as np
+
+ r, pad_x, pad_y, ow, oh = transform
+ for res in results if isinstance(results, list) else [results]:
+ if not isinstance(res, dict):
+ continue
+ b = res.get("boxes")
+ if b is None or len(b) == 0:
+ continue
+ b = np.asarray(b, dtype=np.float32).copy()
+ b[:, [0, 2]] = ((b[:, [0, 2]] - pad_x) / r).clip(0, ow)
+ b[:, [1, 3]] = ((b[:, [1, 3]] - pad_y) / r).clip(0, oh)
+ res["boxes"] = b
+
def do_load_model(self, model_name, **kwargs):
"""Load a pre-trained model by name from TorchVision, Transformers (via Optimum ONNX), or a local ONNX path."""
processor_name = kwargs.get("processor_name")
@@ -369,10 +425,21 @@ def do_forward(self, frames):
fmt = self.input_format
if fmt == "auto" and self._input_is_nchw():
self.input_format = "nchw"
- img = self._apply_input_format(frames.astype(np.float32) / 255.0, is_batch)
+ # Letterbox to the model's fixed input size for inference, keeping
+ # the transform so boxes map back to the original frame -- lets the
+ # caller feed full-res frames and overlay on them.
+ proc, transform = self._letterbox(frames, is_batch)
+ img = self._apply_input_format(proc.astype(np.float32) / 255.0, is_batch)
+ if "float16" in self.session.get_inputs()[0].type:
+ img = img.astype(np.float16)
outputs = self.session.run(self.output_names, {self.input_names[0]: img})
raw = outputs if len(outputs) > 1 else outputs[0]
- return self._apply_post_process(raw, is_batch)
+ if isinstance(raw, np.ndarray) and raw.dtype != np.float32:
+ raw = raw.astype(np.float32)
+ results = self._apply_post_process(raw, is_batch)
+ if transform is not None:
+ self._unletterbox(results, transform)
+ return results
else:
raise ValueError("Unsupported model type.")
diff --git a/plugins/python/engine/optical_flow_engine.py b/plugins/python/engine/optical_flow_engine.py
new file mode 100644
index 0000000..c90e190
--- /dev/null
+++ b/plugins/python/engine/optical_flow_engine.py
@@ -0,0 +1,86 @@
+# OpticalFlowEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class OpticalFlowEngine(PyTorchEngine):
+ """
+ PyTorch engine for dense optical flow estimation using RAFT.
+
+ Supports torchvision RAFT model variants:
+ raft_large (most accurate)
+ raft_small (fastest)
+ """
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from torchvision.models.optical_flow import (
+ raft_large,
+ raft_small,
+ Raft_Large_Weights,
+ Raft_Small_Weights,
+ )
+
+ if model_name == "raft_small":
+ weights = Raft_Small_Weights.DEFAULT
+ self.model = raft_small(weights=weights)
+ else:
+ weights = Raft_Large_Weights.DEFAULT
+ self.model = raft_large(weights=weights)
+
+ self.transforms = weights.transforms()
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.model.eval()
+ self.logger.info(f"RAFT model '{model_name}' loaded on {self.device}")
+ except Exception as e:
+ raise ValueError(f"Failed to load RAFT model '{model_name}': {e}")
+
+ def do_forward(self, prev_frame, curr_frame):
+ import torch
+
+ try:
+ H, W = curr_frame.shape[:2]
+
+ # Convert HWC uint8 -> CHW float tensor
+ prev_t = torch.from_numpy(prev_frame).permute(2, 0, 1).float()
+ curr_t = torch.from_numpy(curr_frame).permute(2, 0, 1).float()
+
+ # RAFT requires dimensions divisible by 8
+ pad_h = (8 - H % 8) % 8
+ pad_w = (8 - W % 8) % 8
+ if pad_h > 0 or pad_w > 0:
+ prev_t = torch.nn.functional.pad(prev_t, (0, pad_w, 0, pad_h))
+ curr_t = torch.nn.functional.pad(curr_t, (0, pad_w, 0, pad_h))
+
+ prev_t, curr_t = self.transforms(prev_t, curr_t)
+ prev_batch = prev_t.unsqueeze(0).to(self.device)
+ curr_batch = curr_t.unsqueeze(0).to(self.device)
+
+ with torch.no_grad():
+ flow_predictions = self.model(prev_batch, curr_batch)
+
+ # RAFT returns a list of flow predictions; take the last (finest)
+ flow = flow_predictions[-1].squeeze(0).cpu().numpy()
+ # flow shape: (2, H', W') -> transpose to (H, W, 2) and crop
+ flow = flow.transpose(1, 2, 0)[:H, :W]
+ return flow
+
+ except Exception as e:
+ self.logger.error(f"Optical flow inference error: {e}")
+ return None
diff --git a/plugins/python/engine/sam_engine.py b/plugins/python/engine/sam_engine.py
new file mode 100644
index 0000000..a104e96
--- /dev/null
+++ b/plugins/python/engine/sam_engine.py
@@ -0,0 +1,106 @@
+# SamEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class SamEngine(PyTorchEngine):
+ """
+ PyTorch engine for Segment Anything Model 2 (SAM2).
+
+ Supports HuggingFace model IDs:
+ facebook/sam2-hiera-large
+ facebook/sam2-hiera-base-plus
+ facebook/sam2-hiera-small
+ facebook/sam2-hiera-tiny
+ """
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from transformers import Sam2Model, Sam2Processor
+
+ self.processor = Sam2Processor.from_pretrained(model_name)
+ self.model = Sam2Model.from_pretrained(model_name)
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.model.eval()
+ self.logger.info(f"SAM2 model '{model_name}' loaded on {self.device}")
+ except Exception as e:
+ raise ValueError(f"Failed to load SAM2 model '{model_name}': {e}")
+
+ def do_forward(self, frames, max_masks=10):
+ import numpy as np
+ import torch
+ from PIL import Image
+
+ is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
+ if not is_batch:
+ frames = frames[np.newaxis]
+
+ results = []
+ for frame in frames:
+ try:
+ pil_img = Image.fromarray(frame.astype(np.uint8))
+ H, W = frame.shape[:2]
+
+ # Automatic mask generation: grid of input points
+ grid_size = int(np.ceil(np.sqrt(max_masks)))
+ xs = np.linspace(0, W - 1, grid_size).astype(int)
+ ys = np.linspace(0, H - 1, grid_size).astype(int)
+ points = [[int(x), int(y)] for y in ys for x in xs][:max_masks]
+ input_points = [points]
+
+ inputs = self.processor(
+ images=pil_img,
+ input_points=input_points,
+ return_tensors="pt",
+ )
+ inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+ with torch.no_grad():
+ outputs = self.model(**inputs)
+
+ masks = self.processor.post_process_masks(
+ outputs.pred_masks,
+ inputs["original_sizes"],
+ inputs["reshaped_input_sizes"],
+ )
+ scores = outputs.iou_scores
+
+ mask_list = []
+ if len(masks) > 0:
+ frame_masks = masks[0].cpu().numpy()
+ frame_scores = scores[0].cpu().numpy()
+ for j in range(min(frame_masks.shape[0], max_masks)):
+ best_idx = frame_scores[j].argmax()
+ mask = frame_masks[j, best_idx]
+ score = float(frame_scores[j, best_idx])
+ mask_list.append(
+ {"mask_idx": j, "score": score, "shape": list(mask.shape)}
+ )
+
+ results.append(
+ {
+ "masks": mask_list,
+ "raw_masks": masks[0].cpu().numpy() if len(masks) > 0 else None,
+ }
+ )
+ except Exception as e:
+ self.logger.error(f"SAM inference error on frame: {e}")
+ results.append({"masks": [], "raw_masks": None})
+
+ return results[0] if not is_batch else results
diff --git a/plugins/python/engine/sepformer_engine.py b/plugins/python/engine/sepformer_engine.py
new file mode 100644
index 0000000..63555b5
--- /dev/null
+++ b/plugins/python/engine/sepformer_engine.py
@@ -0,0 +1,101 @@
+# SepformerEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+import os
+
+from .pytorch_engine import PyTorchEngine
+
+
+class SepformerEngine(PyTorchEngine):
+ def __init__(self):
+ super().__init__()
+ self.sample_rate = 0
+
+ def do_load_model(self, model_name, **kwargs):
+ from speechbrain.pretrained import SepformerSeparation
+ from huggingface_hub import snapshot_download
+
+ if not model_name:
+ return
+ self.logger.info(f"Loading Sepformer-WhamR model on device: {self.device}")
+ savedir = "pretrained_models/sepformer-whamr"
+ repo_id = "speechbrain/sepformer-whamr"
+ try:
+ # Download the model files manually to avoid deprecated argument issues
+ if not os.path.exists(savedir):
+ snapshot_download(repo_id=repo_id, local_dir=savedir)
+ # Load from local directory
+ self.model = SepformerSeparation.from_hparams(
+ source=savedir, savedir=savedir, run_opts={"device": self.device}
+ )
+ self.sample_rate = 8000 # Hz, as per SpeechBrain Sepformer models
+ self.sources = ["source0", "source1"] # 2 sources for separation
+ except Exception as e:
+ self.logger.error(f"Failed to load Sepformer-WhamR model: {e}")
+
+ def separate_sources(
+ self,
+ mix,
+ segment=10.0,
+ overlap=0.1,
+ ):
+ import torch
+ from torchaudio.transforms import Fade
+
+ device = mix.device
+ batch, length = mix.shape # For SpeechBrain, input is (batch, time)
+ chunk_len = int(self.sample_rate * segment * (1 + overlap))
+ start = 0
+ end = chunk_len
+ overlap_frames = int(overlap * self.sample_rate)
+ fade = Fade(fade_in_len=0, fade_out_len=overlap_frames, fade_shape="linear")
+
+ final = torch.zeros(batch, len(self.sources), length, device=device)
+
+ min_chunk_samples = int(self.sample_rate * 0.5) # Avoid tiny chunks
+
+ while start < length - overlap_frames:
+ actual_end = min(end, length)
+ chunk_length = actual_end - start
+ if chunk_length < min_chunk_samples:
+ break
+
+ chunk = mix[:, start:actual_end]
+ if chunk_length < chunk_len:
+ pad = chunk_len - chunk_length
+ chunk = torch.nn.functional.pad(chunk, (0, pad))
+
+ # Add small epsilon noise to avoid zero std
+ chunk += 1e-8 * torch.randn_like(chunk)
+
+ with torch.no_grad():
+ out = self.model.separate_batch(chunk) # (batch, time, sources)
+ out = out.permute(0, 2, 1) # (batch, sources, time)
+
+ out = out[:, :, :chunk_length]
+ out = fade(out)
+ final[:, :, start:actual_end] += out
+ if start == 0:
+ fade.fade_in_len = overlap_frames
+ start += int(chunk_len - overlap_frames)
+ else:
+ start += chunk_len
+ end += chunk_len
+ if end >= length:
+ fade.fade_out_len = 0
+ return final
diff --git a/plugins/python/engine/super_res_engine.py b/plugins/python/engine/super_res_engine.py
new file mode 100644
index 0000000..7c87f40
--- /dev/null
+++ b/plugins/python/engine/super_res_engine.py
@@ -0,0 +1,92 @@
+# SuperResEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class SuperResEngine(PyTorchEngine):
+ """
+ PyTorch engine for image super-resolution using Real-ESRGAN.
+
+ Supports model variants:
+ real-esrgan-x4 (4x upscale, general purpose)
+ real-esrgan-x2 (2x upscale)
+ """
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from basicsr.archs.rrdbnet_arch import RRDBNet
+ from realesrgan import RealESRGANer
+
+ scale = 4
+ if "x2" in model_name:
+ scale = 2
+
+ model = RRDBNet(
+ num_in_ch=3,
+ num_out_ch=3,
+ num_feat=64,
+ num_block=23,
+ num_grow_ch=32,
+ scale=scale,
+ )
+
+ model_url = (
+ "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth"
+ if scale == 4
+ else "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth"
+ )
+
+ gpu_id = 0 if str(self.device) != "cpu" else None
+ self.upsampler = RealESRGANer(
+ scale=scale,
+ model_path=model_url,
+ model=model,
+ tile=0,
+ tile_pad=10,
+ pre_pad=0,
+ half=False,
+ gpu_id=gpu_id,
+ )
+ self._scale = scale
+ self.logger.info(f"Real-ESRGAN model '{model_name}' (scale={scale}) loaded")
+ except Exception as e:
+ raise ValueError(f"Failed to load Real-ESRGAN model '{model_name}': {e}")
+
+ def do_forward(self, frames):
+ import cv2
+ import numpy as np
+
+ is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
+ if not is_batch:
+ frames = frames[np.newaxis]
+
+ results = []
+ for frame in frames:
+ try:
+ # Real-ESRGAN expects BGR input
+ bgr = cv2.cvtColor(frame.astype(np.uint8), cv2.COLOR_RGB2BGR)
+ output, _ = self.upsampler.enhance(bgr, outscale=self._scale)
+ # Convert back to RGB
+ rgb_out = cv2.cvtColor(output, cv2.COLOR_BGR2RGB)
+ results.append(rgb_out)
+ except Exception as e:
+ self.logger.error(f"Super-resolution inference error: {e}")
+ results.append(None)
+
+ return results[0] if not is_batch else results
diff --git a/plugins/python/engine/vlm_engine.py b/plugins/python/engine/vlm_engine.py
new file mode 100644
index 0000000..903853d
--- /dev/null
+++ b/plugins/python/engine/vlm_engine.py
@@ -0,0 +1,138 @@
+# VlmEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class VlmEngine(PyTorchEngine):
+ """
+ PyTorch engine for Vision-Language Models.
+
+ Supports HuggingFace VLM model IDs via AutoProcessor + AutoModelForVision2Seq:
+ llava-hf/llava-1.5-7b-hf
+ Qwen/Qwen2-VL-7B-Instruct
+ OpenGVLab/InternVL2-8B
+ """
+
+ def __init__(self):
+ super().__init__()
+ self.processor = None
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ import torch
+ from transformers import AutoProcessor, AutoModelForVision2Seq
+
+ self.processor = AutoProcessor.from_pretrained(model_name)
+ self.model = AutoModelForVision2Seq.from_pretrained(
+ model_name,
+ torch_dtype=torch.float16,
+ device_map=self.device,
+ )
+ self.model.eval()
+ self.logger.info(f"VLM model '{model_name}' loaded on {self.device}")
+ except Exception as e:
+ raise ValueError(f"Failed to load VLM model '{model_name}': {e}")
+
+ def do_forward(
+ self,
+ frame,
+ prompt="Describe this image in detail.",
+ system_prompt=None,
+ max_tokens=256,
+ temperature=0.7,
+ ):
+ """
+ Run VLM inference on a single video frame.
+
+ Args:
+ frame: numpy RGB array (H, W, 3).
+ prompt: user prompt text.
+ system_prompt: optional system prompt.
+ max_tokens: maximum tokens to generate.
+ temperature: sampling temperature.
+
+ Returns:
+ Generated text string, or None on failure.
+ """
+ import numpy as np
+ from PIL import Image
+
+ try:
+ pil_img = Image.fromarray(frame.astype(np.uint8))
+ text = self.do_generate(
+ pil_img, prompt, system_prompt, max_tokens, temperature
+ )
+ return text
+ except Exception as e:
+ self.logger.error(f"VLM inference error: {e}")
+ return None
+
+ def do_generate(self, image, prompt, system_prompt, max_tokens, temperature):
+ """
+ Apply chat template, process image + text, and generate a response.
+
+ Args:
+ image: PIL Image.
+ prompt: user prompt text.
+ system_prompt: optional system prompt.
+ max_tokens: maximum new tokens.
+ temperature: sampling temperature.
+
+ Returns:
+ Generated text string.
+ """
+ import torch
+
+ messages = []
+ if system_prompt:
+ messages.append({"role": "system", "content": system_prompt})
+ messages.append(
+ {
+ "role": "user",
+ "content": [
+ {"type": "image"},
+ {"type": "text", "text": prompt},
+ ],
+ }
+ )
+
+ # Apply chat template if the processor supports it
+ if hasattr(self.processor, "apply_chat_template"):
+ text_input = self.processor.apply_chat_template(
+ messages, add_generation_prompt=True
+ )
+ else:
+ text_input = prompt
+
+ inputs = self.processor(text=text_input, images=image, return_tensors="pt")
+ inputs = {k: v.to(self.model.device) for k, v in inputs.items()}
+
+ with torch.no_grad():
+ output_ids = self.model.generate(
+ **inputs,
+ max_new_tokens=max_tokens,
+ temperature=temperature,
+ do_sample=temperature > 0,
+ )
+
+ # Decode only the newly generated tokens
+ input_len = inputs.get("input_ids", torch.tensor([])).shape[-1]
+ generated = output_ids[0][input_len:]
+ text = self.processor.decode(generated, skip_special_tokens=True)
+ return text.strip()
diff --git a/plugins/python/engine/whisper_engine.py b/plugins/python/engine/whisper_engine.py
new file mode 100644
index 0000000..33d9803
--- /dev/null
+++ b/plugins/python/engine/whisper_engine.py
@@ -0,0 +1,34 @@
+# WhisperEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class WhisperEngine(PyTorchEngine):
+ def do_load_model(self, model_name, **kwargs):
+ from faster_whisper import WhisperModel
+
+ if not model_name:
+ return
+ compute_type = "float16" if self.device.startswith("cuda") else "int8"
+ self.logger.info(
+ f"Loading Whisper model on device: {self.device} with compute_type: {compute_type}"
+ )
+ self.model = WhisperModel(
+ model_name, device=self.device, compute_type=compute_type
+ )
diff --git a/plugins/python/engine/yolo_advanced_engine.py b/plugins/python/engine/yolo_advanced_engine.py
new file mode 100644
index 0000000..28c7273
--- /dev/null
+++ b/plugins/python/engine/yolo_advanced_engine.py
@@ -0,0 +1,1060 @@
+# YoloAdvancedEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from collections import deque
+
+from .pytorch_engine import PyTorchEngine
+
+BOT_OK = BYTE_OK = CFG_OK = True
+BOTSORT = BYTETracker = get_cfg = Boxes = None
+
+
+def _init_ultralytics():
+ global BOT_OK, BYTE_OK, CFG_OK, BOTSORT, BYTETracker, get_cfg, Boxes
+ try:
+ from ultralytics.trackers.bot_sort import BOTSORT as _BS
+
+ BOTSORT = _BS
+ except Exception:
+ BOT_OK = False
+ try:
+ from ultralytics.trackers.byte_tracker import BYTETracker as _BT
+
+ BYTETracker = _BT
+ except Exception:
+ BYTE_OK = False
+ try:
+ from ultralytics.cfg import get_cfg as _gcfg
+
+ get_cfg = _gcfg
+ except Exception:
+ CFG_OK = False
+ try:
+ from ultralytics.engine.results import Boxes as _Boxes
+
+ Boxes = _Boxes
+ except Exception:
+ pass
+
+
+def eye3():
+ import numpy as np
+
+ return np.eye(3, dtype=np.float32)
+
+
+def estimate_global_motion(
+ prev_gray,
+ gray,
+ gmc_mode,
+ gmc_scale,
+ gft_max_corners,
+ gft_quality,
+ gft_min_dist,
+ lk_win,
+ lk_levels,
+ ransac_thresh,
+ frame_idx,
+ verbose=False,
+):
+ import cv2
+ import numpy as np
+
+ if gmc_mode == "off":
+ if verbose:
+ print(f"[frame {frame_idx}] GMC OFF → I")
+ return eye3()
+
+ def down(img):
+ if gmc_scale == 1.0:
+ return img
+ w = max(2, int(img.shape[1] * gmc_scale))
+ h = max(2, int(img.shape[0] * gmc_scale))
+ return cv2.resize(img, (w, h), interpolation=cv2.INTER_AREA)
+
+ pg, cg = down(prev_gray), down(gray)
+ pts_prev = cv2.goodFeaturesToTrack(
+ pg,
+ maxCorners=gft_max_corners,
+ qualityLevel=gft_quality,
+ minDistance=gft_min_dist,
+ )
+ if pts_prev is None or len(pts_prev) < 6:
+ if verbose:
+ print(f"[frame {frame_idx}] GMC: insufficient corners → I")
+ return eye3()
+
+ pts_curr, st, _ = cv2.calcOpticalFlowPyrLK(
+ pg,
+ cg,
+ pts_prev,
+ None,
+ winSize=(lk_win, lk_win),
+ maxLevel=lk_levels,
+ criteria=(cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 30, 0.01),
+ )
+ if pts_curr is None or st is None:
+ if verbose:
+ print(f"[frame {frame_idx}] GMC: LK failed → I")
+ return eye3()
+
+ m = st.reshape(-1).astype(bool)
+ if m.sum() < (4 if gmc_mode == "homography" else 3):
+ if verbose:
+ print(f"[frame {frame_idx}] GMC: not enough inliers → I")
+ return eye3()
+
+ src = pts_prev[m]
+ dst = pts_curr[m]
+ if gmc_scale != 1.0:
+ s = 1.0 / gmc_scale
+ src *= s
+ dst *= s
+
+ if gmc_mode == "homography":
+ H, _ = cv2.findHomography(
+ src,
+ dst,
+ cv2.RANSAC,
+ ransacReprojThreshold=ransac_thresh,
+ maxIters=1000,
+ )
+ H = H.astype(np.float32) if H is not None else eye3()
+ else:
+ A, _ = cv2.estimateAffine2D(
+ src, dst, ransacReprojThreshold=ransac_thresh, maxIters=1000
+ )
+ H = np.vstack([A, [0, 0, 1]]).astype(np.float32) if A is not None else eye3()
+
+ avg = float(np.mean(np.linalg.norm(dst - src, axis=1))) if len(src) > 0 else 0.0
+ if verbose:
+ print(
+ f"[frame {frame_idx}] GMC {gmc_mode} inliers={int(m.sum())} avg_motion={avg:.2f}px"
+ )
+ return H
+
+
+def warp_points(points_xy, M):
+ import numpy as np
+
+ if not points_xy:
+ return []
+ P = np.c_[
+ np.array(points_xy, dtype=np.float32), np.ones((len(points_xy), 1), np.float32)
+ ]
+ Q = (M @ P.T).T
+ Q = Q[:, :2] / np.clip(Q[:, 2:3], 1e-6, None)
+ return [tuple(q) for q in Q]
+
+
+def classwise_keep(result, person_thr, ball_thr):
+ import numpy as np
+
+ if result is None or result.boxes is None or len(result.boxes) == 0:
+ return np.zeros((0, 6), np.float32), np.zeros((0, 6), np.float32)
+
+ b = result.boxes
+ xyxy = b.xyxy.cpu().numpy()
+ conf = (
+ b.conf.cpu().numpy() if b.conf is not None else np.ones((len(b),), np.float32)
+ )
+ cls = (
+ b.cls.cpu().numpy().astype(int)
+ if b.cls is not None
+ else np.zeros((len(b),), np.int32)
+ )
+
+ keep_p = (cls == 0) & (conf >= person_thr)
+ keep_b = (cls == 32) & (conf >= ball_thr)
+
+ dets_p = np.c_[xyxy[keep_p], conf[keep_p], cls[keep_p]].astype(np.float32)
+ dets_b = np.c_[xyxy[keep_b], conf[keep_b], cls[keep_b]].astype(np.float32)
+ return dets_p, dets_b
+
+
+def dets_to_boxes(dets_xyxy_conf_cls, frame_shape):
+ _init_ultralytics()
+ if dets_xyxy_conf_cls is None or dets_xyxy_conf_cls.size == 0:
+ import torch as _torch
+
+ data = _torch.zeros((0, 6), dtype=_torch.float32)
+ return Boxes(data, frame_shape)
+ import torch as _torch
+
+ data = _torch.from_numpy(dets_xyxy_conf_cls).to(_torch.float32)
+ return Boxes(data, frame_shape)
+
+
+def clamp_imgsz_for_device(imgsz, device_str):
+ if device_str in ("cpu", "auto"):
+ return min(imgsz, 1280)
+ return imgsz
+
+
+def expand_roi(xyxy, scale, W, H, min_side=256, max_side=1920):
+ x1, y1, x2, y2 = map(float, xyxy)
+ cx, cy = (x1 + x2) * 0.5, (y1 + y2) * 0.5
+ w, h = (x2 - x1), (y2 - y1)
+ side = max(w, h) * float(scale)
+ side = max(min_side, min(side, max_side))
+ x1n = max(0, int(cx - side * 0.5))
+ y1n = max(0, int(cy - side * 0.5))
+ x2n = min(W - 1, int(cx + side * 0.5))
+ y2n = min(H - 1, int(cy + side * 0.5))
+ return x1n, y1n, x2n, y2n
+
+
+def _normalize_tracker_args(args, kind="byte"):
+ def has(a):
+ return hasattr(args, a) and getattr(args, a) is not None
+
+ def setif(a, v):
+ setattr(args, a, v)
+
+ def copy_if_missing(target, *sources, default=None):
+ if not has(target):
+ for s in sources:
+ if has(s):
+ setif(target, getattr(args, s))
+ return
+ if default is not None:
+ setif(target, default)
+
+ copy_if_missing("track_high_thresh", "track_thresh", default=0.5)
+ copy_if_missing("track_thresh", "track_high_thresh", default=0.5)
+ copy_if_missing("track_low_thresh", default=0.1)
+ copy_if_missing("new_track_thresh", default=0.4)
+ copy_if_missing("match_thresh", default=0.8)
+ copy_if_missing("asso_thresh", "match_thresh", default=0.8)
+ copy_if_missing("track_buffer", default=30)
+ copy_if_missing("frame_rate", default=30)
+ copy_if_missing("fuse_score", default=True)
+ copy_if_missing("fuse_score_coef", default=1.0)
+ copy_if_missing("mot20", default=False)
+
+ if kind == "botsort":
+ copy_if_missing("with_reid", default=True)
+ copy_if_missing("proximity_thresh", default=0.5)
+ copy_if_missing("appearance_thresh", default=0.25)
+ if has("cmc_method") and not has("gmc_method"):
+ setif("gmc_method", getattr(args, "cmc_method"))
+ copy_if_missing("gmc_method", default="sparseOptFlow")
+
+
+class ByteTrackWrapper:
+ """Ultralytics BYTETracker; load params from YAML using get_cfg, accept Boxes."""
+
+ def __init__(self, yaml_path, frame_rate):
+ if not BYTE_OK:
+ raise RuntimeError("Ultralytics BYTETracker not available")
+ if not CFG_OK:
+ raise RuntimeError(
+ "Ultralytics get_cfg not available; update ultralytics package."
+ )
+ args = get_cfg(yaml_path)
+ args.frame_rate = int(frame_rate)
+ _normalize_tracker_args(args, kind="byte")
+ self.tracker = BYTETracker(args, frame_rate=int(frame_rate))
+
+ def update(self, boxes: Boxes, frame):
+ return self.tracker.update(boxes, frame)
+
+
+class BoTSORTWrapper:
+ """Ultralytics BOTSORT; load params via get_cfg, accept Boxes, safe ReID encoder."""
+
+ def __init__(self, yaml_path, frame_rate, enable_reid=True):
+ if not BOT_OK:
+ raise RuntimeError("Ultralytics BOTSORT not available")
+ if not CFG_OK:
+ raise RuntimeError(
+ "Ultralytics get_cfg not available; update ultralytics package."
+ )
+ args = get_cfg(yaml_path)
+ args.frame_rate = int(frame_rate)
+ args.with_reid = bool(enable_reid)
+ _normalize_tracker_args(args, kind="botsort")
+ self.tracker = BOTSORT(args, frame_rate=int(frame_rate))
+ self._install_safe_encoder()
+
+ def _install_safe_encoder(self):
+ import cv2
+ import numpy as np
+ import torch as _torch
+
+ def _safe_hsv_hist(img_bgr, bboxes):
+ feats = []
+ if img_bgr is None or bboxes is None:
+ return feats
+
+ if hasattr(bboxes, "detach"):
+ bb = bboxes.detach().cpu().numpy()
+ elif isinstance(bboxes, np.ndarray):
+ bb = bboxes
+ else:
+ try:
+ bb = np.asarray(bboxes, dtype=np.float32)
+ except Exception:
+ bb = None
+
+ H, W = img_bgr.shape[:2]
+
+ def process_one(b):
+ b = np.asarray(b, dtype=np.float32).reshape(-1)
+ x1, y1, x2, y2 = map(float, b[:4])
+ x1i = max(0, min(int(x1), W - 1))
+ y1i = max(0, min(int(y1), H - 1))
+ x2i = max(0, min(int(x2), W - 1))
+ y2i = max(0, min(int(y2), H - 1))
+ if x2i <= x1i or y2i <= y1i:
+ return _torch.zeros(512, dtype=_torch.float32)
+ crop = img_bgr[y1i:y2i, x1i:x2i]
+ if crop.size == 0:
+ return _torch.zeros(512, dtype=_torch.float32)
+ hsv = cv2.cvtColor(crop, cv2.COLOR_BGR2HSV)
+ hist = cv2.calcHist(
+ [hsv], [0, 1, 2], None, [8, 8, 8], [0, 180, 0, 256, 0, 256]
+ ).flatten()
+ norm = np.linalg.norm(hist) + 1e-6
+ hist = (hist / norm).astype(np.float32)
+ return _torch.from_numpy(hist)
+
+ if bb is not None:
+ for b in bb:
+ feats.append(process_one(b))
+ else:
+ for b in bboxes:
+ feats.append(process_one(b))
+ return feats
+
+ self.tracker.encoder = lambda img, tlbrs: _safe_hsv_hist(img, tlbrs)
+
+ def update(self, boxes: Boxes, frame):
+ return self.tracker.update(boxes, frame)
+
+
+def _callable_or_attr(obj, name):
+ v = getattr(obj, name, None)
+ if v is None:
+ return None
+ return v() if callable(v) else v
+
+
+def tlbr_of(tr):
+ a = _callable_or_attr(tr, "tlbr")
+ if a is not None:
+ a = a.tolist() if hasattr(a, "tolist") else a
+ if len(a) == 4:
+ return a
+ a = _callable_or_attr(tr, "tlwh")
+ if a is not None:
+ a = a.tolist() if hasattr(a, "tolist") else a
+ if len(a) == 4:
+ x, y, w, h = a
+ return [x, y, x + w, y + h]
+ if hasattr(tr, "bbox"):
+ b = tr.bbox
+ return b.tolist() if hasattr(b, "tolist") else list(b)
+ return None
+
+
+class BallState:
+ def __init__(self):
+ self.cx = None
+ self.cy = None
+ self.vx = 0.0
+ self.vy = 0.0
+ self.frame = -1
+
+ def predict(self, frame_idx, decay=0.85):
+ if self.cx is None:
+ return None
+ return (self.cx + decay * self.vx, self.cy + decay * self.vy)
+
+ def update_from_xyxy(self, xyxy, frame_idx):
+ x1, y1, x2, y2 = map(float, xyxy)
+ cx, cy = 0.5 * (x1 + x2), 0.5 * (y1 + y2)
+ if self.cx is not None and self.frame >= 0:
+ dt = max(1, frame_idx - self.frame)
+ self.vx = (cx - self.cx) / dt
+ self.vy = (cy - self.cy) / dt
+ self.cx, self.cy, self.frame = cx, cy, frame_idx
+
+ def update_from_center(self, cx, cy, frame_idx):
+ if self.cx is not None and self.frame >= 0:
+ dt = max(1, frame_idx - self.frame)
+ self.vx = (cx - self.cx) / dt
+ self.vy = (cy - self.cy) / dt
+ self.cx, self.cy, self.frame = float(cx), float(cy), int(frame_idx)
+
+
+def _aspect_round_penalty(w, h):
+ import numpy as np
+
+ ar = w / max(h, 1e-6)
+ roundness = np.exp(-((ar - 1.0) ** 2) / 0.15)
+ return 1.0 - float(roundness)
+
+
+def _size_penalty(w, h, H, W):
+ import numpy as np
+
+ s = max(w, h)
+ tgt = 0.03 * min(H, W)
+ return float(np.clip(abs(s - tgt) / (tgt + 1e-6), 0.0, 2.0)) * 0.5
+
+
+def select_best_ball(
+ dets_b,
+ frame_shape,
+ ball_state,
+ frame_idx,
+ w_conf=1.0,
+ w_dist=0.015,
+ w_size=0.5,
+ w_round=0.4,
+):
+ import numpy as np
+
+ if dets_b is None or len(dets_b) == 0:
+ return None
+ H, W = frame_shape
+ pred = ball_state.predict(frame_idx)
+ scores = []
+ for d in dets_b:
+ x1, y1, x2, y2, conf, _ = d
+ cx, cy = 0.5 * (x1 + x2), 0.5 * (y1 + y2)
+ w, h = (x2 - x1), (y2 - y1)
+ s_conf = float(conf)
+ if pred is None:
+ dist_pen = 0.0
+ else:
+ px, py = pred
+ dist = np.hypot(cx - px, cy - py)
+ dist_pen = float(dist / (0.5 * (H + W)))
+ size_pen = _size_penalty(w, h, H, W)
+ round_pen = _aspect_round_penalty(w, h)
+ s = (
+ w_conf * s_conf
+ - w_dist * dist_pen
+ - w_size * size_pen
+ - w_round * round_pen
+ )
+ scores.append(s)
+ if not scores:
+ return None
+ return int(np.argmax(scores))
+
+
+def nms_class(dets, iou_thr=0.5):
+ import numpy as np
+
+ if dets is None or len(dets) == 0:
+ return dets
+ boxes = dets[:, :4].copy()
+ scores = dets[:, 4].copy()
+ order = scores.argsort()[::-1]
+ keep = []
+
+ def iou(a, b):
+ xx1 = np.maximum(a[0], b[0])
+ yy1 = np.maximum(a[1], b[1])
+ xx2 = np.minimum(a[2], b[2])
+ yy2 = np.minimum(a[3], b[3])
+ w = np.maximum(0.0, xx2 - xx1)
+ h = np.maximum(0.0, yy2 - yy1)
+ inter = w * h
+ area_a = (a[2] - a[0]) * (a[3] - a[1])
+ area_b = (b[2] - b[0]) * (b[3] - b[1])
+ return inter / (area_a + area_b - inter + 1e-6)
+
+ while order.size > 0:
+ i = order[0]
+ keep.append(i)
+ if order.size == 1:
+ break
+ ious = np.array([iou(boxes[i], boxes[j]) for j in order[1:]])
+ remain = np.where(ious <= iou_thr)[0]
+ order = order[remain + 1]
+ return dets[keep]
+
+
+def add_trail_point(seq: deque, x: int, y: int, k: int, densify=True, max_gap=5):
+ if len(seq) > 0 and densify:
+ _, _, k_prev = seq[-1]
+ gap = k - k_prev
+ if 1 < gap <= max_gap:
+ x_prev, y_prev, _ = seq[-1]
+ for t in range(1, gap):
+ alpha = t / gap
+ xi = int(round((1 - alpha) * x_prev + alpha * x))
+ yi = int(round((1 - alpha) * y_prev + alpha * y))
+ seq.append((xi, yi, k_prev + t))
+ seq.append((int(x), int(y), int(k)))
+
+
+def iou_xyxy(a, b):
+ if a is None or b is None:
+ return 0.0
+ ax1, ay1, ax2, ay2 = a
+ bx1, by1, bx2, by2 = b
+ xx1 = max(ax1, bx1)
+ yy1 = max(ay1, by1)
+ xx2 = min(ax2, bx2)
+ yy2 = min(ay2, by2)
+ w = max(0.0, xx2 - xx1)
+ h = max(0.0, yy2 - yy1)
+ inter = w * h
+ if inter <= 0:
+ return 0.0
+ area_a = max(0.0, (ax2 - ax1)) * max(0.0, (ay2 - ay1))
+ area_b = max(0.0, (bx2 - bx1)) * max(0.0, (by2 - by1))
+ denom = area_a + area_b - inter + 1e-6
+ return float(inter / denom)
+
+
+def lerp(a, b, t):
+ return a * (1.0 - t) + b * t
+
+
+def gate_accept(
+ center,
+ cand_box,
+ trail_seq,
+ last_shown_box,
+ recent_speed,
+ frame_shape,
+ args,
+ pred=None,
+ from_track=True,
+):
+ import numpy as np
+
+ if center is None:
+ return False
+
+ H, W = frame_shape
+ hard_cap = float(args.ball_max_jump_rel) * float(min(H, W))
+
+ if len(trail_seq) == 0:
+ dist_prev = 0.0
+ else:
+ x_prev, y_prev, _ = trail_seq[-1]
+ dist_prev = float(np.hypot(center[0] - x_prev, center[1] - y_prev))
+
+ if dist_prev > hard_cap:
+ return False
+
+ base_gate = max(
+ float(args.ball_gate_min), float(args.ball_gate_rel) * float(min(H, W))
+ )
+ gate_px = base_gate if from_track else base_gate * 1.25
+ pass_prev = (len(trail_seq) == 0) or (dist_prev <= gate_px)
+
+ pred_ok = False
+ if getattr(args, "ball_gate_use_pred", False) and pred is not None:
+ d_pred = float(np.hypot(center[0] - pred[0], center[1] - pred[1]))
+ pred_ok = d_pred <= gate_px * 1.25
+
+ if not from_track:
+ return pass_prev or pred_ok
+
+ iou_ok = (last_shown_box is None) or (
+ iou_xyxy(cand_box, last_shown_box) >= float(args.ball_min_iou)
+ )
+
+ speed_ok = True
+ if recent_speed is not None and recent_speed > 0:
+ speed_ok = dist_prev <= float(args.ball_speed_mult) * float(recent_speed + 1e-6)
+
+ return (pass_prev and iou_ok and speed_ok) or pred_ok
+
+
+def safe_int_pair(wx, wy, W, H):
+ import numpy as np
+
+ if wx is None or wy is None:
+ return None
+ if not (np.isfinite(wx) and np.isfinite(wy)):
+ return None
+ try:
+ xi = int(round(float(wx)))
+ yi = int(round(float(wy)))
+ except Exception:
+ return None
+ if abs(xi) > 10 * W or abs(yi) > 10 * H:
+ return None
+ return xi, yi
+
+
+class YoloAdvancedEngine(PyTorchEngine):
+ def __init__(self, device=None, **kwargs):
+ super().__init__(device=device)
+ # Then set self.device_str = device if device else 'auto'
+ self.device_str = device if device else "auto"
+ self.det_model = None
+ self.fb_model = None
+ self.people_tracker = None
+ self.ball_tracker = None
+ self.ball_state = BallState()
+ self.single_ball_trail = deque(maxlen=kwargs.get("trail", 200))
+ self.cum_H_history = [eye3()]
+ self.cum_H = eye3()
+ self.prev_gray = None
+ self.last_ball_xyxy = None
+ self.last_shown_box = None
+ self.recent_speed = None
+ self.ema_cxcy = None
+ self.coast_streak = 0
+ self.det_reject_streak = 0
+ self.dropped_by_gate = 0
+ self.coast_used = 0
+ self.frame_idx = 0
+ self.frame_rate = kwargs.get("frame_rate", 30.0)
+ # Set all params from kwargs
+ self.device_str = kwargs.get("device", "auto")
+ self.imgsz = kwargs.get("imgsz", 1280)
+ self.conf = kwargs.get("conf", 0.25)
+ self.iou = kwargs.get("iou", 0.45)
+ self.classes = kwargs.get("classes", [0, 32])
+ self.person_conf_keep = kwargs.get("person_conf_keep", 0.25)
+ self.ball_conf_keep = kwargs.get("ball_conf_keep", 0.04)
+ self.ball_mode = kwargs.get("ball_mode", True)
+ self.hires_fallback = kwargs.get("hires_fallback", True)
+ self.hires_imgsz = kwargs.get("hires_imgsz", 1536)
+ self.fallback_every = kwargs.get("fallback_every", 6)
+ self.fallback_tiles = kwargs.get("fallback_tiles", False)
+ self.tile_size = kwargs.get("tile_size", 1280)
+ self.tile_overlap = kwargs.get("tile_overlap", 256)
+ self.fallback_budget_ms = kwargs.get("fallback_budget_ms", 300)
+ self.ball_roi_boost = kwargs.get("ball_roi_boost", False)
+ self.roi_scale = kwargs.get("roi_scale", 2.5)
+ self.roi_min = kwargs.get("roi_min", 256)
+ self.roi_max = kwargs.get("roi_max", 1920)
+ self.tracker_people = kwargs.get("tracker_people", "botsort_people_reid.yaml")
+ self.tracker_ball = kwargs.get("tracker_ball", "bytetrack_ball.yaml")
+ self.people_reid = kwargs.get("people_reid", True)
+ self.trail = kwargs.get("trail", 200)
+ self.gmc = kwargs.get("gmc", "affine")
+ self.gmc_scale = kwargs.get("gmc_scale", 0.5)
+ self.gft_max_corners = kwargs.get("gft_max_corners", 400)
+ self.gft_quality = kwargs.get("gft_quality", 0.01)
+ self.gft_min_dist = kwargs.get("gft_min_dist", 8)
+ self.lk_win = kwargs.get("lk_win", 21)
+ self.lk_levels = kwargs.get("lk_levels", 3)
+ self.ransac_thresh = kwargs.get("ransac_thresh", 3.0)
+ self.ball_gate_rel = kwargs.get("ball_gate_rel", 0.06)
+ self.ball_gate_min = kwargs.get("ball_gate_min", 12)
+ self.ball_gate_use_pred = kwargs.get("ball_gate_use_pred", False)
+ self.ball_min_iou = kwargs.get("ball_min_iou", 0.20)
+ self.ball_max_jump_rel = kwargs.get("ball_max_jump_rel", 0.12)
+ self.ball_speed_mult = kwargs.get("ball_speed_mult", 3.0)
+ self.ball_smooth_ema = kwargs.get("ball_smooth_ema", 0.0)
+ self.det_override_conf = kwargs.get("det_override_conf", 0.28)
+ self.det_override_after = kwargs.get("det_override_after", 2)
+ self.reacquire_frames = kwargs.get("reacquire_frames", 6)
+ self.ball_coast = kwargs.get("ball_coast", False)
+ self.coast_max = kwargs.get("coast_max", 6)
+ self.coast_decay = kwargs.get("coast_decay", 0.90)
+ self.verbose = kwargs.get("verbose", False)
+
+ def do_load_model(self, model_name, **kwargs):
+ _init_ultralytics()
+ try:
+ from ultralytics import YOLO
+
+ # YOLO load unchanged...
+ self.det_model = YOLO(f"{model_name}.pt")
+ self.execute_with_stream(lambda: self.det_model.to(self.device))
+ self.logger.info(
+ f"YOLO primary model '{model_name}' loaded on {self.device}"
+ )
+
+ if self.hires_fallback:
+ self.fb_model = YOLO(f"{model_name}.pt")
+ self.execute_with_stream(lambda: self.fb_model.to(self.device))
+
+ self.model = self.det_model # Alias for base compat
+
+ # Trackers with fallback
+ if self.tracker_people:
+ try:
+ self.people_tracker = BoTSORTWrapper(
+ self.tracker_people, self.frame_rate, self.people_reid
+ )
+ self.logger.info(
+ f"People tracker loaded from {self.tracker_people}"
+ )
+ except Exception as te:
+ self.logger.warning(f"People tracker failed ({te}); disabling.")
+ self.people_tracker = None
+
+ if self.tracker_ball:
+ try:
+ self.ball_tracker = ByteTrackWrapper(
+ self.tracker_ball, self.frame_rate
+ )
+ self.logger.info(f"Ball tracker loaded from {self.tracker_ball}")
+ except Exception as te:
+ self.logger.warning(f"Ball tracker failed ({te}); disabling.")
+ self.ball_tracker = None
+
+ # ... kwargs update unchanged ...
+ return self.tracker_people and self.tracker_ball
+
+ except Exception as e:
+ self.logger.error(f"Core model load failed: {e}")
+ return False # No raise—let base handle
+
+ def do_forward(self, frames):
+ import cv2
+ import numpy as np
+
+ is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
+ if is_batch:
+ frame_bgr = frames[0] # Assume single for stateful; extend if needed
+ else:
+ frame_bgr = np.array(frames, copy=True)
+ gray = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2GRAY)
+ Hh, Ww = frame_bgr.shape[:2]
+
+ # GMC
+ if self.prev_gray is not None:
+ H = estimate_global_motion(
+ self.prev_gray,
+ gray,
+ self.gmc,
+ self.gmc_scale,
+ self.gft_max_corners,
+ self.gft_quality,
+ self.gft_min_dist,
+ self.lk_win,
+ self.lk_levels,
+ self.ransac_thresh,
+ self.frame_idx,
+ self.verbose,
+ )
+ self.cum_H = H @ self.cum_H
+ self.cum_H_history.append(self.cum_H.copy())
+ self.prev_gray = gray
+
+ # Detection
+ det_res = self.execute_with_stream(
+ lambda: self.det_model.predict(
+ frame_bgr,
+ imgsz=self.imgsz,
+ conf=self.conf,
+ iou=self.iou,
+ classes=self.classes,
+ device=self.device_str if self.device_str != "auto" else None,
+ verbose=False,
+ )[0]
+ )
+
+ dets_p, dets_b = classwise_keep(
+ det_res, self.person_conf_keep, self.ball_conf_keep
+ )
+
+ dets_b = nms_class(dets_b, iou_thr=0.35)
+ best_idx = select_best_ball(
+ dets_b, frame_bgr.shape[:2], self.ball_state, self.frame_idx
+ )
+ if best_idx is not None:
+ dets_b = dets_b[[best_idx]]
+ self.ball_state.update_from_xyxy(dets_b[0, :4], self.frame_idx)
+ self.last_ball_xyxy = dets_b[0, :4].astype(int).tolist()
+ else:
+ dets_b = dets_b[:0]
+
+ # Fallback logic
+ if (
+ self.ball_mode
+ and self.hires_fallback
+ and dets_b.shape[0] == 0
+ and self.frame_idx % max(1, self.fallback_every) == 0
+ ):
+ clamp_imgsz = clamp_imgsz_for_device(self.hires_imgsz, self.device_str)
+ collected = []
+
+ if (
+ self.ball_roi_boost
+ and self.last_ball_xyxy is not None
+ and self.fb_model is not None
+ ):
+ x1r, y1r, x2r, y2r = expand_roi(
+ self.last_ball_xyxy,
+ self.roi_scale,
+ Ww,
+ Hh,
+ min_side=self.roi_min,
+ max_side=self.roi_max,
+ )
+ crop = frame_bgr[y1r:y2r, x1r:x2r]
+ pred = self.execute_with_stream(
+ lambda: self.fb_model.predict(
+ crop,
+ imgsz=min(max(x2r - x1r, y2r - y1r), clamp_imgsz),
+ conf=max(self.ball_conf_keep, 0.02),
+ iou=max(self.iou, 0.50),
+ classes=[32],
+ device=self.device_str if self.device_str != "auto" else None,
+ verbose=False,
+ )[0]
+ )
+ if pred.boxes is not None and len(pred.boxes) > 0:
+ b = pred.boxes
+ xyxy = b.xyxy.cpu().numpy()
+ conf = (
+ b.conf.cpu().numpy()
+ if b.conf is not None
+ else np.ones((len(b),), np.float32)
+ )
+ cls = np.full((len(b),), 32, dtype=np.float32)
+ xyxy[:, [0, 2]] += x1r
+ xyxy[:, [1, 3]] += y1r
+ collected.append(np.c_[xyxy, conf, cls])
+
+ if self.fb_model is not None and not collected:
+ pred = self.execute_with_stream(
+ lambda: self.fb_model.predict(
+ frame_bgr,
+ imgsz=clamp_imgsz,
+ conf=max(self.ball_conf_keep, 0.02),
+ iou=max(self.iou, 0.50),
+ classes=[32],
+ device=self.device_str if self.device_str != "auto" else None,
+ verbose=False,
+ )[0]
+ )
+ if pred.boxes is not None and len(pred.boxes) > 0:
+ b = pred.boxes
+ xyxy = b.xyxy.cpu().numpy()
+ conf = (
+ b.conf.cpu().numpy()
+ if b.conf is not None
+ else np.ones((len(b),), np.float32)
+ )
+ cls = np.full((len(b),), 32, dtype=np.float32)
+ collected.append(np.c_[xyxy, conf, cls])
+
+ if collected:
+ dets_b = np.vstack(collected).astype(np.float32)
+ dets_b = nms_class(dets_b, iou_thr=0.35)
+ best_idx = select_best_ball(
+ dets_b, frame_bgr.shape[:2], self.ball_state, self.frame_idx
+ )
+ if best_idx is not None:
+ dets_b = dets_b[[best_idx]]
+ self.ball_state.update_from_xyxy(dets_b[0, :4], self.frame_idx)
+ self.last_ball_xyxy = dets_b[0, :4].astype(int).tolist()
+ else:
+ dets_b = dets_b[:0]
+
+ frame_shape = frame_bgr.shape[:2]
+ boxes_p = dets_to_boxes(dets_p, frame_shape)
+ boxes_b = dets_to_boxes(dets_b, frame_shape)
+ tracks_p = (
+ self.people_tracker.update(boxes_p, frame_bgr)
+ if self.people_tracker
+ else []
+ )
+ tracks_b = (
+ self.ball_tracker.update(boxes_b, frame_bgr) if self.ball_tracker else []
+ )
+
+ # Ball candidate selection and gating
+ cand_box = None
+ ball_center_candidate = None
+ cand_conf = None
+ from_track = False
+
+ if len(tracks_b) >= 1:
+ tr = tracks_b[0]
+ tb = tlbr_of(tr)
+ if tb is not None:
+ x1, y1, x2, y2 = map(int, tb)
+ cand_box = [x1, y1, x2, y2]
+ ball_center_candidate = ((x1 + x2) // 2, (y1 + y2) // 2)
+ from_track = True
+ cand_conf = None
+ elif dets_b.shape[0] == 1:
+ x1, y1, x2, y2 = map(int, dets_b[0, :4])
+ cand_box = [x1, y1, x2, y2]
+ ball_center_candidate = ((x1 + x2) // 2, (y1 + y2) // 2)
+ cand_conf = float(dets_b[0, 4])
+ from_track = False
+
+ pred_pos = (
+ self.ball_state.predict(self.frame_idx) if self.ball_gate_use_pred else None
+ )
+ accept = gate_accept(
+ ball_center_candidate,
+ cand_box,
+ self.single_ball_trail,
+ self.last_shown_box,
+ self.recent_speed,
+ frame_bgr.shape[:2],
+ self, # Use self as args
+ pred=pred_pos,
+ from_track=from_track,
+ )
+
+ # Det override logic
+ if (not accept) and (ball_center_candidate is not None) and (not from_track):
+ self.det_reject_streak += 1
+ gap_frames = (
+ (self.frame_idx - self.single_ball_trail[-1][2])
+ if len(self.single_ball_trail)
+ else 9999
+ )
+ Hmin = float(min(Hh, Ww))
+ base_gate = max(float(self.ball_gate_min), float(self.ball_gate_rel) * Hmin)
+
+ if cand_conf is not None and cand_conf >= float(self.det_override_conf):
+ x_prev, y_prev = (
+ (self.single_ball_trail[-1][0], self.single_ball_trail[-1][1])
+ if self.single_ball_trail
+ else (ball_center_candidate[0], ball_center_candidate[1])
+ )
+ dist_prev = float(
+ np.hypot(
+ ball_center_candidate[0] - x_prev,
+ ball_center_candidate[1] - y_prev,
+ )
+ )
+ if (
+ (self.det_reject_streak >= int(self.det_override_after))
+ or (gap_frames >= int(self.reacquire_frames))
+ or (dist_prev <= 2.5 * base_gate)
+ ):
+ accept = True # force accept the detection
+ else:
+ self.det_reject_streak = 0
+
+ if accept and ball_center_candidate is not None:
+ cx_raw, cy_raw = ball_center_candidate
+ alpha = float(self.ball_smooth_ema)
+ if 0.0 < alpha <= 1.0:
+ if self.ema_cxcy is None:
+ self.ema_cxcy = (float(cx_raw), float(cy_raw))
+ else:
+ self.ema_cxcy = (
+ lerp(self.ema_cxcy[0], float(cx_raw), alpha),
+ lerp(self.ema_cxcy[1], float(cy_raw), alpha),
+ )
+ cx, cy = int(round(self.ema_cxcy[0])), int(round(self.ema_cxcy[1]))
+ else:
+ cx, cy = cx_raw, cy_raw
+
+ if from_track:
+ x1, y1, x2, y2 = cand_box
+ # No drawing here, as it's engine
+
+ # Update trail
+ if (
+ len(self.single_ball_trail) >= 1
+ and self.single_ball_trail[-1][2] <= self.frame_idx - 2
+ ):
+ x_prev, y_prev, k_prev = self.single_ball_trail[-1]
+ if self.frame_idx - k_prev == 2:
+ add_trail_point(
+ self.single_ball_trail,
+ (x_prev + cx) // 2,
+ (y_prev + cy) // 2,
+ k_prev + 1,
+ densify=False,
+ )
+
+ add_trail_point(
+ self.single_ball_trail, cx, cy, self.frame_idx, densify=True, max_gap=5
+ )
+
+ if from_track:
+ self.ball_state.update_from_xyxy(cand_box, self.frame_idx)
+ else:
+ if cand_box is not None:
+ self.ball_state.update_from_center(cx, cy, self.frame_idx)
+
+ self.last_shown_box = (
+ cand_box if cand_box is not None else self.last_shown_box
+ )
+
+ if len(self.single_ball_trail) >= 2:
+ x0, y0, _ = self.single_ball_trail[-2]
+ step = float(np.hypot(cx - x0, cy - y0))
+ self.recent_speed = (
+ step
+ if self.recent_speed is None
+ else 0.8 * self.recent_speed + 0.2 * step
+ )
+
+ self.coast_streak = 0
+ self.det_reject_streak = 0
+
+ else:
+ if ball_center_candidate is not None:
+ self.dropped_by_gate += 1
+
+ if (
+ self.ball_coast
+ and len(self.single_ball_trail) > 0
+ and self.coast_streak < int(self.coast_max)
+ ):
+ pred = self.ball_state.predict(
+ self.frame_idx, decay=float(self.coast_decay)
+ )
+ if pred is not None and np.all(np.isfinite(pred)):
+ px, py = int(round(pred[0])), int(round(pred[1]))
+ x_prev, y_prev, _ = self.single_ball_trail[-1]
+ hard_cap = float(self.ball_max_jump_rel) * float(min(Hh, Ww))
+ if (
+ 0 <= px < Ww
+ and 0 <= py < Hh
+ and float(np.hypot(px - x_prev, py - y_prev)) <= 1.25 * hard_cap
+ ):
+ add_trail_point(
+ self.single_ball_trail,
+ px,
+ py,
+ self.frame_idx,
+ densify=False,
+ )
+ self.ball_state.update_from_center(px, py, self.frame_idx)
+ if len(self.single_ball_trail) >= 2:
+ step = float(np.hypot(px - x_prev, py - y_prev))
+ self.recent_speed = (
+ step
+ if self.recent_speed is None
+ else 0.8 * self.recent_speed + 0.2 * step
+ )
+ self.coast_streak += 1
+ self.coast_used += 1
+ else:
+ self.coast_streak = 0
+
+ self.frame_idx += 1
+
+ # Return result for decode
+ class AdvancedResult:
+ def __init__(self, tracks_p, tracks_b, ball_trail, boxes):
+ self.tracks_p = tracks_p
+ self.tracks_b = tracks_b
+ self.ball_trail = list(ball_trail)
+ self.boxes = boxes
+
+ return AdvancedResult(tracks_p, tracks_b, self.single_ball_trail, det_res.boxes)
diff --git a/plugins/python/engine/yolo_engine.py b/plugins/python/engine/yolo_engine.py
new file mode 100644
index 0000000..ba53ba5
--- /dev/null
+++ b/plugins/python/engine/yolo_engine.py
@@ -0,0 +1,103 @@
+# YoloEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+import time
+
+from .pytorch_engine import PyTorchEngine
+
+
+class YoloEngine(PyTorchEngine):
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from ultralytics import YOLO
+
+ self.model = YOLO(f"{model_name}.pt")
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.logger.info(f"YOLO model '{model_name}' loaded on {self.device}")
+ except Exception as e:
+ raise ValueError(f"Failed to load YOLO model '{model_name}'. Error: {e}")
+
+ def do_forward(self, frames):
+ import numpy as np
+
+ is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
+ writable_frames = np.array(frames, copy=True)
+ batch_size = writable_frames.shape[0] if is_batch else 1
+
+ model = self.get_model()
+ if model is None:
+ self.logger.error("Model is not loaded.")
+ return None if not is_batch else [None] * batch_size
+
+ try:
+ start_pre = time.time()
+ img_list = (
+ [
+ writable_frames[i] if is_batch else writable_frames
+ for i in range(batch_size)
+ ]
+ if is_batch
+ else [writable_frames]
+ )
+ self.logger.debug(
+ f"Input shape: {writable_frames.shape}, min={writable_frames.min()}, max={writable_frames.max()}"
+ )
+ end_pre = time.time()
+
+ conf = getattr(self, "conf", 0.25)
+ iou = getattr(self, "iou", 0.5)
+ agnostic = getattr(self, "agnostic_nms", True)
+ if self.track:
+ # Ensure tracker persists across batches
+ results = self.execute_with_stream(
+ lambda: model.track(
+ source=img_list,
+ persist=True,
+ imgsz=640,
+ conf=conf,
+ iou=iou,
+ agnostic_nms=agnostic,
+ verbose=True,
+ tracker="botsort.yaml",
+ )
+ )
+ else:
+ results = self.execute_with_stream(
+ lambda: model(
+ img_list,
+ imgsz=640,
+ conf=conf,
+ iou=iou,
+ agnostic_nms=agnostic,
+ verbose=True,
+ )
+ )
+ end_inf = time.time()
+
+ if results is None or (isinstance(results, list) and not results):
+ self.logger.warning("Inference returned None or empty list.")
+ return None if not is_batch else [None] * batch_size
+
+ self.logger.info(
+ f"Preprocessing: {(end_pre - start_pre)*1000:.2f} ms, Inference: {(end_inf - end_pre)*1000:.2f} ms for {batch_size} frames"
+ )
+ return results[0] if not is_batch else results
+
+ except Exception as e:
+ self.logger.error(f"Error during inference: {e}")
+ return None if not is_batch else [None] * batch_size
diff --git a/plugins/python/engine/yolo_pose_engine.py b/plugins/python/engine/yolo_pose_engine.py
new file mode 100644
index 0000000..3d33953
--- /dev/null
+++ b/plugins/python/engine/yolo_pose_engine.py
@@ -0,0 +1,59 @@
+# YoloPoseEngine
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+from .pytorch_engine import PyTorchEngine
+
+
+class YoloPoseEngine(PyTorchEngine):
+ """PyTorch engine for YOLO pose estimation models."""
+
+ def do_load_model(self, model_name, **kwargs):
+ try:
+ from ultralytics import YOLO
+
+ self.model = YOLO(f"{model_name}.pt")
+ self.execute_with_stream(lambda: self.model.to(self.device))
+ self.logger.info(f"YOLO pose model '{model_name}' loaded on {self.device}")
+ except Exception as e:
+ raise ValueError(f"Failed to load YOLO pose model '{model_name}': {e}")
+
+ def do_forward(self, frames):
+ import numpy as np
+
+ is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
+ writable = np.array(frames, copy=True)
+ batch_size = writable.shape[0] if is_batch else 1
+
+ model = self.get_model()
+ if model is None:
+ self.logger.error("Pose model not loaded")
+ return None if not is_batch else [None] * batch_size
+
+ try:
+ img_list = (
+ [writable[i] for i in range(batch_size)] if is_batch else [writable]
+ )
+ results = self.execute_with_stream(
+ lambda: model(img_list, imgsz=640, conf=0.25, verbose=False)
+ )
+ if not results:
+ return None if not is_batch else [None] * batch_size
+ return results[0] if not is_batch else results
+ except Exception as e:
+ self.logger.error(f"Pose inference error: {e}")
+ return None if not is_batch else [None] * batch_size
diff --git a/plugins/python/face.py b/plugins/python/face.py
index 4b03976..3d7d4d8 100644
--- a/plugins/python/face.py
+++ b/plugins/python/face.py
@@ -17,23 +17,22 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
import json
- import os
import gi
gi.require_version("Gst", "1.0")
gi.require_version("GstBase", "1.0")
gi.require_version("GstVideo", "1.0")
- gi.require_version("GstAnalytics", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject, GstAnalytics, GLib
+ from gi.repository import Gst
+ from backend import analytics, GObject
from base_objectdetector import BaseObjectDetector
- from engine.pytorch_engine import PyTorchEngine
+ from engine.face_engine import FaceEngine
from engine.engine_factory import EngineFactory
except ImportError as e:
@@ -44,103 +43,6 @@
FACE_META_HEADER = b"GST-FACE:"
-class FaceEngine(PyTorchEngine):
- """
- PyTorch engine for face detection and recognition using InsightFace.
-
- Supports InsightFace model packs:
- buffalo_l (large, most accurate)
- buffalo_s (small, fastest)
- buffalo_sc (small with recognition)
- """
-
- def do_load_model(self, model_name, **kwargs):
- try:
- from insightface.app import FaceAnalysis
-
- self.app = FaceAnalysis(
- name=model_name,
- providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
- )
- self.app.prepare(ctx_id=0, det_size=(640, 640))
- self.gallery = {}
- self.logger.info(f"InsightFace model '{model_name}' loaded")
- except Exception as e:
- raise ValueError(f"Failed to load InsightFace model '{model_name}': {e}")
-
- def load_gallery(self, gallery_path):
- """Load known face embeddings from a directory of images."""
- import numpy as np
- from PIL import Image
-
- if not gallery_path or not os.path.isdir(gallery_path):
- return
-
- self.gallery = {}
- for fname in os.listdir(gallery_path):
- fpath = os.path.join(gallery_path, fname)
- if not os.path.isfile(fpath):
- continue
- try:
- img = np.array(Image.open(fpath).convert("RGB"))
- faces = self.app.get(img)
- if faces:
- name = os.path.splitext(fname)[0]
- self.gallery[name] = faces[0].embedding
- self.logger.info(f"Loaded gallery face: {name}")
- except Exception as e:
- self.logger.warning(f"Failed to load gallery image '{fname}': {e}")
-
- def do_forward(self, frames, threshold=0.5):
- import numpy as np
-
- is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
- if not is_batch:
- frames = frames[np.newaxis]
-
- results = []
- for frame in frames:
- try:
- faces = self.app.get(frame.astype(np.uint8))
- detections = []
- for face in faces:
- bbox = face.bbox.astype(float).tolist()
- score = float(face.det_score)
- embedding = face.embedding
-
- identity = "unknown"
- best_sim = 0.0
- if self.gallery and embedding is not None:
- for name, gallery_emb in self.gallery.items():
- sim = float(
- np.dot(embedding, gallery_emb)
- / (
- np.linalg.norm(embedding)
- * np.linalg.norm(gallery_emb)
- + 1e-8
- )
- )
- if sim > best_sim:
- best_sim = sim
- if sim >= threshold:
- identity = name
-
- detections.append(
- {
- "bbox": bbox,
- "score": score,
- "identity": identity,
- "similarity": best_sim,
- }
- )
- results.append(detections)
- except Exception as e:
- self.logger.error(f"Face inference error on frame: {e}")
- results.append([])
-
- return results[0] if not is_batch else results
-
-
class FaceTransform(BaseObjectDetector):
"""
GStreamer element for face detection and recognition using InsightFace.
@@ -204,7 +106,7 @@ def do_decode(self, buf, result, stream_idx=0):
self.logger.info(f"Stream {stream_idx}: no faces detected")
return
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
+ meta = analytics.add_relation_meta(buf)
if not meta:
self.logger.error("Failed to add analytics relation metadata")
return
@@ -217,16 +119,16 @@ def do_decode(self, buf, result, stream_idx=0):
identity = det["identity"]
similarity = det["similarity"]
- qk = GLib.quark_from_string(f"stream_{stream_idx}_face_{identity}")
- ret, _ = meta.add_od_mtd(
- qk,
+ mtd = analytics.add_object(
+ meta,
+ f"stream_{stream_idx}_face_{identity}",
x1,
y1,
x2 - x1,
y2 - y1,
score,
)
- if not ret:
+ if mtd is None:
self.logger.error(f"Failed to add od_mtd for face {i}")
continue
@@ -251,10 +153,9 @@ def do_decode(self, buf, result, stream_idx=0):
)
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(FaceTransform)
- __gstelementfactory__ = ("pyml_face", Gst.Rank.NONE, FaceTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_face", FaceTransform)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_face' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/football_analyzer.py b/plugins/python/football_analyzer.py
new file mode 100644
index 0000000..1f4190c
--- /dev/null
+++ b/plugins/python/football_analyzer.py
@@ -0,0 +1,946 @@
+# FootballAnalyzer
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+import os
+import pickle
+
+from log.global_logger import GlobalLogger
+import backend
+
+CAN_REGISTER_ELEMENT = True
+try:
+ import gi
+
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ gi.require_version("GstVideo", "1.0")
+ from gi.repository import Gst, GstBase, GstVideo # noqa: E402
+ from backend import GObject # noqa: E402
+
+ # Define caps before the optional heavy imports so the element's pad
+ # templates still resolve when an optional dep (e.g. supervision) is missing;
+ # only registration is then skipped (CAN_REGISTER_ELEMENT=False).
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ VIDEO_CAPS = Gst.Caps.from_string("video/x-raw, format=BGR")
+
+ from log.logger_factory import LoggerFactory # noqa: E402
+
+except ImportError as e:
+ CAN_REGISTER_ELEMENT = False
+ GlobalLogger().warning(
+ f"The 'pyml_football_analyzer' element will not be available. Error: {e}"
+ )
+
+
+def get_center_of_bbox(bbox):
+ x1, y1, x2, y2 = bbox
+ return int((x1 + x2) / 2), int((y1 + y2) / 2)
+
+
+def get_bbox_width(bbox):
+ return bbox[2] - bbox[0]
+
+
+class Tracker:
+ """Tracker."""
+
+ def __init__(self, model_path):
+ import cv2
+ import supervision as sv
+ from ultralytics import YOLO
+
+ self.model = YOLO(model_path)
+ self.tracker = sv.ByteTrack()
+ self.sift = cv2.SIFT_create()
+ self.matcher = cv2.BFMatcher(cv2.NORM_L2)
+
+ def _foreground_mask(self, shape, frame_tracks, dilation=15):
+ import numpy as np
+
+ h, w = shape[:2]
+ mask = np.full((h, w), 255, dtype=np.uint8)
+ bboxes = []
+ for key in ("players", "referees", "ball"):
+ for obj in frame_tracks.get(key, {}).values():
+ bboxes.append(obj["bbox"])
+ for bbox in bboxes:
+ x1 = max(0, int(bbox[0]) - dilation)
+ y1 = max(0, int(bbox[1]) - dilation)
+ x2 = min(w, int(bbox[2]) + dilation)
+ y2 = min(h, int(bbox[3]) + dilation)
+ mask[y1:y2, x1:x2] = 0
+ return mask
+
+ def get_camera_motion(
+ self,
+ frames,
+ tracks,
+ read_from_stub=False,
+ stub_path=None,
+ ratio=0.75,
+ ransac_thresh=3.0,
+ min_matches=8,
+ ):
+ import cv2
+ import numpy as np
+
+ if read_from_stub and stub_path is not None and os.path.exists(stub_path):
+ with open(stub_path, "rb") as f:
+ return pickle.load(f)
+
+ cumulative = [np.eye(3, dtype=np.float64)]
+ prev_gray = cv2.cvtColor(frames[0], cv2.COLOR_BGR2GRAY)
+ prev_mask = self._foreground_mask(
+ frames[0].shape, {k: tracks[k][0] for k in tracks}
+ )
+ prev_kp, prev_desc = self.sift.detectAndCompute(prev_gray, prev_mask)
+
+ for i in range(1, len(frames)):
+ curr_gray = cv2.cvtColor(frames[i], cv2.COLOR_BGR2GRAY)
+ curr_mask = self._foreground_mask(
+ frames[i].shape, {k: tracks[k][i] for k in tracks}
+ )
+ curr_kp, curr_desc = self.sift.detectAndCompute(curr_gray, curr_mask)
+
+ H_step = np.eye(3, dtype=np.float64)
+ if (
+ prev_desc is not None
+ and curr_desc is not None
+ and len(prev_desc) >= 2
+ and len(curr_desc) >= 2
+ ):
+ knn = self.matcher.knnMatch(prev_desc, curr_desc, k=2)
+ good = [
+ m
+ for pair in knn
+ if len(pair) == 2
+ for m, n in [pair]
+ if m.distance < ratio * n.distance
+ ]
+ if len(good) >= min_matches:
+ pts_prev = np.float32(
+ [prev_kp[m.queryIdx].pt for m in good]
+ ).reshape(-1, 1, 2)
+ pts_curr = np.float32(
+ [curr_kp[m.trainIdx].pt for m in good]
+ ).reshape(-1, 1, 2)
+ H, _ = cv2.findHomography(
+ pts_prev, pts_curr, cv2.RANSAC, ransac_thresh
+ )
+ if H is not None:
+ H_step = H
+
+ cumulative.append(H_step @ cumulative[-1])
+ prev_kp, prev_desc = curr_kp, curr_desc
+
+ if stub_path is not None:
+ with open(stub_path, "wb") as f:
+ pickle.dump(cumulative, f)
+ return cumulative
+
+ def detect_frames(self, frames):
+ batch_size = 20
+ detections = []
+ for i in range(0, len(frames), batch_size):
+ detections_batch = self.model.predict(frames[i : i + batch_size], conf=0.1)
+ detections += detections_batch
+ return detections
+
+ def get_object_tracks(self, frames, read_from_stub=False, stub_path=None):
+ import supervision as sv
+
+ if read_from_stub and stub_path is not None and os.path.exists(stub_path):
+ with open(stub_path, "rb") as f:
+ tracks = pickle.load(f)
+ return tracks
+
+ detections = self.detect_frames(frames)
+ tracks = {"players": [], "referees": [], "ball": []}
+
+ per_frame = []
+ class_votes = {}
+ for detection in detections:
+ cls_names = detection.names
+ cls_names_inv = {v: k for k, v in cls_names.items()}
+
+ detection_supervision = sv.Detections.from_ultralytics(detection)
+
+ for object_ind, class_id in enumerate(detection_supervision.class_id):
+ if cls_names[class_id] == "goalkeeper":
+ detection_supervision.class_id[object_ind] = cls_names_inv["player"]
+
+ tracked = self.tracker.update_with_detections(detection_supervision)
+ per_frame.append((tracked, detection_supervision, cls_names, cls_names_inv))
+
+ for fd in tracked:
+ cls_name = cls_names[fd[3]]
+ track_id = fd[4]
+ if cls_name in ("player", "referee"):
+ v = class_votes.setdefault(track_id, {"player": 0, "referee": 0})
+ v[cls_name] += 1
+
+ track_class = {
+ tid: ("player" if v["player"] >= v["referee"] else "referee")
+ for tid, v in class_votes.items()
+ }
+
+ for frame_num, (tracked, raw_detections, cls_names, cls_names_inv) in enumerate(
+ per_frame
+ ):
+ tracks["players"].append({})
+ tracks["referees"].append({})
+ tracks["ball"].append({})
+
+ for fd in tracked:
+ bbox = fd[0].tolist()
+ track_id = fd[4]
+ stable_cls = track_class.get(track_id)
+ if stable_cls == "player":
+ tracks["players"][frame_num][track_id] = {"bbox": bbox}
+ elif stable_cls == "referee":
+ tracks["referees"][frame_num][track_id] = {"bbox": bbox}
+
+ for fd in raw_detections:
+ if fd[3] == cls_names_inv["ball"]:
+ tracks["ball"][frame_num][1] = {"bbox": fd[0].tolist()}
+
+ if stub_path is not None:
+ with open(stub_path, "wb") as f:
+ pickle.dump(tracks, f)
+
+ return tracks
+
+ def draw_ellipse(self, frame, bbox, color, track_id=None):
+ import cv2
+
+ y2 = int(bbox[3])
+ x_center, _ = get_center_of_bbox(bbox)
+ width = get_bbox_width(bbox)
+
+ cv2.ellipse(
+ frame,
+ center=(x_center, y2),
+ axes=(int(width), int(0.35 * width)),
+ angle=0.0,
+ startAngle=-45,
+ endAngle=235,
+ color=color,
+ thickness=2,
+ lineType=cv2.LINE_4,
+ )
+
+ rectangle_width = 40
+ rectangle_height = 20
+ x1_rect = x_center - rectangle_width // 2
+ x2_rect = x_center + rectangle_width // 2
+ y1_rect = (y2 - rectangle_height // 2) + 15
+ y2_rect = (y2 + rectangle_height // 2) + 15
+
+ if track_id is not None:
+ cv2.rectangle(
+ frame,
+ (int(x1_rect), int(y1_rect)),
+ (int(x2_rect), int(y2_rect)),
+ color,
+ cv2.FILLED,
+ )
+ x1_text = x1_rect + 12
+ if track_id > 99:
+ x1_text -= 10
+ cv2.putText(
+ frame,
+ f"{track_id}",
+ (int(x1_text), int(y1_rect + 15)),
+ cv2.FONT_HERSHEY_SIMPLEX,
+ 0.6,
+ (0, 0, 0),
+ 2,
+ )
+ return frame
+
+ def draw_traingle(self, frame, bbox, color):
+ import cv2
+ import numpy as np
+
+ y = int(bbox[1])
+ x, _ = get_center_of_bbox(bbox)
+ triangle_points = np.array(
+ [
+ [x, y],
+ [x - 10, y - 20],
+ [x + 10, y - 20],
+ ]
+ )
+ cv2.drawContours(frame, [triangle_points], 0, color, cv2.FILLED)
+ cv2.drawContours(frame, [triangle_points], 0, (0, 0, 0), 2)
+ return frame
+
+ def classify_jersey(self, frame, bbox):
+ import cv2
+
+ x1, y1, x2, y2 = [int(v) for v in bbox]
+ h_box, w_box = y2 - y1, x2 - x1
+ if h_box <= 0 or w_box <= 0:
+ return None
+ jy1 = y1 + int(0.15 * h_box)
+ jy2 = y1 + int(0.55 * h_box)
+ jx1 = x1 + int(0.25 * w_box)
+ jx2 = x1 + int(0.75 * w_box)
+ H, W = frame.shape[:2]
+ jy1, jy2 = max(0, jy1), min(H, jy2)
+ jx1, jx2 = max(0, jx1), min(W, jx2)
+ if jy2 - jy1 < 3 or jx2 - jx1 < 3:
+ return None
+ patch = frame[jy1:jy2, jx1:jx2]
+ hsv = cv2.cvtColor(patch, cv2.COLOR_BGR2HSV)
+ s_v = (hsv[..., 1] > 80) & (hsv[..., 2] > 50)
+ h = hsv[..., 0]
+ red = (((h <= 10) | (h >= 170)) & s_v).sum()
+ blue = ((h >= 100) & (h <= 130) & s_v).sum()
+ min_pixels = max(20, int(0.02 * patch.shape[0] * patch.shape[1]))
+ if red < min_pixels and blue < min_pixels:
+ return None
+ return "red" if red >= blue else "blue"
+
+ def _ref_bottom_center(self, bbox, H_inv):
+ import cv2
+ import numpy as np
+
+ xc, _ = get_center_of_bbox(bbox)
+ yb = int(bbox[3])
+ pt = cv2.perspectiveTransform(np.array([[[xc, yb]]], dtype=np.float32), H_inv)[
+ 0
+ ][0]
+ return float(pt[0]), float(pt[1])
+
+ def _minimap_extent(self, tracks, camera_motion):
+ import numpy as np
+
+ xs, ys = [], []
+ n = len(tracks["players"])
+ for i in range(n):
+ H_inv = (
+ np.linalg.inv(camera_motion[i])
+ if camera_motion is not None
+ else np.eye(3)
+ )
+ for key in ("players", "referees"):
+ for p in tracks[key][i].values():
+ x, y = self._ref_bottom_center(p["bbox"], H_inv)
+ xs.append(x)
+ ys.append(y)
+ if not xs:
+ return None
+ min_x, max_x = min(xs), max(xs)
+ min_y, max_y = min(ys), max(ys)
+ pad_x = 0.05 * max(1.0, max_x - min_x)
+ pad_y = 0.05 * max(1.0, max_y - min_y)
+ return min_x - pad_x, min_y - pad_y, max_x + pad_x, max_y + pad_y
+
+ def _make_minimap_bg(self, mm_w, mm_h):
+ import cv2
+ import numpy as np
+
+ bg = np.full((mm_h, mm_w, 3), (40, 110, 40), dtype=np.uint8)
+ cv2.rectangle(bg, (2, 2), (mm_w - 3, mm_h - 3), (240, 240, 240), 2)
+ cv2.line(bg, (mm_w // 2, 2), (mm_w // 2, mm_h - 3), (240, 240, 240), 1)
+ cv2.circle(bg, (mm_w // 2, mm_h // 2), max(10, mm_h // 8), (240, 240, 240), 1)
+ return bg
+
+ def _project_to_minimap(self, extent, mm_w, mm_h, x, y):
+ min_x, min_y, max_x, max_y = extent
+ dx = max(1e-6, max_x - min_x)
+ dy = max(1e-6, max_y - min_y)
+ scale = min((mm_w - 10) / dx, (mm_h - 10) / dy)
+ off_x = (mm_w - scale * dx) / 2.0
+ off_y = (mm_h - scale * dy) / 2.0
+ return int(off_x + (x - min_x) * scale), int(off_y + (y - min_y) * scale)
+
+ def _smooth_points(self, pts, window):
+ import numpy as np
+
+ if window <= 1 or len(pts) < 2:
+ return pts
+ pts = np.asarray(pts, dtype=np.float32)
+ n = len(pts)
+ half = window // 2
+ smoothed = np.empty_like(pts)
+ for i in range(n):
+ lo = max(0, i - half)
+ hi = min(n, i + half + 1)
+ smoothed[i] = pts[lo:hi].mean(axis=0)
+ return smoothed
+
+ def draw_trail(self, frame, points, color):
+ import cv2
+ import numpy as np
+
+ if len(points) < 2:
+ return frame
+ pts = np.array(points, dtype=np.int32).reshape(-1, 1, 2)
+ cv2.polylines(
+ frame, [pts], isClosed=False, color=color, thickness=2, lineType=cv2.LINE_AA
+ )
+ return frame
+
+ def _point_to_bbox_distance(self, px, py, bbox):
+ import numpy as np
+
+ x1, y1, x2, y2 = bbox
+ dx = max(x1 - px, 0.0, px - x2)
+ dy = max(y1 - py, 0.0, py - y2)
+ return float(np.hypot(dx, dy))
+
+ def _ball_contact(self, player_dict, ball_bbox, contact_pad_ratio):
+ bx, by = get_center_of_bbox(ball_bbox)
+ best_tid, best_d, best_bbox = None, float("inf"), None
+ for tid, player in player_dict.items():
+ d = self._point_to_bbox_distance(bx, by, player["bbox"])
+ if d < best_d:
+ best_tid, best_d, best_bbox = tid, d, player["bbox"]
+ if best_bbox is None:
+ return None
+ w_box = best_bbox[2] - best_bbox[0]
+ h_box = best_bbox[3] - best_bbox[1]
+ if best_d > contact_pad_ratio * max(w_box, h_box):
+ return None
+ return best_tid
+
+ def _count_total_contacts(self, tracks, contact_gap_frames, contact_pad_ratio):
+ totals = {}
+ last_contact_frame = {}
+ for frame_num, (player_dict, ball_dict) in enumerate(
+ zip(tracks["players"], tracks["ball"])
+ ):
+ ball = ball_dict.get(1)
+ if ball is None or not player_dict:
+ continue
+ tid = self._ball_contact(player_dict, ball["bbox"], contact_pad_ratio)
+ if tid is None:
+ continue
+ last = last_contact_frame.get(tid)
+ if last is None or (frame_num - last) > contact_gap_frames:
+ totals[tid] = totals.get(tid, 0) + 1
+ last_contact_frame[tid] = frame_num
+ return totals
+
+ def draw_player_hud(
+ self, frame, player_id, contacts, distance_m, color, headshot=None
+ ):
+ import cv2
+
+ x, y = 10, 10
+ bg_color = (131, 41, 92)
+ text_color = (47, 186, 64)
+ if headshot is not None:
+ hh, hw = headshot.shape[:2]
+ w, h = hw + 280, max(110, hh + 20)
+ text_x = x + hw + 20
+ else:
+ w, h = 320, 100
+ text_x = x + 12
+ cv2.rectangle(frame, (x, y), (x + w, y + h), bg_color, cv2.FILLED)
+ cv2.rectangle(frame, (x, y), (x + w, y + h), color, 2)
+ if headshot is not None:
+ hy, hx = y + 10, x + 10
+ frame[hy : hy + headshot.shape[0], hx : hx + headshot.shape[1]] = headshot
+ cv2.rectangle(
+ frame,
+ (hx, hy),
+ (hx + headshot.shape[1], hy + headshot.shape[0]),
+ color,
+ 2,
+ )
+ cv2.putText(
+ frame,
+ "Player #8",
+ (text_x, y + 28),
+ cv2.FONT_HERSHEY_SIMPLEX,
+ 0.7,
+ text_color,
+ 2,
+ )
+ cv2.putText(
+ frame,
+ f"Ball contacts: {contacts}",
+ (text_x, y + 58),
+ cv2.FONT_HERSHEY_SIMPLEX,
+ 0.6,
+ text_color,
+ 1,
+ )
+ cv2.putText(
+ frame,
+ f"Distance: {distance_m:.1f} m",
+ (text_x, y + 85),
+ cv2.FONT_HERSHEY_SIMPLEX,
+ 0.6,
+ text_color,
+ 1,
+ )
+ return frame
+
+ def draw_annotations(
+ self,
+ video_frames,
+ tracks,
+ camera_motion=None,
+ trail_length=30,
+ contact_gap_frames=5,
+ contact_pad_ratio=0.25,
+ player_height_m=1.8,
+ headshot_path=None,
+ headshot_size=90,
+ logo_path=None,
+ logo_height=80,
+ logo_margin=15,
+ trail_smooth_window=11,
+ show_minimap=True,
+ minimap_size=(320, 200),
+ minimap_margin=15,
+ ):
+ import cv2
+ import numpy as np
+
+ output_video_frames = []
+ player_trails = {}
+ team_votes = {}
+ team_bgr = {"red": (0, 0, 255), "blue": (255, 0, 0)}
+ default_color = (200, 200, 200)
+
+ frames_count = {}
+ for frame_players in tracks["players"]:
+ for tid in frame_players:
+ frames_count[tid] = frames_count.get(tid, 0) + 1
+ total_contacts = self._count_total_contacts(
+ tracks, contact_gap_frames, contact_pad_ratio
+ )
+
+ heights = [
+ p["bbox"][3] - p["bbox"][1]
+ for frame_players in tracks["players"]
+ for p in frame_players.values()
+ if p["bbox"][3] > p["bbox"][1]
+ ]
+ px_per_meter = float(np.median(heights)) / player_height_m if heights else 1.0
+
+ headshot = None
+ if headshot_path is not None and os.path.exists(headshot_path):
+ img = cv2.imread(headshot_path)
+ if img is not None:
+ headshot = cv2.resize(
+ img, (headshot_size, headshot_size), interpolation=cv2.INTER_AREA
+ )
+
+ logo_bgr, logo_alpha = None, None
+ if logo_path is not None and os.path.exists(logo_path):
+ img = cv2.imread(logo_path, cv2.IMREAD_UNCHANGED)
+ if img is not None:
+ scale = logo_height / img.shape[0]
+ new_w = max(1, int(round(img.shape[1] * scale)))
+ img = cv2.resize(
+ img, (new_w, logo_height), interpolation=cv2.INTER_LANCZOS4
+ )
+ if img.ndim == 3 and img.shape[2] == 4:
+ logo_bgr = img[..., :3]
+ logo_alpha = (img[..., 3:4].astype(np.float32)) / 255.0
+ else:
+ logo_bgr = (
+ img if img.ndim == 3 else cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+ )
+
+ minimap_bg, minimap_extent = None, None
+ if show_minimap:
+ minimap_extent = self._minimap_extent(tracks, camera_motion)
+ if minimap_extent is not None:
+ minimap_bg = self._make_minimap_bg(minimap_size[0], minimap_size[1])
+
+ if total_contacts:
+ focal_tid = max(
+ total_contacts,
+ key=lambda t: (total_contacts[t], frames_count.get(t, 0)),
+ )
+ elif frames_count:
+ focal_tid = max(frames_count, key=frames_count.get)
+ else:
+ focal_tid = None
+
+ last_ref_pt = {}
+ player_distance = {}
+ player_contacts = {}
+ last_contact_frame = {}
+ for frame_num, frame in enumerate(video_frames):
+ frame = frame.copy()
+
+ player_dict = tracks["players"][frame_num]
+ ball_dict = tracks["ball"][frame_num]
+ referee_dict = tracks["referees"][frame_num]
+
+ H_cum = camera_motion[frame_num] if camera_motion is not None else np.eye(3)
+ H_inv = np.linalg.inv(H_cum)
+
+ active_ids = set(player_dict.keys())
+ for track_id, player in player_dict.items():
+ x_center, _ = get_center_of_bbox(player["bbox"])
+ y_bottom = int(player["bbox"][3])
+ ref_pt = cv2.perspectiveTransform(
+ np.array([[[x_center, y_bottom]]], dtype=np.float32), H_inv
+ )[0][0]
+ ref_tuple = (float(ref_pt[0]), float(ref_pt[1]))
+ player_trails.setdefault(track_id, []).append(ref_tuple)
+ if len(player_trails[track_id]) > trail_length:
+ player_trails[track_id] = player_trails[track_id][-trail_length:]
+
+ if track_id in last_ref_pt:
+ dx = ref_tuple[0] - last_ref_pt[track_id][0]
+ dy = ref_tuple[1] - last_ref_pt[track_id][1]
+ player_distance[track_id] = player_distance.get(
+ track_id, 0.0
+ ) + float(np.hypot(dx, dy))
+ last_ref_pt[track_id] = ref_tuple
+
+ vote = self.classify_jersey(frame, player["bbox"])
+ if vote is not None:
+ counts = team_votes.setdefault(track_id, {"red": 0, "blue": 0})
+ counts[vote] += 1
+ for track_id in list(player_trails.keys()):
+ if track_id not in active_ids:
+ del player_trails[track_id]
+ last_ref_pt.pop(track_id, None)
+
+ ball = ball_dict.get(1)
+ if ball is not None and player_dict:
+ tid = self._ball_contact(player_dict, ball["bbox"], contact_pad_ratio)
+ if tid is not None:
+ last = last_contact_frame.get(tid)
+ if last is None or (frame_num - last) > contact_gap_frames:
+ player_contacts[tid] = player_contacts.get(tid, 0) + 1
+ last_contact_frame[tid] = frame_num
+
+ focal_color = (131, 41, 92)
+
+ def color_for(track_id):
+ if track_id == focal_tid:
+ return focal_color
+ counts = team_votes.get(track_id)
+ if not counts or (counts["red"] == 0 and counts["blue"] == 0):
+ return default_color
+ return (
+ team_bgr["red"]
+ if counts["red"] >= counts["blue"]
+ else team_bgr["blue"]
+ )
+
+ for track_id, ref_points in player_trails.items():
+ smoothed_ref = self._smooth_points(ref_points, trail_smooth_window)
+ pts = cv2.perspectiveTransform(
+ np.asarray(smoothed_ref, dtype=np.float32).reshape(-1, 1, 2), H_cum
+ ).reshape(-1, 2)
+ frame = self.draw_trail(frame, pts.tolist(), color_for(track_id))
+
+ for track_id, player in player_dict.items():
+ frame = self.draw_ellipse(frame, player["bbox"], color_for(track_id))
+
+ for _, referee in referee_dict.items():
+ frame = self.draw_ellipse(frame, referee["bbox"], (0, 255, 255))
+
+ for track_id, ball in ball_dict.items():
+ frame = self.draw_traingle(frame, ball["bbox"], (0, 255, 0))
+
+ if focal_tid is not None:
+ frame = self.draw_player_hud(
+ frame,
+ focal_tid,
+ player_contacts.get(focal_tid, 0),
+ player_distance.get(focal_tid, 0.0) / px_per_meter,
+ color_for(focal_tid),
+ headshot=headshot,
+ )
+
+ if logo_bgr is not None:
+ lh, lw = logo_bgr.shape[:2]
+ fh, fw = frame.shape[:2]
+ x0 = max(0, fw - lw - logo_margin)
+ y0 = logo_margin
+ x1, y1 = x0 + lw, y0 + lh
+ if logo_alpha is not None:
+ roi = frame[y0:y1, x0:x1].astype(np.float32)
+ blended = (
+ roi * (1.0 - logo_alpha)
+ + logo_bgr.astype(np.float32) * logo_alpha
+ )
+ frame[y0:y1, x0:x1] = blended.astype(np.uint8)
+ else:
+ frame[y0:y1, x0:x1] = logo_bgr
+
+ if minimap_bg is not None and minimap_extent is not None:
+ mm = minimap_bg.copy()
+ mm_w, mm_h = minimap_size
+ for tid, player in player_dict.items():
+ rx, ry = self._ref_bottom_center(player["bbox"], H_inv)
+ mx, my = self._project_to_minimap(
+ minimap_extent, mm_w, mm_h, rx, ry
+ )
+ dot_color = color_for(tid)
+ radius = 6 if tid == focal_tid else 4
+ cv2.circle(mm, (mx, my), radius, dot_color, cv2.FILLED)
+ cv2.circle(mm, (mx, my), radius, (0, 0, 0), 1)
+ for referee in referee_dict.values():
+ rx, ry = self._ref_bottom_center(referee["bbox"], H_inv)
+ mx, my = self._project_to_minimap(
+ minimap_extent, mm_w, mm_h, rx, ry
+ )
+ cv2.circle(mm, (mx, my), 3, (0, 255, 255), cv2.FILLED)
+ cv2.circle(mm, (mx, my), 3, (0, 0, 0), 1)
+ ball = ball_dict.get(1)
+ if ball is not None:
+ bx, by = get_center_of_bbox(ball["bbox"])
+ bref = cv2.perspectiveTransform(
+ np.array([[[bx, by]]], dtype=np.float32), H_inv
+ )[0][0]
+ mx, my = self._project_to_minimap(
+ minimap_extent, mm_w, mm_h, float(bref[0]), float(bref[1])
+ )
+ cv2.circle(mm, (mx, my), 4, (0, 255, 0), cv2.FILLED)
+ cv2.circle(mm, (mx, my), 4, (0, 0, 0), 1)
+ fh, fw = frame.shape[:2]
+ x0 = max(0, fw - mm_w - minimap_margin)
+ y0 = max(0, fh - mm_h - minimap_margin)
+ frame[y0 : y0 + mm_h, x0 : x0 + mm_w] = mm
+
+ output_video_frames.append(frame)
+
+ return output_video_frames
+
+
+class FootballAnalyzer(GstBase.BaseTransform):
+ """
+ Buffers every incoming video frame, then on EOS runs the full batch
+ pipeline (YOLO detection, ByteTrack with whole-clip class voting,
+ SIFT/RANSAC camera motion, annotated drawing with trails / HUD /
+ logo / minimap) and pushes the annotated frames downstream before
+ forwarding EOS.
+ """
+
+ __gstmetadata__ = (
+ "Football Analyzer",
+ "Filter/Effect/Video",
+ "Runs football_analysis (YOLO + ByteTrack + SIFT camera motion + "
+ "annotated drawing) on the full clip and emits annotated frames on EOS",
+ "Marcus Edel ",
+ )
+
+ if backend.BACKEND == "gst":
+ src_template = Gst.PadTemplate.new(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ VIDEO_CAPS.copy(),
+ )
+ sink_template = Gst.PadTemplate.new(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.ALWAYS,
+ VIDEO_CAPS.copy(),
+ )
+ __gsttemplates__ = (src_template, sink_template)
+
+ model_path = GObject.Property(
+ type=str,
+ default="",
+ nick="Model Path",
+ blurb="Path to the YOLO weights (must be set before processing)",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ headshot_path = GObject.Property(
+ type=str,
+ default="",
+ nick="Headshot Path",
+ blurb="Optional headshot image for the focal-player HUD",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ logo_path = GObject.Property(
+ type=str,
+ default="",
+ nick="Logo Path",
+ blurb="Optional top-right logo overlay",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ tracks_stub_path = GObject.Property(
+ type=str,
+ default="",
+ nick="Tracks Stub Path",
+ blurb="Optional pickle path for cached object tracks (read & written)",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ camera_motion_stub_path = GObject.Property(
+ type=str,
+ default="",
+ nick="Camera Motion Stub Path",
+ blurb="Optional pickle path for cached camera-motion homographies (read & written)",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ show_minimap = GObject.Property(
+ type=bool,
+ default=True,
+ nick="Show Minimap",
+ blurb="Render the bottom-right minimap overlay",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ def __init__(self):
+ super().__init__()
+ self.logger = LoggerFactory.get(LoggerFactory.LOGGER_TYPE_GST)
+ self._frames = []
+ self._pts = []
+ self._duration = []
+ self._width = 0
+ self._height = 0
+ self._tracker = None
+
+ def _ensure_tracker(self):
+ if self._tracker is not None:
+ return self._tracker
+ if not self.model_path or not os.path.exists(self.model_path):
+ raise FileNotFoundError(f"YOLO model not found: {self.model_path!r}")
+ self.logger.info(f"Loading FootballAnalyzer Tracker from {self.model_path}")
+ self._tracker = Tracker(self.model_path)
+ return self._tracker
+
+ def do_set_caps(self, incaps, outcaps):
+ info = GstVideo.VideoInfo.new_from_caps(incaps)
+ self._width = info.width
+ self._height = info.height
+ return True
+
+ def do_transform_ip(self, buf):
+ import numpy as np
+
+ try:
+ ok, mapinfo = buf.map(Gst.MapFlags.READ)
+ if not ok:
+ self.logger.error("Failed to map incoming buffer for read")
+ return Gst.FlowReturn.ERROR
+ try:
+ frame = (
+ np.frombuffer(mapinfo.data, dtype=np.uint8)
+ .reshape(self._height, self._width, 3)
+ .copy()
+ )
+ finally:
+ buf.unmap(mapinfo)
+
+ self._frames.append(frame)
+ self._pts.append(buf.pts)
+ self._duration.append(buf.duration)
+ return Gst.FlowReturn.OK
+
+ except Exception as e:
+ self.logger.error(f"FootballAnalyzer chain error: {e}")
+ return Gst.FlowReturn.ERROR
+
+ def do_sink_event(self, event):
+ if event.type == Gst.EventType.EOS:
+ try:
+ self._run_pipeline_and_push()
+ except Exception as e:
+ self.logger.error(f"FootballAnalyzer EOS processing failed: {e}")
+ # Forward EOS regardless so the pipeline shuts down cleanly.
+ return GstBase.BaseTransform.do_sink_event(self, event)
+
+ def _run_pipeline_and_push(self):
+ import numpy as np
+
+ if not self._frames:
+ self.logger.info("FootballAnalyzer: no frames buffered, skipping")
+ return
+
+ tracker = self._ensure_tracker()
+ n = len(self._frames)
+ self.logger.info(f"FootballAnalyzer: running pipeline on {n} frames")
+
+ tracks_stub = self.tracks_stub_path or None
+ cam_stub = self.camera_motion_stub_path or None
+ headshot = self.headshot_path or None
+ logo = self.logo_path or None
+
+ tracks = tracker.get_object_tracks(
+ self._frames,
+ read_from_stub=tracks_stub is not None and os.path.exists(tracks_stub),
+ stub_path=tracks_stub,
+ )
+ camera_motion = tracker.get_camera_motion(
+ self._frames,
+ tracks,
+ read_from_stub=cam_stub is not None and os.path.exists(cam_stub),
+ stub_path=cam_stub,
+ )
+ annotated = tracker.draw_annotations(
+ self._frames,
+ tracks,
+ camera_motion=camera_motion,
+ headshot_path=headshot,
+ logo_path=logo,
+ show_minimap=self.show_minimap,
+ )
+
+ if len(annotated) != n:
+ self.logger.warning(
+ f"draw_annotations returned {len(annotated)} frames for {n} inputs; "
+ "padding/truncating to match"
+ )
+ if len(annotated) < n:
+ annotated = list(annotated) + [annotated[-1]] * (n - len(annotated))
+ else:
+ annotated = annotated[:n]
+
+ srcpad = self.srcpad
+ for i, out in enumerate(annotated):
+ data = np.ascontiguousarray(out, dtype=np.uint8).tobytes()
+ outbuf = Gst.Buffer.new_allocate(None, len(data), None)
+ outbuf.fill(0, data)
+ outbuf.pts = self._pts[i]
+ outbuf.duration = self._duration[i]
+ ret = srcpad.push(outbuf)
+ if ret != Gst.FlowReturn.OK:
+ self.logger.error(
+ f"Pushing annotated frame {i} failed with {ret}; aborting"
+ )
+ break
+
+ self._frames.clear()
+ self._pts.clear()
+ self._duration.clear()
+
+
+# GStreamer factory registration runs only under the gst backend.
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_football_analyzer", FootballAnalyzer
+ )
+elif not CAN_REGISTER_ELEMENT:
+ GlobalLogger().warning(
+ "The 'pyml_football_analyzer' element will not be registered because "
+ "required modules are missing."
+ )
diff --git a/plugins/python/football_overlay.py b/plugins/python/football_overlay.py
new file mode 100644
index 0000000..430d1c9
--- /dev/null
+++ b/plugins/python/football_overlay.py
@@ -0,0 +1,1162 @@
+# FootballOverlay
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+import os
+
+from log.global_logger import GlobalLogger
+import backend
+
+CAN_REGISTER_ELEMENT = True
+try:
+ import re
+ import gi
+
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ gi.require_version("GstVideo", "1.0")
+ gi.require_version("GstAnalytics", "1.0")
+ gi.require_version("GLib", "2.0")
+ from gi.repository import (
+ Gst,
+ GstBase,
+ GstVideo,
+ GstAnalytics,
+ GLib,
+ ) # noqa: E402
+ from backend import GObject # noqa: E402
+
+ from log.logger_factory import LoggerFactory # noqa: E402
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ OVERLAY_CAPS = Gst.Caps.from_string(
+ "video/x-raw, format=(string){ RGBA, ARGB, BGRA, ABGR }"
+ )
+
+except ImportError as e:
+ CAN_REGISTER_ELEMENT = False
+ GlobalLogger().warning(
+ f"The 'pyml_football_overlay' element will not be available. Error: {e}"
+ )
+
+
+_FORMAT_ORDER = {
+ "RGBA": (0, 1, 2, 3),
+ "ARGB": (3, 0, 1, 2),
+ "BGRA": (2, 1, 0, 3),
+ "ABGR": (3, 2, 1, 0),
+}
+
+_PALETTE = [
+ (239, 71, 111, 255),
+ (255, 209, 102, 255),
+ (6, 214, 160, 255),
+ (17, 138, 178, 255),
+ (255, 107, 107, 255),
+ (78, 205, 196, 255),
+ (199, 125, 255, 255),
+ (255, 159, 28, 255),
+ (46, 196, 182, 255),
+ (118, 200, 247, 255),
+]
+
+_REFEREE_RGBA = (255, 215, 0, 255)
+_BALL_RGBA = (0, 230, 0, 255)
+_PLAYER_RGBA = (0, 200, 255, 255)
+_RED_TEAM_RGBA = (255, 40, 40, 255)
+_BLUE_TEAM_RGBA = (40, 90, 255, 255)
+_DEFAULT_RGBA = (235, 235, 235, 255)
+_BLACK_RGBA = (0, 0, 0, 255)
+_HUD_BG_RGBA = (92, 41, 131, 255)
+_HUD_TEXT_RGBA = (64, 186, 47, 255)
+_HIGHLIGHT_RGBA = (255, 255, 255, 255)
+
+
+def _is_ball(label):
+ return "ball" in label
+
+
+def _is_referee(label):
+ return "referee" in label or label == "ref"
+
+
+class FootballOverlay(GstBase.BaseTransform):
+ """
+ Metadata-driven broadcast overlay (football_analysis style), streaming.
+
+ Reads upstream GstAnalytics detection/tracking metadata and draws: an
+ ellipse + optional id badge per subject, a gold ellipse for referees, a
+ green triangle on the ball, fading motion trails, and a focal-player HUD
+ with a headshot, accumulated ball contacts, and distance travelled.
+ """
+
+ __gstmetadata__ = (
+ "Football Overlay",
+ "Filter/Effect/Video",
+ "Broadcast-style detection/tracking overlay (ellipses, ball triangle, "
+ "trails, headshot HUD with ball contacts + distance) from GstAnalytics",
+ "Marcus Edel ",
+ )
+
+ if backend.BACKEND == "gst":
+ src_template = Gst.PadTemplate.new(
+ "src", Gst.PadDirection.SRC, Gst.PadPresence.ALWAYS, OVERLAY_CAPS.copy()
+ )
+ sink_template = Gst.PadTemplate.new(
+ "sink", Gst.PadDirection.SINK, Gst.PadPresence.ALWAYS, OVERLAY_CAPS.copy()
+ )
+ __gsttemplates__ = (src_template, sink_template)
+
+ show_labels = GObject.Property(
+ type=bool,
+ default=True,
+ nick="Show Labels",
+ blurb="Draw the class name above each object",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ show_ids = GObject.Property(
+ type=bool,
+ default=True,
+ nick="Show Track IDs",
+ blurb="Draw the track-id badge under each tracked object",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ trails = GObject.Property(
+ type=bool,
+ default=True,
+ nick="Show Trails",
+ blurb="Draw a fading motion trail behind each tracked object",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ trail_length = GObject.Property(
+ type=int,
+ default=30,
+ minimum=2,
+ maximum=300,
+ nick="Trail Length",
+ blurb="Number of recent positions kept in each motion trail",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ show_ball = GObject.Property(
+ type=bool,
+ default=False,
+ nick="Show Ball",
+ blurb="Draw the marker on the ball (the ball is still tracked for "
+ "contact counting either way)",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ show_hud = GObject.Property(
+ type=bool,
+ default=True,
+ nick="Show HUD",
+ blurb="Draw the focal-player HUD (headshot, label, contacts, distance)",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ headshot_path = GObject.Property(
+ type=str,
+ default="data/Chinedu-Obasi_2684938.jpg",
+ nick="Headshot Path",
+ blurb="Image shown in the HUD (empty to disable)",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ headshot_size = GObject.Property(
+ type=int,
+ default=90,
+ minimum=16,
+ maximum=512,
+ nick="Headshot Size",
+ blurb="Headshot square size in pixels",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ player_label = GObject.Property(
+ type=str,
+ default="Player #8",
+ nick="Player Label",
+ blurb="Static label drawn in the HUD",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ contact_pad_ratio = GObject.Property(
+ type=float,
+ default=0.25,
+ minimum=0.0,
+ maximum=5.0,
+ nick="Contact Pad Ratio",
+ blurb="Ball counts as a contact within this fraction of the player box size",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ contact_gap_frames = GObject.Property(
+ type=int,
+ default=5,
+ minimum=0,
+ maximum=1000,
+ nick="Contact Gap Frames",
+ blurb="Min frames between counted contacts for the same player",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ player_height = GObject.Property(
+ type=float,
+ default=1.8,
+ minimum=0.1,
+ maximum=10.0,
+ nick="Player Height (m)",
+ blurb="Assumed real-world height used to convert pixels to metres",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ min_confidence = GObject.Property(
+ type=float,
+ default=0.0,
+ minimum=0.0,
+ maximum=1.0,
+ nick="Min Confidence",
+ blurb="Skip detections whose confidence is below this threshold",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ class_names = GObject.Property(
+ type=str,
+ default="",
+ nick="Class Names",
+ blurb="Comma-separated names to map numeric labels (label_N) from the "
+ "onnx/objectdetector path, e.g. 'ball,goalkeeper,player,referee'",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ team_colors = GObject.Property(
+ type=bool,
+ default=True,
+ nick="Team Colors",
+ blurb="Colour players by jersey team (red/blue, per-track majority vote); "
+ "off draws all players one colour",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ draw_from_detections = GObject.Property(
+ type=bool,
+ default=False,
+ nick="Draw From Detections",
+ blurb="Draw ellipses on the raw per-frame detection boxes instead of the "
+ "tracker's boxes -- no Kalman drift, coasted phantoms or track-split "
+ "doubles. Team colour is then classified per frame; the HUD still uses "
+ "tracker metadata if present",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ merge_iou = GObject.Property(
+ type=float,
+ default=0.5,
+ minimum=0.0,
+ maximum=1.0,
+ nick="Merge IoU",
+ blurb="Collapse overlapping boxes (across classes) into one before "
+ "drawing, so one player isn't circled twice; a box is merged when its "
+ "IoU or containment with a kept box exceeds this (0 disables)",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ position_smoothing = GObject.Property(
+ type=float,
+ default=0.5,
+ minimum=0.0,
+ maximum=0.95,
+ nick="Position Smoothing",
+ blurb="Temporal EMA on drawn box positions (0=off, higher=smoother but "
+ "more lag). Boxes are associated frame-to-frame by proximity, so this "
+ "damps detection jitter and the steps from a detection interval > 1",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ highlight_focal = GObject.Property(
+ type=bool,
+ default=True,
+ nick="Highlight Focal Player",
+ blurb="Mark the focal player (the one shown in the HUD) on the pitch "
+ "with a chevron above their head and a bolder ellipse",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ focal_track_id = GObject.Property(
+ type=int,
+ default=-1,
+ minimum=-1,
+ maximum=100000,
+ nick="Focal Track ID",
+ blurb="Pin the focal/highlighted player to this track id; -1 = auto "
+ "(the player tracked the most, with hysteresis so it stays stable)",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ def __init__(self):
+ super().__init__()
+ self.logger = LoggerFactory.get(LoggerFactory.LOGGER_TYPE_GST)
+ self.set_in_place(True)
+ self.width = 0
+ self.height = 0
+ self._order = _FORMAT_ORDER["RGBA"]
+ # per-track state, accumulated across frames
+ self._trail = {}
+ self._last_pt = {}
+ self._distance_px = {}
+ self._heights = []
+ self._widths = []
+ self._ell_w = {} # track_id -> smoothed ellipse half-width (px)
+ self._contacts = {}
+ self._last_contact_frame = {}
+ self._frames_seen = {}
+ self._track_label = {}
+ self._class_votes = {} # track_id -> {label: count}, for stable class
+ self._team_votes = {} # track_id -> {"red": n, "blue": n}, jersey team
+ self._frame = 0
+ self._focal = None # current focal track id (sticky, for hysteresis)
+ self._headshot = None
+ self._headshot_loaded = False
+ self._inv_order = [0, 1, 2, 3] # buffer-channel -> logical RGBA index
+ # Position-smoothing slots: {"box": np[x1,y1,x2,y2]} kept across frames
+ # and matched by proximity, so the drawn ellipse can be low-passed.
+ self._smooth_slots = []
+
+ def do_set_caps(self, incaps, outcaps):
+ info = GstVideo.VideoInfo.new_from_caps(incaps)
+ self.width = info.width
+ self.height = info.height
+ fmt = info.finfo.name if info.finfo else "RGBA"
+ self._order = _FORMAT_ORDER.get(fmt, _FORMAT_ORDER["RGBA"])
+ # buffer channel j holds logical[self._order[j]]; invert so we can pull
+ # logical R,G,B out of the buffer for jersey colour classification.
+ self._inv_order = [self._order.index(c) for c in range(4)]
+ self._headshot_loaded = False # re-load in the new channel order
+ self.logger.info(f"FootballOverlay caps: {fmt} {self.width}x{self.height}")
+ return True
+
+ def _map_label(self, label):
+ if self.class_names:
+ m = re.match(r"label_(\d+)$", label)
+ if m:
+ names = [s.strip() for s in self.class_names.split(",") if s.strip()]
+ i = int(m.group(1))
+ if 0 <= i < len(names):
+ return names[i]
+ return label
+
+ def _parse_label(self, full_label):
+ core = full_label
+ m = re.match(r"stream_\d+_(.*)$", full_label)
+ if m:
+ core = m.group(1)
+ m = re.match(r"(.+)_id_(\d+)$", core)
+ if m:
+ return self._map_label(m.group(1)), int(m.group(2))
+ m = re.match(r"id_(\d+)$", core)
+ if m:
+ return "object", int(m.group(1))
+ return self._map_label(core or "object"), None
+
+ def _read_metadata(self, buf):
+ entries = []
+ meta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
+ if not meta:
+ return entries
+ for index in range(GstAnalytics.relation_get_length(meta)):
+ ret, od_mtd = meta.get_od_mtd(index)
+ if not ret or od_mtd is None:
+ continue
+ full_label = GLib.quark_to_string(od_mtd.get_obj_type())
+ presence, x, y, w, h, score = od_mtd.get_location()
+ if not presence:
+ continue
+ label, track_id = self._parse_label(full_label)
+ entries.append(
+ {
+ "label": label.lower(),
+ "track_id": track_id,
+ "confidence": score,
+ "box": (x, y, x + w, y + h),
+ }
+ )
+ return entries
+
+ @staticmethod
+ def _point_to_bbox_distance(px, py, box):
+ x1, y1, x2, y2 = box
+ dx = max(x1 - px, 0.0, px - x2)
+ dy = max(y1 - py, 0.0, py - y2)
+ return (dx * dx + dy * dy) ** 0.5
+
+ def _ball_contact(self, players, ball_box):
+ """Closest player to the ball, if within contact_pad_ratio of its size."""
+ bx = (ball_box[0] + ball_box[2]) / 2.0
+ by = (ball_box[1] + ball_box[3]) / 2.0
+ best_tid, best_d, best_box = None, float("inf"), None
+ for tid, box in players.items():
+ d = self._point_to_bbox_distance(bx, by, box)
+ if d < best_d:
+ best_tid, best_d, best_box = tid, d, box
+ if best_box is None:
+ return None
+ w = best_box[2] - best_box[0]
+ h = best_box[3] - best_box[1]
+ if best_d > self.contact_pad_ratio * max(w, h):
+ return None
+ return best_tid
+
+ def _update_tracks(self, entries, det_ball_box=None):
+ self._frame += 1
+ active = set()
+ players = {}
+ ball_box = None
+ # Accumulate per-track class votes first so the stable label below
+ # already reflects this frame.
+ for e in entries:
+ tid = e["track_id"]
+ if tid is None:
+ continue
+ v = self._class_votes.setdefault(tid, {})
+ v[e["label"]] = v.get(e["label"], 0) + 1
+ for e in entries:
+ tid = e["track_id"]
+ if tid is None:
+ continue
+ label = self._stable_label(tid, e["label"])
+ if _is_ball(label):
+ ball_box = e["box"]
+ continue
+ active.add(tid)
+ players[tid] = e["box"]
+ self._track_label[tid] = label
+ self._frames_seen[tid] = self._frames_seen.get(tid, 0) + 1
+ x1, y1, x2, y2 = e["box"]
+ foot = (int((x1 + x2) / 2), int(y2))
+ if y2 - y1 > 0:
+ self._heights.append(y2 - y1)
+ if len(self._heights) > 600:
+ self._heights = self._heights[-600:]
+ self._update_ellipse_width(tid, x2 - x1)
+ prev = self._last_pt.get(tid)
+ if prev is not None:
+ self._distance_px[tid] = (
+ self._distance_px.get(tid, 0.0)
+ + ((foot[0] - prev[0]) ** 2 + (foot[1] - prev[1]) ** 2) ** 0.5
+ )
+ self._last_pt[tid] = foot
+ trail = self._trail.setdefault(tid, [])
+ trail.append(foot)
+ if len(trail) > self.trail_length:
+ del trail[: -self.trail_length]
+
+ # Fall back to the detected ball if no tracked ball this frame.
+ if ball_box is None:
+ ball_box = det_ball_box
+
+ # Ball contacts (debounced per player), like football_analyzer.
+ if ball_box is not None and players:
+ tid = self._ball_contact(players, ball_box)
+ if tid is not None:
+ last = self._last_contact_frame.get(tid)
+ if last is None or (self._frame - last) > self.contact_gap_frames:
+ self._contacts[tid] = self._contacts.get(tid, 0) + 1
+ self._last_contact_frame[tid] = self._frame
+
+ for tid in list(self._trail.keys()):
+ if tid not in active:
+ del self._trail[tid]
+ self._last_pt.pop(tid, None)
+ self._ell_w.pop(tid, None)
+ return active
+
+ def _update_ellipse_width(self, track_id, raw_w):
+ # Smooth (and outlier-reject) the per-track ellipse width so a single
+ # oversized box -- two players merged, or a drifting keep-alive
+ # prediction -- can't balloon the circle for one frame.
+ if raw_w <= 0:
+ return
+ if self._widths:
+ self._widths.append(raw_w)
+ if len(self._widths) > 600:
+ self._widths = self._widths[-600:]
+ srt = sorted(self._widths)
+ med = srt[len(srt) // 2]
+ clamped = min(max(raw_w, 0.5 * med), 1.8 * med)
+ else:
+ self._widths.append(raw_w)
+ clamped = raw_w
+ prev = self._ell_w.get(track_id)
+ # EMA: slow enough to keep the circle size steady frame-to-frame, fast
+ # enough to still follow real perspective changes as players move.
+ self._ell_w[track_id] = (
+ clamped if prev is None else 0.25 * clamped + 0.75 * prev
+ )
+
+ def _px_per_meter(self):
+ import numpy as np
+
+ if not self._heights:
+ return None
+ return float(np.median(self._heights)) / max(0.1, self.player_height)
+
+ def _focal_track(self):
+ # Pin to an explicit track id if requested.
+ if self.focal_track_id >= 0:
+ return (
+ self.focal_track_id
+ if self.focal_track_id in self._frames_seen
+ else self._focal
+ )
+ keys = set(self._frames_seen)
+ if not keys:
+ return None
+
+ # Only consider *sustained* tracks. Otherwise a track that flickered for
+ # a few frames -- common when detection/tracking churns -- can win on a
+ # single ball contact and then show ~0 distance (it was barely tracked).
+ # The floor scales with elapsed frames, with a small absolute minimum.
+ floor = max(10, int(0.2 * self._frame))
+ candidates = [t for t in keys if self._frames_seen.get(t, 0) >= floor] or list(
+ keys
+ )
+
+ # Rank by ball contacts (the player most involved with the ball), with
+ # frames-seen as a tiebreak / pre-contact fallback (before anyone has
+ # touched the ball, the most-tracked player is shown).
+ def score(t):
+ return (self._contacts.get(t, 0), self._frames_seen.get(t, 0))
+
+ best = max(candidates, key=score)
+ # Stability: keep the current focal unless a challenger has *strictly
+ # more* contacts, so the highlight/HUD don't flip on ties or noise.
+ cur = self._focal
+ if (
+ cur is not None
+ and cur in candidates
+ and self._contacts.get(best, 0) <= self._contacts.get(cur, 0)
+ ):
+ best = cur
+ self._focal = best
+ return best
+
+ def _stable_label(self, track_id, fallback=""):
+ # Majority-voted class over the track's history — smooths frame-to-frame
+ # misclassifications (e.g. a player briefly tagged 'referee'), so the
+ # gold referee marking doesn't flicker.
+ votes = self._class_votes.get(track_id)
+ if not votes:
+ return fallback
+ return max(votes, key=votes.get)
+
+ def _c(self, rgba):
+ return tuple(rgba[i] for i in self._order)
+
+ def _team_color(self, track_id):
+ # Confident kit colour from a track's accumulated jersey votes, else
+ # None. Red/blue -> team; "ref" (distinctive non-team kit) -> gold.
+ # Requires a minimum number of votes AND a clear majority, so a few
+ # noisy frames can't decide the colour.
+ if track_id is None:
+ return None
+ c = self._team_votes.get(track_id)
+ if not c:
+ return None
+ red, blue, ref = c.get("red", 0), c.get("blue", 0), c.get("ref", 0)
+ total = red + blue + ref
+ if total < 4:
+ return None
+ colors = {_RED_TEAM_RGBA: red, _BLUE_TEAM_RGBA: blue, _REFEREE_RGBA: ref}
+ color, n = max(colors.items(), key=lambda kv: kv[1])
+ return color if n >= 0.6 * total else None
+
+ def _is_referee_track(self, track_id, fallback_label):
+ # A track is a referee only if referee *clearly dominates* its class
+ # votes. Referees are rare, so a mostly-player track with a few stray
+ # 'referee' mislabels stays a player (won't get the gold circle).
+ votes = self._class_votes.get(track_id) if track_id is not None else None
+ if not votes:
+ return _is_referee(fallback_label)
+ total = sum(votes.values())
+ ref = sum(c for lbl, c in votes.items() if _is_referee(lbl))
+ return total > 0 and ref >= 3 and ref >= 0.6 * total
+
+ def _color_for(self, label, track_id):
+ # Colour by the track's *accumulated* jersey team (robust to per-frame
+ # noise). Referee/player only decides the fallback when the team is
+ # undecided: a referee keeps gold (stays visible), a player isn't drawn.
+ if _is_ball(label):
+ return _BALL_RGBA
+ if self.team_colors:
+ team = self._team_color(track_id)
+ if team is not None:
+ return team
+ return _REFEREE_RGBA if self._is_referee_track(track_id, label) else None
+ return _REFEREE_RGBA if _is_referee(label) else _PLAYER_RGBA
+
+ @staticmethod
+ def _overlap(a, b):
+ # max(IoU, intersection-over-smaller-area): catches both heavy overlap
+ # and a small duplicate box sitting inside a larger one.
+ ax1, ay1, ax2, ay2 = a
+ bx1, by1, bx2, by2 = b
+ iw = max(0.0, min(ax2, bx2) - max(ax1, bx1))
+ ih = max(0.0, min(ay2, by2) - max(ay1, by1))
+ inter = iw * ih
+ if inter <= 0.0:
+ return 0.0
+ area_a = max(0.0, ax2 - ax1) * max(0.0, ay2 - ay1)
+ area_b = max(0.0, bx2 - bx1) * max(0.0, by2 - by1)
+ union = area_a + area_b - inter
+ iou = inter / union if union > 0.0 else 0.0
+ smaller = min(area_a, area_b)
+ contain = inter / smaller if smaller > 0.0 else 0.0
+ return max(iou, contain)
+
+ @staticmethod
+ def _feet_close(a, b):
+ # True when two boxes' foot points (bottom-centre, where the ellipse is
+ # drawn) are within ~0.4 of the smaller box width. The ellipse is ~2x the
+ # box width, so near-coincident feet = one player circled twice even when
+ # the boxes' IoU is low. Genuinely adjacent players are ~a full width
+ # apart at the feet, so they're not merged.
+ fax, fay = (a[0] + a[2]) / 2.0, a[3]
+ fbx, fby = (b[0] + b[2]) / 2.0, b[3]
+ ref = max(1.0, min(a[2] - a[0], b[2] - b[0]))
+ return ((fax - fbx) ** 2 + (fay - fby) ** 2) ** 0.5 < 0.4 * ref
+
+ def _merge_overlaps(self, entries):
+ # Class-agnostic greedy suppression: keep the most confident box, drop
+ # any later box that overlaps it past merge_iou OR sits at the same feet.
+ # Collapses a player circled twice (e.g. player+goalkeeper on one person,
+ # or two offset boxes) into one. The ball is never merged against players.
+ if self.merge_iou <= 0.0 or len(entries) < 2:
+ return entries
+ ordered = sorted(entries, key=lambda e: e["confidence"], reverse=True)
+ kept = []
+ for e in ordered:
+ if _is_ball(e["label"]):
+ kept.append(e)
+ continue
+ if any(
+ not _is_ball(k["label"])
+ and (
+ self._overlap(e["box"], k["box"]) >= self.merge_iou
+ or self._feet_close(e["box"], k["box"])
+ )
+ for k in kept
+ ):
+ continue
+ kept.append(e)
+ return kept
+
+ def _assign_track_ids(self, draw_entries, track_entries):
+ # Give each drawn box a stable track id: track-mode boxes already carry
+ # one; detection-mode boxes borrow the id of the best-overlapping track
+ # (greedy, each track used once) so detection circles can use the
+ # tracker's persistent id for the badge and the accumulated colour.
+ ids = [e["track_id"] for e in draw_entries]
+ if not track_entries:
+ return ids
+ pairs = []
+ for di, e in enumerate(draw_entries):
+ if e["track_id"] is not None or _is_ball(e["label"]):
+ continue
+ for t in track_entries:
+ if _is_ball(t["label"]):
+ continue
+ ov = self._overlap(e["box"], t["box"])
+ if ov >= 0.3:
+ pairs.append((ov, di, t["track_id"]))
+ pairs.sort(key=lambda p: p[0], reverse=True)
+ used_draw, used_track = set(), set()
+ for _ov, di, tid in pairs:
+ if di in used_draw or tid in used_track:
+ continue
+ ids[di] = tid
+ used_draw.add(di)
+ used_track.add(tid)
+ return ids
+
+ def _smooth_boxes(self, np, entries):
+ # Temporal EMA on the boxes we're about to draw. Each box is matched to
+ # the nearest slot from last frame (by centre, within a size-relative
+ # gate) and pulled toward the new detection; slots not matched this
+ # frame are dropped (no phantoms). Damps jitter and interval steps. The
+ # ball is passed through unsmoothed so it never lags.
+ import numpy as np
+
+ a = float(self.position_smoothing)
+ if a <= 0.0 or not entries:
+ return entries
+ used = set()
+ out = []
+ for e in entries:
+ if _is_ball(e["label"]):
+ out.append(e)
+ continue
+ box = np.array(e["box"], dtype=np.float64)
+ cx, cy = (box[0] + box[2]) / 2.0, (box[1] + box[3]) / 2.0
+ # Generous gate so a coherent interval-step jump still associates
+ # (and glides) without grabbing a different nearby player.
+ gate = 1.5 * max(box[2] - box[0], box[3] - box[1], 1.0)
+ best, best_d = None, gate
+ for idx, slot in enumerate(self._smooth_slots):
+ if idx in used:
+ continue
+ sb = slot["box"]
+ d = (
+ ((sb[0] + sb[2]) / 2.0 - cx) ** 2
+ + ((sb[1] + sb[3]) / 2.0 - cy) ** 2
+ ) ** 0.5
+ if d < best_d:
+ best, best_d = idx, d
+ if best is None:
+ self._smooth_slots.append({"box": box.copy()})
+ used.add(len(self._smooth_slots) - 1)
+ smoothed = box
+ else:
+ used.add(best)
+ slot = self._smooth_slots[best]
+ slot["box"] = a * slot["box"] + (1.0 - a) * box
+ smoothed = slot["box"]
+ ne = dict(e)
+ ne["box"] = (
+ float(smoothed[0]),
+ float(smoothed[1]),
+ float(smoothed[2]),
+ float(smoothed[3]),
+ )
+ out.append(ne)
+ self._smooth_slots = [s for i, s in enumerate(self._smooth_slots) if i in used]
+ return out
+
+ def _detection_color(self, cv2, np, frame, label, box):
+ # Colour a raw detection box (no track id) by its jersey team, classified
+ # from this frame -- referees included. When the jersey isn't clearly a
+ # team colour, a referee falls back to gold (so real refs stay visible)
+ # and a player isn't drawn (matching the track-mode behaviour).
+ ref = _is_referee(label)
+ if not self.team_colors:
+ return _REFEREE_RGBA if ref else _PLAYER_RGBA
+ vote = self._classify_jersey(cv2, np, frame, box)
+ if vote == "red":
+ return _RED_TEAM_RGBA
+ if vote == "blue":
+ return _BLUE_TEAM_RGBA
+ if vote == "ref":
+ return _REFEREE_RGBA
+ return _REFEREE_RGBA if ref else None
+
+ def _classify_jersey(self, cv2, np, frame, box):
+ # Dominant jersey colour in the torso patch -> "red"/"blue"/"ref"/None
+ # (HSV). "ref" is a distinctive non-team kit colour (yellow/orange or
+ # pink/magenta) -- chosen to avoid grass-green and the red/blue teams --
+ # so the referee is identified by its kit colour, not the class label.
+ import cv2
+ import numpy as np
+
+ x1, y1, x2, y2 = (int(v) for v in box)
+ h_box, w_box = y2 - y1, x2 - x1
+ if h_box <= 0 or w_box <= 0:
+ return None
+ jy1, jy2 = y1 + int(0.15 * h_box), y1 + int(0.55 * h_box)
+ jx1, jx2 = x1 + int(0.25 * w_box), x1 + int(0.75 * w_box)
+ H, W = frame.shape[:2]
+ jy1, jy2 = max(0, jy1), min(H, jy2)
+ jx1, jx2 = max(0, jx1), min(W, jx2)
+ if jy2 - jy1 < 3 or jx2 - jx1 < 3:
+ return None
+ # logical RGB from the buffer's channel order, then HSV
+ rgb = np.ascontiguousarray(frame[jy1:jy2, jx1:jx2][:, :, self._inv_order[:3]])
+ hsv = cv2.cvtColor(rgb, cv2.COLOR_RGB2HSV)
+ s_v = (hsv[..., 1] > 80) & (hsv[..., 2] > 50)
+ h = hsv[..., 0]
+ red = int((((h <= 10) | (h >= 170)) & s_v).sum())
+ blue = int(((h >= 100) & (h <= 130) & s_v).sum())
+ # Referee kit: yellow/orange (~18-34) or pink/magenta (~145-165). These
+ # bands skip grass-green (~40-90) and the red/blue team bands.
+ ref = int(((((h >= 18) & (h <= 34)) | ((h >= 145) & (h <= 165))) & s_v).sum())
+ min_pixels = max(20, int(0.02 * rgb.shape[0] * rgb.shape[1]))
+ counts = {"red": red, "blue": blue, "ref": ref}
+ best = max(counts, key=counts.get)
+ if counts[best] < min_pixels:
+ return None
+ return best
+
+ def _load_headshot(self, cv2, np):
+ import cv2
+ import numpy as np
+
+ if self._headshot_loaded:
+ return self._headshot
+ self._headshot_loaded = True
+ self._headshot = None
+ path = self.headshot_path
+ if not path or not os.path.exists(path):
+ if path:
+ self.logger.warning(f"headshot not found: {path}")
+ return None
+ img = cv2.imread(path) # BGR
+ if img is None:
+ return None
+ sz = int(self.headshot_size)
+ img = cv2.resize(img, (sz, sz), interpolation=cv2.INTER_AREA)
+ rgb = img[:, :, ::-1] # BGR -> RGB
+ alpha = np.full((sz, sz, 1), 255, dtype=np.uint8)
+ rgba = np.concatenate([rgb, alpha], axis=2).astype(np.uint8) # logical RGBA
+
+ self._headshot = np.ascontiguousarray(rgba[:, :, list(self._order)])
+ return self._headshot
+
+ def _draw_trail(self, cv2, np, frame, points, rgba):
+ import cv2
+ import numpy as np
+
+ if len(points) < 2:
+ return
+ pts = np.array(points, dtype=np.int32).reshape(-1, 1, 2)
+ cv2.polylines(frame, [pts], False, self._c(rgba), 2, cv2.LINE_AA)
+
+ def _draw_ellipse(self, cv2, frame, box, rgba, track_id):
+ import cv2
+
+ x1, y1, x2, y2 = box
+ y_bottom = int(y2)
+ x_center = int((x1 + x2) / 2)
+ # Prefer the per-track smoothed width so the ellipse stays stable even
+ # when a single detection box is momentarily oversized.
+ smoothed = self._ell_w.get(track_id)
+ width = max(1, int(smoothed if smoothed is not None else x2 - x1))
+ color = self._c(rgba)
+ cv2.ellipse(
+ frame,
+ (x_center, y_bottom),
+ (width, max(1, int(0.35 * width))),
+ 0.0,
+ -45,
+ 235,
+ color,
+ 2,
+ cv2.LINE_AA,
+ )
+ if self.show_ids and track_id is not None:
+ rect_w, rect_h = 40, 18
+ x1r = x_center - rect_w // 2
+ x2r = x_center + rect_w // 2
+ y1r = y_bottom - rect_h // 2 + 15
+ y2r = y_bottom + rect_h // 2 + 15
+ cv2.rectangle(frame, (x1r, y1r), (x2r, y2r), color, cv2.FILLED)
+ tx = x1r + 12 - (10 if track_id > 99 else 0)
+ cv2.putText(
+ frame,
+ str(track_id),
+ (tx, y1r + 14),
+ cv2.FONT_HERSHEY_SIMPLEX,
+ 0.5,
+ self._c(_BLACK_RGBA),
+ 2,
+ cv2.LINE_AA,
+ )
+
+ def _draw_triangle(self, cv2, np, frame, box, rgba):
+ import cv2
+ import numpy as np
+
+ x1, y1, x2, y2 = box
+ x = int((x1 + x2) / 2)
+ y = int(y1)
+ pts = np.array([[x, y], [x - 10, y - 20], [x + 10, y - 20]], dtype=np.int32)
+ cv2.drawContours(frame, [pts], 0, self._c(rgba), cv2.FILLED)
+ cv2.drawContours(frame, [pts], 0, self._c(_BLACK_RGBA), 2)
+
+ def _draw_focal_marker(self, cv2, np, frame, box):
+ # Broadcast-style "selected player" chevron floating above the head,
+ # plus a bolder ellipse, to flag the focal (HUD) player on the pitch.
+ import cv2
+ import numpy as np
+
+ x1, y1, x2, y2 = box
+ cx = int((x1 + x2) / 2)
+ tip_y = int(y1) - 10
+ s = 16
+ pts = np.array(
+ [
+ [cx, tip_y],
+ [cx - s, tip_y - int(s * 1.5)],
+ [cx + s, tip_y - int(s * 1.5)],
+ ],
+ dtype=np.int32,
+ )
+ cv2.drawContours(frame, [pts], 0, self._c(_HIGHLIGHT_RGBA), cv2.FILLED)
+ cv2.drawContours(frame, [pts], 0, self._c(_BLACK_RGBA), 2)
+ # Bolder ring at the feet to reinforce the selection.
+ x_center = int((x1 + x2) / 2)
+ width = max(1, int(x2 - x1))
+ cv2.ellipse(
+ frame,
+ (x_center, int(y2)),
+ (width, max(1, int(0.35 * width))),
+ 0.0,
+ -45,
+ 235,
+ self._c(_HIGHLIGHT_RGBA),
+ 4,
+ cv2.LINE_AA,
+ )
+
+ def _draw_label(self, cv2, frame, box, label, rgba):
+ import cv2
+
+ x1, y1, _, _ = box
+ cv2.putText(
+ frame,
+ label,
+ (int(x1), max(12, int(y1) - 6)),
+ cv2.FONT_HERSHEY_SIMPLEX,
+ 0.5,
+ self._c(rgba),
+ 1,
+ cv2.LINE_AA,
+ )
+
+ def _draw_hud(self, cv2, frame, contacts, distance_m, rgba, headshot):
+ import cv2
+
+ font = cv2.FONT_HERSHEY_SIMPLEX
+ x, y = 10, 10
+ if headshot is not None:
+ hh, hw = headshot.shape[:2]
+ w, h = hw + 280, max(110, hh + 20)
+ text_x = x + hw + 20
+ else:
+ w, h = 320, 100
+ text_x = x + 12
+ cv2.rectangle(frame, (x, y), (x + w, y + h), self._c(_HUD_BG_RGBA), cv2.FILLED)
+ cv2.rectangle(frame, (x, y), (x + w, y + h), self._c(rgba), 2)
+ if headshot is not None:
+ hy, hx = y + 10, x + 10
+ fh, fw = frame.shape[:2]
+ hh = min(hh, fh - hy)
+ hw = min(hw, fw - hx)
+ if hh > 0 and hw > 0:
+ frame[hy : hy + hh, hx : hx + hw] = headshot[:hh, :hw]
+ cv2.rectangle(frame, (hx, hy), (hx + hw, hy + hh), self._c(rgba), 2)
+ tc = self._c(_HUD_TEXT_RGBA)
+ cv2.putText(
+ frame, self.player_label, (text_x, y + 28), font, 0.7, tc, 2, cv2.LINE_AA
+ )
+ cv2.putText(
+ frame,
+ f"Ball contacts: {contacts}",
+ (text_x, y + 58),
+ font,
+ 0.6,
+ tc,
+ 1,
+ cv2.LINE_AA,
+ )
+ cv2.putText(
+ frame,
+ f"Distance: {distance_m:.1f} m",
+ (text_x, y + 85),
+ font,
+ 0.6,
+ tc,
+ 1,
+ cv2.LINE_AA,
+ )
+
+ def do_transform_ip(self, buf):
+ import numpy as np
+
+ try:
+ import numpy as np
+
+ all_entries = self._read_metadata(buf)
+ # The buffer carries both the detector's boxes (track_id None) and
+ # the tracker's boxes (track_id set). Tracking state/HUD always use
+ # the tracked entries; what we *draw* depends on draw_from_detections.
+ track_entries = [e for e in all_entries if e["track_id"] is not None]
+ det_entries = [e for e in all_entries if e["track_id"] is None]
+
+ # Ball position for contact counting: prefer a tracked ball, else
+ # fall back to the strongest ball *detection* (the ball is small and
+ # fast, so it often isn't tracked) -- so contacts still get counted.
+ det_ball_box = None
+ best_ball = -1.0
+ for e in det_entries:
+ if _is_ball(e["label"]) and e["confidence"] > best_ball:
+ best_ball, det_ball_box = e["confidence"], e["box"]
+
+ # Per-track state (votes, contacts, distance, focal) from the tracker.
+ active = self._update_tracks(
+ track_entries if track_entries else all_entries, det_ball_box
+ )
+
+ if self.draw_from_detections:
+ draw_entries = list(det_entries)
+ # Bridge missed detections: the detector occasionally drops a
+ # player for a frame, which would flicker the circle. The tracker
+ # is still coasting that player (Kalman keep-alive), so draw any
+ # confirmed track that has no detection this frame -- detections
+ # still drive everything they cover; tracks only fill the gaps.
+ if track_entries:
+ covered = set()
+ for d in det_entries:
+ if _is_ball(d["label"]):
+ continue
+ for t in track_entries:
+ if t["track_id"] in covered or _is_ball(t["label"]):
+ continue
+ if self._overlap(d["box"], t["box"]) >= 0.3:
+ covered.add(t["track_id"])
+ draw_entries += [
+ t
+ for t in track_entries
+ if not _is_ball(t["label"]) and t["track_id"] not in covered
+ ]
+ else:
+ draw_entries = track_entries if track_entries else det_entries
+ # min-confidence gates only what we *draw* (tracks carry conf 1.0, so
+ # they're unaffected); the contact math above used the raw detections.
+ if self.min_confidence > 0.0:
+ draw_entries = [
+ e for e in draw_entries if e["confidence"] >= self.min_confidence
+ ]
+ # Collapse overlapping boxes so one player isn't circled twice,
+ # then low-pass the positions so the circle glides.
+ draw_entries = self._merge_overlaps(draw_entries)
+ draw_entries = self._smooth_boxes(np, draw_entries)
+ if not all_entries:
+ return Gst.FlowReturn.OK
+
+ import cv2
+
+ ok, mapinfo = buf.map(Gst.MapFlags.WRITE)
+ if not ok:
+ self.logger.error("Failed to map buffer for writing")
+ return Gst.FlowReturn.ERROR
+ try:
+ frame = np.frombuffer(
+ mapinfo.data, dtype=np.uint8, count=self.height * self.width * 4
+ ).reshape(self.height, self.width, 4)
+
+ # Jersey team voting first, so trails/ellipses use this frame's
+ # vote (track mode; detection mode classifies per box at draw).
+ # Referees are voted on too -- their colour comes from the jersey
+ # (gold only as the fallback), not the class label.
+ if self.team_colors:
+ for e in track_entries:
+ tid = e["track_id"]
+ lab = self._stable_label(tid, e["label"])
+ if _is_ball(lab):
+ continue
+ vote = self._classify_jersey(cv2, np, frame, e["box"])
+ if vote:
+ tv = self._team_votes.setdefault(
+ tid, {"red": 0, "blue": 0, "ref": 0}
+ )
+ tv[vote] = tv.get(vote, 0) + 1
+
+ if self.trails:
+ for tid in active:
+ rgba = self._color_for(self._track_label.get(tid, ""), tid)
+ if rgba is None:
+ continue
+ self._draw_trail(cv2, np, frame, self._trail.get(tid, []), rgba)
+
+ # Which drawn box is the focal (HUD) player? Match the focal
+ # track's box to the nearest drawn box so we can highlight it
+ # even when drawing from detections (no track id on the box).
+ focal_idx = None
+ if self.highlight_focal:
+ focal_tid = self._focal_track()
+ focal_box = None
+ if focal_tid is not None:
+ for t in track_entries:
+ if t["track_id"] == focal_tid:
+ focal_box = t["box"]
+ break
+ if focal_box is not None:
+ best = 0.0
+ for i, e in enumerate(draw_entries):
+ if _is_ball(e["label"]):
+ continue
+ ov = self._overlap(e["box"], focal_box)
+ if ov > best:
+ best, focal_idx = ov, i
+
+ # Stable track id per drawn box (detection boxes borrow the id of
+ # the track they overlap) -- used for the id badge and to look up
+ # the track's accumulated colour.
+ draw_ids = self._assign_track_ids(draw_entries, track_entries)
+
+ for i, e in enumerate(draw_entries):
+ box = e["box"]
+ badge_id = draw_ids[i]
+ # Use the track's stable identity (class + accumulated team
+ # votes) for colour whenever the box maps to a track -- in
+ # detection mode that's the box's matched track id. This
+ # makes colour robust to per-frame label/jersey noise. Only
+ # an unmatched detection falls back to this frame's guess.
+ color_tid = e["track_id"] if e["track_id"] is not None else badge_id
+ if color_tid is not None:
+ label = self._stable_label(color_tid, e["label"])
+ else:
+ label = e["label"]
+ if _is_ball(label):
+ if self.show_ball:
+ self._draw_triangle(cv2, np, frame, box, _BALL_RGBA)
+ continue
+ if color_tid is not None:
+ rgba = self._color_for(label, color_tid)
+ else:
+ rgba = self._detection_color(cv2, np, frame, label, box)
+ if rgba is None:
+ continue
+ self._draw_ellipse(cv2, frame, box, rgba, badge_id)
+ if i == focal_idx:
+ self._draw_focal_marker(cv2, np, frame, box)
+ if self.show_labels:
+ self._draw_label(cv2, frame, box, label, rgba)
+
+ if self.show_hud:
+ focal = self._focal_track()
+ if focal is not None:
+ ppm = self._px_per_meter()
+ dist_m = (
+ (self._distance_px.get(focal, 0.0) / ppm) if ppm else 0.0
+ )
+ hud_rgba = (
+ self._color_for(self._track_label.get(focal, ""), focal)
+ or _DEFAULT_RGBA
+ )
+ self._draw_hud(
+ cv2,
+ frame,
+ self._contacts.get(focal, 0),
+ dist_m,
+ hud_rgba,
+ self._load_headshot(cv2, np),
+ )
+ finally:
+ buf.unmap(mapinfo)
+
+ return Gst.FlowReturn.OK
+
+ except Exception as e:
+ self.logger.error(f"FootballOverlay transform error: {e}")
+ return Gst.FlowReturn.ERROR
+
+
+# GStreamer factory registration runs only under the gst backend.
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_football_overlay", FootballOverlay
+ )
+elif not CAN_REGISTER_ELEMENT:
+ GlobalLogger().warning(
+ "The 'pyml_football_overlay' element will not be registered because "
+ "required modules are missing."
+ )
diff --git a/plugins/python/inference.py b/plugins/python/inference.py
index e112092..e7fe8d5 100644
--- a/plugins/python/inference.py
+++ b/plugins/python/inference.py
@@ -17,18 +17,11 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- from gi.repository import Gst, GObject
-
from video_transform import VideoTransform
- from utils.muxed_buffer_processor import MuxedBufferProcessor
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -45,7 +38,7 @@ class GenericInferenceTransform(VideoTransform):
engine-name: pytorch (default), onnx, tensorflow, tflite, openvino
Example:
- gst-launch-1.0 filesrc location=data/people.mp4 ! decodebin name=d \
+ python pyml-launch.py filesrc location=data/people.mp4 ! decodebin name=d \
d. ! queue ! videoconvert ! videoscale \
! "video/x-raw,format=RGB,width=640,height=480" \
! pyml_inference engine-name=onnx model-name=yolo11m.onnx device=cpu \
@@ -67,35 +60,23 @@ def do_start(self):
)
return result
- def do_transform_ip(self, buf):
- try:
- processor = MuxedBufferProcessor(
- self.logger, self.width, self.height, 30, 1
- )
- frames, _, num_sources, _ = processor.extract_frames(buf, self.sinkpad)
- if frames is None:
- return Gst.FlowReturn.ERROR
-
- frame = frames[0] if num_sources > 1 else frames
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Run the engine on the frame and log the result. The frame is unchanged."""
+ frame = frames[0] if num_sources > 1 else frames
- if not self.engine:
- return Gst.FlowReturn.OK
+ if not self.engine:
+ return
- result = self.engine.do_forward(frame)
- if result is not None:
- self.logger.info(f"inference result: {result}")
+ result = self.engine.do_forward(frame)
+ if result is not None:
+ self.logger.info(f"inference result: {result}")
- return Gst.FlowReturn.OK
- except Exception as e:
- self.logger.error(f"inference error: {e}")
- return Gst.FlowReturn.ERROR
-
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(GenericInferenceTransform)
- __gstelementfactory__ = ("pyml_inference", Gst.Rank.NONE, GenericInferenceTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_inference", GenericInferenceTransform
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_inference' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/kafkasink.py b/plugins/python/kafkasink.py
index 349589f..121d74d 100644
--- a/plugins/python/kafkasink.py
+++ b/plugins/python/kafkasink.py
@@ -17,22 +17,21 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
import json
import os
- from confluent_kafka import Producer
import gi
gi.require_version("Gst", "1.0")
- gi.require_version("GstAnalytics", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject, GLib, GstAnalytics # noqa: E402
+ from gi.repository import Gst # noqa: E402
from log.logger_factory import LoggerFactory # noqa: E402
from utils.runtime_utils import runtime_check_gstreamer_version # noqa: E402
+ from backend import analytics, GObject # noqa: E402
except ImportError as e:
CAN_REGISTER_ELEMENT = False
GlobalLogger().warning(
@@ -191,6 +190,10 @@ def do_get_property(self, prop: GObject.GParamSpec):
def initialize_producer(self):
if self.broker and self.topic:
try:
+ # Imported lazily so a missing confluent_kafka does not break the
+ # element scan; it is only needed once a producer is created.
+ from confluent_kafka import Producer
+
config = {
"bootstrap.servers": self.broker,
"linger.ms": int(self.linger_ms),
@@ -237,40 +240,21 @@ def extract_metadata(self, buffer):
"""Extract object detection metadata from GstBuffer using GstAnalyticsRelationMeta."""
metadata = []
- meta = GstAnalytics.buffer_get_analytics_relation_meta(buffer)
+ meta = analytics.get_relation_meta(buffer)
if not meta:
- self.logger.warning("No GstAnalytics metadata found on buffer.")
+ self.logger.warning("No analytics metadata found on buffer.")
return metadata
try:
- count = GstAnalytics.relation_get_length(meta)
- for index in range(count):
- ret, od_mtd = meta.get_od_mtd(index)
- if not ret or od_mtd is None:
- break
-
- label_quark = od_mtd.get_obj_type()
- label = GLib.quark_to_string(label_quark)
- location = od_mtd.get_location()
-
- presence, x, y, w, h, loc_conf_lvl = location
- if presence:
- x1 = x
- y1 = y
- x2 = x + w
- y2 = y + h
-
- metadata.append(
- {
- "label": label,
- "confidence": loc_conf_lvl,
- "box": {"x1": x1, "y1": y1, "x2": x2, "y2": y2},
- }
- )
- else:
- self.logger.warning(
- "Presence flag in location is False. Skipping this entry."
- )
+ for obj in analytics.read_objects(meta):
+ x, y, w, h = obj["x"], obj["y"], obj["w"], obj["h"]
+ metadata.append(
+ {
+ "label": obj["label"],
+ "confidence": obj["score"],
+ "box": {"x1": x, "y1": y, "x2": x + w, "y2": y + h},
+ }
+ )
except Exception as e:
self.logger.error(f"Error while extracting metadata: {e}")
@@ -332,10 +316,11 @@ def do_finalize(self):
self.producer.flush()
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(KafkaSink)
- __gstelementfactory__ = (KafkaSink.GST_PLUGIN_NAME, Gst.Rank.NONE, KafkaSink)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ KafkaSink.GST_PLUGIN_NAME, KafkaSink
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
- "The 'pyml_kafkasink' element will not be registered because confluent_kafka module is missing."
+ "The 'pyml_kafkasink' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/llm.py b/plugins/python/llm.py
index ced1265..ccce00d 100644
--- a/plugins/python/llm.py
+++ b/plugins/python/llm.py
@@ -17,13 +17,10 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- from gi.repository import Gst, GObject # noqa: E402
from base_llm import BaseLlm
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -39,10 +36,9 @@ class LLM(BaseLlm):
)
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(LLM)
- __gstelementfactory__ = ("pyml_llm", Gst.Rank.NONE, LLM)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_llm", LLM)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_llm' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/llm_remote.py b/plugins/python/llm_remote.py
index 62a65ff..4187adc 100644
--- a/plugins/python/llm_remote.py
+++ b/plugins/python/llm_remote.py
@@ -17,6 +17,7 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
@@ -25,7 +26,8 @@
gi.require_version("Gst", "1.0")
gi.require_version("GstBase", "1.0")
gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject, GstBase
+ from gi.repository import Gst, GstBase
+ from backend import GObject
from log.logger_factory import LoggerFactory
except ImportError as e:
@@ -48,7 +50,7 @@ class LlmRemote(GstBase.Aggregator):
timeout: HTTP request timeout in seconds (default: 120)
Example (Ollama):
- gst-launch-1.0 filesrc location=prompt.txt ! "text/x-raw,format=utf8" \
+ python pyml-launch.py filesrc location=prompt.txt ! "text/x-raw,format=utf8" \
! pyml_llm_remote url=http://localhost:11434/api/generate model-name=llama3 \
! fakesink
"""
@@ -60,20 +62,22 @@ class LlmRemote(GstBase.Aggregator):
"Aaron Boxer ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- Gst.Caps.from_string("text/x-raw,format=utf8"),
- ),
- Gst.PadTemplate.new(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.REQUEST,
- Gst.Caps.from_string("text/x-raw,format=utf8"),
- ),
- )
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string("text/x-raw,format=utf8"),
+ ),
+ Gst.PadTemplate.new(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.REQUEST,
+ Gst.Caps.from_string("text/x-raw,format=utf8"),
+ ),
+ )
def __init__(self):
super().__init__()
@@ -267,10 +271,9 @@ def push_generated_text(self, inbuf, generated_text):
return Gst.FlowReturn.ERROR
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(LlmRemote)
- __gstelementfactory__ = ("pyml_llm_remote", Gst.Rank.NONE, LlmRemote)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_llm_remote", LlmRemote)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_llm_remote' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/llm_stream_filter.py b/plugins/python/llm_stream_filter.py
index 4a2b062..aad478a 100644
--- a/plugins/python/llm_stream_filter.py
+++ b/plugins/python/llm_stream_filter.py
@@ -18,6 +18,7 @@
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
@@ -26,11 +27,9 @@
gi.require_version("Gst", "1.0")
gi.require_version("GstBase", "1.0")
gi.require_version("GstVideo", "1.0")
- gi.require_version("GLib", "2.0")
- gi.require_version("GstAnalytics", "1.0")
- from gi.repository import Gst, GObject, GstAnalytics, GLib
+ from gi.repository import Gst
- from utils.muxed_buffer_processor import MuxedBufferProcessor
+ from backend import analytics, frameio, GObject
from video_transform import VideoTransform
from engine.engine_manager import EngineManager
from utils.caption_utils import load_captions
@@ -55,20 +54,22 @@ class LLMStreamFilter(VideoTransform):
"Aaron Boxer ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new(
- "video_src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- Gst.Caps.from_string("video/x-raw"),
- ),
- Gst.PadTemplate.new(
- "text_src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.REQUEST,
- Gst.Caps.from_string("text/x-raw, format=utf8"),
- ),
- )
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new(
+ "video_src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string("video/x-raw"),
+ ),
+ Gst.PadTemplate.new(
+ "text_src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.REQUEST,
+ Gst.Caps.from_string("text/x-raw, format=utf8"),
+ ),
+ )
num_streams = GObject.Property(
type=int,
@@ -277,15 +278,8 @@ def do_transform_ip(self, buf):
import torch
try:
- muxed_processor = MuxedBufferProcessor(
- self.logger,
- self.width,
- self.height,
- framerate_num=30,
- framerate_denom=1,
- )
- frames, id_str, num_sources, format = muxed_processor.extract_frames(
- buf, self.sinkpad
+ frames, num_sources, _ = frameio.read_frames(
+ buf, self.sinkpad, self.width, self.height
)
if frames is None:
self.logger.error("Failed to extract frames")
@@ -337,11 +331,12 @@ def do_transform_ip(self, buf):
# Add metadata and push text buffers for selected streams
for idx, caption in enumerate(captions):
if idx in self.selected_streams:
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
+ meta = analytics.add_relation_meta(buf)
if meta:
- qk = GLib.quark_from_string(f"stream_{idx}_{caption}")
- ret, mtd = meta.add_one_cls_mtd(idx, qk)
- if ret:
+ mtd = analytics.add_classification(
+ meta, idx, f"stream_{idx}_{caption}"
+ )
+ if mtd is not None:
self.logger.info(f"Stream {idx}: Added caption {caption}")
else:
self.logger.error(f"Stream {idx}: Failed to add metadata")
@@ -368,10 +363,11 @@ def do_transform_ip(self, buf):
return Gst.FlowReturn.ERROR
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(LLMStreamFilter)
- __gstelementfactory__ = ("pyml_llmstreamfilter", Gst.Rank.NONE, LLMStreamFilter)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_llmstreamfilter", LLMStreamFilter
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_llmstreamfilter' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/mariantranslate.py b/plugins/python/mariantranslate.py
index cfd4b6f..9af52d1 100644
--- a/plugins/python/mariantranslate.py
+++ b/plugins/python/mariantranslate.py
@@ -17,13 +17,10 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- from gi.repository import Gst, GObject # noqa: E402
from base_translate import BaseTranslate
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -74,10 +71,11 @@ def do_translate_text(self, text):
return ""
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(MarianTranslate)
- __gstelementfactory__ = ("pyml_mariantranslate", Gst.Rank.NONE, MarianTranslate)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_mariantranslate", MarianTranslate
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_mariantranslate' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/maskrcnn.py b/plugins/python/maskrcnn.py
index d04ef10..cd60df1 100644
--- a/plugins/python/maskrcnn.py
+++ b/plugins/python/maskrcnn.py
@@ -17,19 +17,12 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- gi.require_version("GstAnalytics", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject, GstAnalytics, GLib # noqa: E402
-
from base_objectdetector import BaseObjectDetector
+ from tasks.maskrcnn import MaskRCNNTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
GlobalLogger().warning(
@@ -37,7 +30,7 @@
)
-class MaskRCNN(BaseObjectDetector):
+class MaskRCNN(BaseObjectDetector, MaskRCNNTask):
"""
GStreamer element for Mask R-CNN model inference on video frames.
"""
@@ -49,65 +42,10 @@ class MaskRCNN(BaseObjectDetector):
"Aaron Boxer ",
)
- def do_decode(self, buf, output, stream_idx=0):
- """
- Processes the Mask R-CNN model's output detections and adds metadata to the GStreamer buffer,
- tagged with the stream index.
- """
- boxes = output["boxes"]
- labels = output["labels"]
- scores = output["scores"]
- masks = output["masks"] # Additional mask outputs for Mask R-CNN
-
- self.logger.info(
- f"Processing buffer at address: {hex(id(buf))} for stream {stream_idx}"
- )
- self.logger.info(f"Stream {stream_idx} - Processing {len(boxes)} detections")
-
- # Add analytics metadata to the buffer
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
- if not meta:
- self.logger.error(f"Stream {stream_idx} - Failed to add analytics metadata")
- return
-
- for i, (box, label, score, mask) in enumerate(
- zip(boxes, labels, scores, masks)
- ):
- x1, y1, x2, y2 = box
- self.logger.info(
- f"Stream {stream_idx} - Detection {i}: Box coordinates (x1={x1}, y1={y1}, x2={x2}, y2={y2}), "
- f"Label={label}, Score={score:.2f}"
- )
-
- # Use stream_idx in the quark string to differentiate streams
- qk_string = f"stream_{stream_idx}_label_{label}"
- qk = GLib.quark_from_string(qk_string)
- ret, mtd = meta.add_od_mtd(qk, x1, y1, x2 - x1, y2 - y1, score)
- if ret:
- self.logger.info(
- f"Stream {stream_idx} - Successfully added object detection metadata with quark {qk_string} and mtd {mtd}"
- )
- else:
- self.logger.error(
- f"Stream {stream_idx} - Failed to add object detection metadata"
- )
-
- attached_meta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
- if attached_meta:
- count = GstAnalytics.relation_get_length(attached_meta)
- self.logger.info(
- f"Stream {stream_idx} - Metadata successfully attached to buffer at address: {hex(id(buf))} with {count} relations"
- )
- else:
- self.logger.warning(
- f"Stream {stream_idx} - Failed to retrieve attached metadata immediately after addition for buffer: {hex(id(buf))}"
- )
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(MaskRCNN)
- __gstelementfactory__ = ("pyml_maskrcnn", Gst.Rank.NONE, MaskRCNN)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_maskrcnn", MaskRCNN)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_maskrcnn' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/objectdetector.py b/plugins/python/objectdetector.py
index a429eb5..5f6e750 100644
--- a/plugins/python/objectdetector.py
+++ b/plugins/python/objectdetector.py
@@ -16,22 +16,9 @@
# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
# Boston, MA 02110-1301, USA.
-from log.global_logger import GlobalLogger
-
-CAN_REGISTER_ELEMENT = True
-try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject # noqa: E402
- from base_objectdetector import BaseObjectDetector
-except ImportError as e:
- CAN_REGISTER_ELEMENT = False
- GlobalLogger().warning(
- f"The 'objectdetector_pylm' element will not be available. Error: {e}"
- )
+from base_objectdetector import BaseObjectDetector
+from backend import GObject
+import backend
class ObjectDetector(BaseObjectDetector):
@@ -46,17 +33,45 @@ class ObjectDetector(BaseObjectDetector):
"Aaron Boxer ",
)
+ confidence = GObject.Property(
+ type=float,
+ default=0.25,
+ minimum=0.0,
+ maximum=1.0,
+ nick="Confidence Threshold",
+ blurb="Minimum detection confidence for the decoder post-process "
+ "(anchor_free); lower = more (and weaker) detections",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ nms_iou = GObject.Property(
+ type=float,
+ default=0.45,
+ minimum=0.0,
+ maximum=1.0,
+ nick="NMS IoU",
+ blurb="NMS IoU threshold for the decoder post-process; higher keeps "
+ "more overlapping boxes",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
def __init__(self):
super().__init__()
self.logger.info(
"ObjectDetector created without a model. Please set the 'model-name' property."
)
+ def do_forward(self, frames):
+ # Push decoder thresholds to the engine before it post-processes.
+ if self.engine:
+ self.engine.conf = self.confidence
+ self.engine.iou = self.nms_iou
+ return super().do_forward(frames)
+
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(ObjectDetector)
- __gstelementfactory__ = ("pyml_objectdetector", Gst.Rank.NONE, ObjectDetector)
-else:
- GlobalLogger().warning(
- "The 'pyml_objectdetector' element will not be registered because base_objectdetector module is missing."
+# The class is backend-agnostic: under g2g the host imports this module and
+# instantiates ObjectDetector directly, so no GObject registration applies.
+# GStreamer factory registration runs only under the gst backend.
+if backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_objectdetector", ObjectDetector
)
diff --git a/plugins/python/ocr.py b/plugins/python/ocr.py
index 16e2fc4..c0ba4e8 100644
--- a/plugins/python/ocr.py
+++ b/plugins/python/ocr.py
@@ -17,24 +17,16 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import ctypes
- import json
-
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- from gi.repository import Gst, GObject
-
from video_transform import VideoTransform
from utils.format_converter import FormatConverter
- from utils.muxed_buffer_processor import MuxedBufferProcessor
- from engine.pytorch_engine import PyTorchEngine
+ from engine.ocr_engine import OcrEngine
from engine.engine_factory import EngineFactory
+ from backend import GObject
+ from tasks.ocr import OcrTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -44,81 +36,7 @@
OCR_META_HEADER = b"GST-OCR:"
-class OcrEngine(PyTorchEngine):
- """
- PyTorch engine for TrOCR text recognition.
-
- Supports HuggingFace model IDs:
- microsoft/trocr-base-printed
- microsoft/trocr-large-printed
- microsoft/trocr-base-handwritten
- """
-
- def do_load_model(self, model_name, **kwargs):
- try:
- from transformers import TrOCRProcessor, VisionEncoderDecoderModel
-
- self.processor = TrOCRProcessor.from_pretrained(model_name)
- self.model = VisionEncoderDecoderModel.from_pretrained(model_name)
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.model.eval()
- self.logger.info(f"TrOCR model '{model_name}' loaded on {self.device}")
- except Exception as e:
- raise ValueError(f"Failed to load TrOCR model '{model_name}': {e}")
-
- def do_forward(self, frames):
- import numpy as np
- import torch
- from PIL import Image
-
- is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
- if not is_batch:
- frames = frames[np.newaxis]
-
- results = []
- for frame in frames:
- try:
- pil_img = Image.fromarray(frame.astype(np.uint8))
- H, W = frame.shape[:2]
-
- # Split frame into horizontal strips for text region detection
- strip_height = max(H // 4, 32)
- texts = []
- regions = []
- for y_start in range(0, H, strip_height):
- y_end = min(y_start + strip_height, H)
- strip = pil_img.crop((0, y_start, W, y_end))
- pixel_values = self.processor(
- images=strip, return_tensors="pt"
- ).pixel_values.to(self.device)
-
- with torch.no_grad():
- generated_ids = self.model.generate(pixel_values)
-
- text = self.processor.batch_decode(
- generated_ids, skip_special_tokens=True
- )[0].strip()
- if text:
- texts.append(text)
- regions.append(
- {
- "x": 0,
- "y": y_start,
- "w": W,
- "h": y_end - y_start,
- "text": text,
- }
- )
-
- results.append({"texts": texts, "regions": regions})
- except Exception as e:
- self.logger.error(f"OCR inference error on frame: {e}")
- results.append({"texts": [], "regions": []})
-
- return results[0] if not is_batch else results
-
-
-class OCRTransform(VideoTransform):
+class OCRTransform(VideoTransform, OcrTask):
"""
GStreamer element for optical character recognition on video frames.
@@ -130,6 +48,8 @@ class OCRTransform(VideoTransform):
(JSON with recognized text and regions).
"""
+ META_HEADER = OCR_META_HEADER
+
__gstmetadata__ = (
"OCR",
"Transform",
@@ -168,112 +88,10 @@ def engine_name(self):
def engine_name(self, value):
raise ValueError("'engine_name' is read-only for pyml_ocr")
- def do_transform_ip(self, buf):
- try:
- processor = MuxedBufferProcessor(
- self.logger, self.width, self.height, 30, 1
- )
- frames, _, num_sources, fmt = processor.extract_frames(buf, self.sinkpad)
- if frames is None:
- return Gst.FlowReturn.ERROR
-
- result = self._do_forward(frames)
- if result is None:
- return Gst.FlowReturn.ERROR
-
- if num_sources == 1:
- self._apply_ocr(buf, result, fmt, frames)
- else:
- if isinstance(result, list) and len(result) > 0:
- self._apply_ocr(
- buf, result[0], fmt, frames[0] if frames.ndim == 4 else frames
- )
-
- return Gst.FlowReturn.OK
-
- except Exception as e:
- self.logger.error(f"OCR transform error: {e}")
- return Gst.FlowReturn.ERROR
-
- def _do_forward(self, frames):
- if self.engine:
- return self.engine.do_forward(frames)
- return None
-
- def _apply_ocr(self, buf, result, fmt, frame):
- """Draw recognized text on frame and append OCR metadata."""
- import cv2
- import numpy as np
-
- regions = result.get("regions", [])
-
- # Draw text overlays before appending read-only metadata memory
- if self.draw_text and regions:
- overlay = frame.copy()
- for region in regions:
- x, y, w, h = region["x"], region["y"], region["w"], region["h"]
- text = region["text"]
- cv2.rectangle(overlay, (x, y), (x + w, y + h), (0, 255, 0), 2)
- font_scale = max(0.4, min(w / 300.0, 1.0))
- cv2.putText(
- overlay,
- text,
- (x + 4, y + h - 8),
- cv2.FONT_HERSHEY_SIMPLEX,
- font_scale,
- (0, 255, 0),
- 1,
- cv2.LINE_AA,
- )
-
- output = self._convert_rgb_to_format(overlay, fmt)
- if output is not None:
- success, map_info = buf.map(Gst.MapFlags.WRITE)
- if success:
- try:
- frame_bytes = np.ascontiguousarray(output).tobytes()
- dst = (ctypes.c_char * map_info.size).from_buffer(map_info.data)
- ctypes.memmove(
- dst, frame_bytes, min(len(frame_bytes), map_info.size)
- )
- finally:
- buf.unmap(map_info)
-
- # Append OCR results as a custom buffer memory chunk
- if regions:
- meta_bytes = OCR_META_HEADER + json.dumps(regions).encode("utf-8")
- tmp = Gst.Buffer.new_allocate(None, len(meta_bytes), None)
- tmp.fill(0, meta_bytes)
- buf.append_memory(tmp.get_memory(0))
-
- @staticmethod
- def _convert_rgb_to_format(rgb, fmt):
- """Convert an RGB numpy array to the target GStreamer video format."""
- import cv2
- import numpy as np
-
- if fmt == "RGB":
- return rgb
- elif fmt == "BGR":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
- elif fmt == "RGBA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- elif fmt == "BGRA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- elif fmt == "ARGB":
- rgba = cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- return np.roll(rgba, 1, axis=-1)
- elif fmt == "ABGR":
- bgra = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- return np.roll(bgra, 1, axis=-1)
- else:
- return rgb
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(OCRTransform)
- __gstelementfactory__ = ("pyml_ocr", Gst.Rank.NONE, OCRTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_ocr", OCRTransform)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_ocr' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/optical_flow.py b/plugins/python/optical_flow.py
index 2c27d5a..1b44974 100644
--- a/plugins/python/optical_flow.py
+++ b/plugins/python/optical_flow.py
@@ -17,23 +17,16 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import ctypes
-
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- from gi.repository import Gst, GObject
-
from video_transform import VideoTransform
from utils.format_converter import FormatConverter
- from utils.muxed_buffer_processor import MuxedBufferProcessor
- from engine.pytorch_engine import PyTorchEngine
+ from engine.optical_flow_engine import OpticalFlowEngine
from engine.engine_factory import EngineFactory
+ from backend import frameio, GObject
+ from tasks.optical_flow import OpticalFlowTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -50,74 +43,7 @@
}
-class OpticalFlowEngine(PyTorchEngine):
- """
- PyTorch engine for dense optical flow estimation using RAFT.
-
- Supports torchvision RAFT model variants:
- raft_large (most accurate)
- raft_small (fastest)
- """
-
- def do_load_model(self, model_name, **kwargs):
- try:
- from torchvision.models.optical_flow import (
- raft_large,
- raft_small,
- Raft_Large_Weights,
- Raft_Small_Weights,
- )
-
- if model_name == "raft_small":
- weights = Raft_Small_Weights.DEFAULT
- self.model = raft_small(weights=weights)
- else:
- weights = Raft_Large_Weights.DEFAULT
- self.model = raft_large(weights=weights)
-
- self.transforms = weights.transforms()
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.model.eval()
- self.logger.info(f"RAFT model '{model_name}' loaded on {self.device}")
- except Exception as e:
- raise ValueError(f"Failed to load RAFT model '{model_name}': {e}")
-
- def do_forward(self, prev_frame, curr_frame):
- import torch
-
- try:
- H, W = curr_frame.shape[:2]
-
- # Convert HWC uint8 -> CHW float tensor
- prev_t = torch.from_numpy(prev_frame).permute(2, 0, 1).float()
- curr_t = torch.from_numpy(curr_frame).permute(2, 0, 1).float()
-
- # RAFT requires dimensions divisible by 8
- pad_h = (8 - H % 8) % 8
- pad_w = (8 - W % 8) % 8
- if pad_h > 0 or pad_w > 0:
- prev_t = torch.nn.functional.pad(prev_t, (0, pad_w, 0, pad_h))
- curr_t = torch.nn.functional.pad(curr_t, (0, pad_w, 0, pad_h))
-
- prev_t, curr_t = self.transforms(prev_t, curr_t)
- prev_batch = prev_t.unsqueeze(0).to(self.device)
- curr_batch = curr_t.unsqueeze(0).to(self.device)
-
- with torch.no_grad():
- flow_predictions = self.model(prev_batch, curr_batch)
-
- # RAFT returns a list of flow predictions; take the last (finest)
- flow = flow_predictions[-1].squeeze(0).cpu().numpy()
- # flow shape: (2, H', W') -> transpose to (H, W, 2) and crop
- flow = flow.transpose(1, 2, 0)[:H, :W]
- return flow
-
- except Exception as e:
- self.logger.error(f"Optical flow inference error: {e}")
- return None
-
-
-class OpticalFlowTransform(VideoTransform):
+class OpticalFlowTransform(VideoTransform, OpticalFlowTask):
"""
GStreamer element for dense optical flow estimation using RAFT.
@@ -167,107 +93,32 @@ def engine_name(self):
def engine_name(self, value):
raise ValueError("'engine_name' is read-only for pyml_optical_flow")
- def do_transform_ip(self, buf):
- try:
- processor = MuxedBufferProcessor(
- self.logger, self.width, self.height, 30, 1
- )
- frames, _, num_sources, fmt = processor.extract_frames(buf, self.sinkpad)
- if frames is None:
- return Gst.FlowReturn.ERROR
-
- frame = frames[0] if frames.ndim == 4 else frames
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Pair this frame with the previous one and draw the flow overlay."""
+ frame = frames[0] if frames.ndim == 4 else frames
- if self._prev_frame is None:
- self._prev_frame = frame.copy()
- return Gst.FlowReturn.OK
-
- flow = self._do_forward(self._prev_frame, frame)
+ # Temporal pairing stays in the shell: hold the previous frame.
+ if self._prev_frame is None:
self._prev_frame = frame.copy()
+ return
- if flow is None:
- return Gst.FlowReturn.OK
-
- if self.visualize:
- self._apply_flow_vis(buf, flow, frame, fmt)
-
- return Gst.FlowReturn.OK
-
- except Exception as e:
- self.logger.error(f"Optical flow transform error: {e}")
- return Gst.FlowReturn.ERROR
+ flow = self.forward(self._prev_frame, frame)
+ self._prev_frame = frame.copy()
- def _do_forward(self, prev_frame, curr_frame):
- if self.engine:
- return self.engine.do_forward(prev_frame, curr_frame)
- return None
+ if flow is None:
+ return
- def _apply_flow_vis(self, buf, flow, frame, fmt):
- """Render flow as a color overlay and write back to buffer."""
- import cv2
- import numpy as np
+ if self.visualize:
+ # Portable task: render the flow overlay frame.
+ output, blob = self.decode(flow, frame, fmt)
+ frameio.write_result(target, output, blob)
- flow_vis = self._flow_to_color(flow)
- blended = cv2.addWeighted(frame, 0.5, flow_vis, 0.5, 0)
- output = self._convert_rgb_to_format(blended, fmt)
- if output is not None:
- success, map_info = buf.map(Gst.MapFlags.WRITE)
- if success:
- try:
- frame_bytes = np.ascontiguousarray(output).tobytes()
- dst = (ctypes.c_char * map_info.size).from_buffer(map_info.data)
- ctypes.memmove(
- dst, frame_bytes, min(len(frame_bytes), map_info.size)
- )
- finally:
- buf.unmap(map_info)
- @staticmethod
- def _flow_to_color(flow):
- """Convert optical flow (H, W, 2) to an RGB color image using HSV encoding."""
- import cv2
- import numpy as np
-
- fx, fy = flow[..., 0], flow[..., 1]
- mag = np.sqrt(fx**2 + fy**2)
- ang = np.arctan2(fy, fx)
-
- hsv = np.zeros((*flow.shape[:2], 3), dtype=np.uint8)
- hsv[..., 0] = ((ang + np.pi) / (2 * np.pi) * 179).astype(np.uint8)
- hsv[..., 1] = 255
- mag_norm = mag / (mag.max() + 1e-8)
- hsv[..., 2] = (mag_norm * 255).astype(np.uint8)
-
- return cv2.cvtColor(hsv, cv2.COLOR_HSV2RGB)
-
- @staticmethod
- def _convert_rgb_to_format(rgb, fmt):
- """Convert an RGB numpy array to the target GStreamer video format."""
- import cv2
- import numpy as np
-
- if fmt == "RGB":
- return rgb
- elif fmt == "BGR":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
- elif fmt == "RGBA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- elif fmt == "BGRA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- elif fmt == "ARGB":
- rgba = cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- return np.roll(rgba, 1, axis=-1)
- elif fmt == "ABGR":
- bgra = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- return np.roll(bgra, 1, axis=-1)
- else:
- return rgb
-
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(OpticalFlowTransform)
- __gstelementfactory__ = ("pyml_optical_flow", Gst.Rank.NONE, OpticalFlowTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_optical_flow", OpticalFlowTransform
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_optical_flow' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/overlay.py b/plugins/python/overlay.py
index e409074..5417f04 100644
--- a/plugins/python/overlay.py
+++ b/plugins/python/overlay.py
@@ -17,6 +17,7 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
from utils.analytics_utils import ANALYTICS_UTILS_AVAILABLE
@@ -46,8 +47,8 @@
GstVideo,
GstGL,
GstVulkan,
- GObject,
) # noqa: E402
+ from backend import GObject
from log.logger_factory import LoggerFactory
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -57,7 +58,9 @@
# Support CPU, OpenGL, and Vulkan buffers
VIDEO_FORMATS = "video/x-raw, format=(string){ RGBA, ARGB, BGRA, ABGR }; video/x-raw(memory:GLMemory), format=(string){ RGBA, ARGB, BGRA, ABGR }; video/x-raw(memory:VulkanMemory), format=(string){ RGBA, ARGB, BGRA, ABGR }"
-OVERLAY_CAPS = Gst.Caps.from_string(VIDEO_FORMATS)
+# Building a Gst object needs Gst.init, which only the gst backend calls.
+if backend.BACKEND == "gst":
+ OVERLAY_CAPS = Gst.Caps.from_string(VIDEO_FORMATS)
class Overlay(GstBase.BaseTransform):
@@ -68,20 +71,21 @@ class Overlay(GstBase.BaseTransform):
"Aaron Boxer ",
)
- src_template = Gst.PadTemplate.new(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- OVERLAY_CAPS.copy(),
- )
+ if backend.BACKEND == "gst":
+ src_template = Gst.PadTemplate.new(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ OVERLAY_CAPS.copy(),
+ )
- sink_template = Gst.PadTemplate.new(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.ALWAYS,
- OVERLAY_CAPS.copy(),
- )
- __gsttemplates__ = (src_template, sink_template)
+ sink_template = Gst.PadTemplate.new(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.ALWAYS,
+ OVERLAY_CAPS.copy(),
+ )
+ __gsttemplates__ = (src_template, sink_template)
meta_path = GObject.Property(
type=str,
@@ -445,10 +449,9 @@ def do_post_process(self, frame_metadata):
self.tracking_display.fade_history()
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(Overlay)
- __gstelementfactory__ = ("pyml_overlay", Gst.Rank.NONE, Overlay)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_overlay", Overlay)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_overlay' element will not be registered because a module is missing."
)
diff --git a/plugins/python/overlay_counter.py b/plugins/python/overlay_counter.py
index b49933e..6ce89f0 100644
--- a/plugins/python/overlay_counter.py
+++ b/plugins/python/overlay_counter.py
@@ -17,11 +17,10 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- from gi.repository import Gst, GObject
-
from overlay import Overlay
from overlay_helper.overlay_utils_interface import Color
except ImportError as e:
@@ -32,7 +31,6 @@
class OverlayCounter(Overlay):
-
__gstmetadata__ = (
"OverlayCounter",
"Filter/Effect/Video",
@@ -68,8 +66,9 @@ def do_post_process(self, frame_metadata):
self.overlay_graphics.draw_text(text, 0, 50, Color(1, 0, 0, 1), 20)
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(OverlayCounter)
- __gstelementfactory__ = ("pyml_overlay_counter", Gst.Rank.NONE, OverlayCounter)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_overlay_counter", OverlayCounter
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning("Failed to register the 'pyml_overlay_counter' element.")
diff --git a/plugins/python/overlay_helper/overlay_skia.py b/plugins/python/overlay_helper/overlay_skia.py
index 5454210..e1a53e1 100644
--- a/plugins/python/overlay_helper/overlay_skia.py
+++ b/plugins/python/overlay_helper/overlay_skia.py
@@ -17,6 +17,7 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
@@ -38,8 +39,8 @@
GstVideo,
GstAnalytics,
GLib,
- GObject,
) # noqa: E402
+ from backend import GObject
from log.logger_factory import LoggerFactory
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -347,10 +348,11 @@ def draw_text_with_cairo(self, cr, label, x, y):
cr.stroke()
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(OverlaySkia)
- __gstelementfactory__ = ("pyml_overlay_skia", Gst.Rank.NONE, OverlaySkia)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_overlay_skia", OverlaySkia
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_overlay_skia' element will not be registered because a required module is missing."
)
diff --git a/plugins/python/pose.py b/plugins/python/pose.py
index cb4a8be..83e5148 100644
--- a/plugins/python/pose.py
+++ b/plugins/python/pose.py
@@ -17,6 +17,7 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
@@ -28,13 +29,12 @@
gi.require_version("Gst", "1.0")
gi.require_version("GstBase", "1.0")
gi.require_version("GstVideo", "1.0")
- gi.require_version("GstAnalytics", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject, GstAnalytics, GLib
+ from gi.repository import Gst
+ from backend import analytics, GObject
from base_objectdetector import BaseObjectDetector
from utils.format_converter import FormatConverter
- from engine.pytorch_engine import PyTorchEngine
+ from engine.yolo_pose_engine import YoloPoseEngine
from engine.engine_factory import EngineFactory
except ImportError as e:
@@ -86,46 +86,6 @@
]
-class YoloPoseEngine(PyTorchEngine):
- """PyTorch engine for YOLO pose estimation models."""
-
- def do_load_model(self, model_name, **kwargs):
- try:
- from ultralytics import YOLO
-
- self.model = YOLO(f"{model_name}.pt")
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.logger.info(f"YOLO pose model '{model_name}' loaded on {self.device}")
- except Exception as e:
- raise ValueError(f"Failed to load YOLO pose model '{model_name}': {e}")
-
- def do_forward(self, frames):
- import numpy as np
-
- is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
- writable = np.array(frames, copy=True)
- batch_size = writable.shape[0] if is_batch else 1
-
- model = self.get_model()
- if model is None:
- self.logger.error("Pose model not loaded")
- return None if not is_batch else [None] * batch_size
-
- try:
- img_list = (
- [writable[i] for i in range(batch_size)] if is_batch else [writable]
- )
- results = self.execute_with_stream(
- lambda: model(img_list, imgsz=640, conf=0.25, verbose=False)
- )
- if not results:
- return None if not is_batch else [None] * batch_size
- return results[0] if not is_batch else results
- except Exception as e:
- self.logger.error(f"Pose inference error: {e}")
- return None if not is_batch else [None] * batch_size
-
-
class YOLOPoseTransform(BaseObjectDetector):
"""
GStreamer element for human pose estimation using YOLO on video frames.
@@ -179,7 +139,7 @@ def do_decode(self, buf, result, stream_idx=0):
return
# Attach person bounding boxes via GstAnalytics (compatible with pyml_overlay)
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
+ meta = analytics.add_relation_meta(buf)
if not meta:
self.logger.error("Failed to add analytics relation metadata")
return
@@ -189,16 +149,16 @@ def do_decode(self, buf, result, stream_idx=0):
x1, y1, x2, y2 = boxes.xyxy[i]
score = boxes.conf[i].item()
- qk = GLib.quark_from_string(f"stream_{stream_idx}_person")
- ret, _ = meta.add_od_mtd(
- qk,
+ mtd = analytics.add_object(
+ meta,
+ f"stream_{stream_idx}_person",
x1.item(),
y1.item(),
x2.item() - x1.item(),
y2.item() - y1.item(),
score,
)
- if not ret:
+ if mtd is None:
self.logger.error(f"Failed to add od_mtd for person {i}")
continue
@@ -291,10 +251,11 @@ def _convert_bgr_to_format(bgr, fmt):
return cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(YOLOPoseTransform)
- __gstelementfactory__ = ("pyml_yolo_pose", Gst.Rank.NONE, YOLOPoseTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_yolo_pose", YOLOPoseTransform
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_yolo_pose' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/pyml_launch.py b/plugins/python/pyml_launch.py
new file mode 100644
index 0000000..c5cb298
--- /dev/null
+++ b/plugins/python/pyml_launch.py
@@ -0,0 +1,532 @@
+# pyml-launch
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Run one pipeline description on whichever backend `PYML_BACKEND` selects.
+
+`gst` hands the pipeline to `gst-launch-1.0` unchanged. `g2g` rewrites it for
+`g2g-launch-py`, which spells five things differently:
+
+ * a `pyml_*` element is the generic `pyelement` host plus the module and class
+ to load;
+ * `pyml_overlay` is g2g's own `analyticsoverlay`, a native element with its own
+ properties rather than a hosted one;
+ * `pyelement` and `analyticsoverlay` both work on RGBA, so a raw-video caps
+ filter that leaves the format open has to pin it;
+ * a hosted element carries no pad templates into g2g, so an element that
+ declares `INPUT_CAPS` / `OUTPUT_CAPS` hands them to its host as properties,
+ which is the only way it can take audio in and give text out;
+ * a sink takes no clock properties and the overlay no `wait-text`, because
+ g2g never does what those switch off.
+
+Everything else is named the same on both: `filesrc`, `decodebin`,
+`videoconvert`, `videoscale`, `autovideosink`. Write the pipeline the way the
+README does, in `gst-launch` spelling, and this translates it. Either way the
+line the backend runs is printed, so it can be pasted back and extended.
+"""
+
+import ast
+import os
+import shlex
+import shutil
+import site
+import subprocess
+import sys
+from pathlib import Path
+from typing import NamedTuple
+
+#: The checkout's `plugins` tree, when this module was loaded from one. An
+#: installed copy sits in site-packages, which carries no element modules.
+_PLUGINS = Path(__file__).resolve().parent.parent
+CHECKOUT_PLUGINS = _PLUGINS if _PLUGINS.name == "plugins" else None
+
+USAGE = "usage: pyml-launch [key=value ...] ! ! ..."
+
+#: The g2g launcher, and the checkout it is built in, looked for beside this one.
+G2G_LAUNCH_NAME = "g2g-launch-py"
+G2G_CHECKOUT_NAME = "glass2glass"
+
+#: The g2g elements that host a gst-python-ml one: a transform, and the N-in
+#: batching host an aggregator needs.
+PY_ELEMENT = "pyelement"
+PY_AGGREGATOR = "pyaggregator"
+
+#: gst-python-ml elements that g2g implements natively instead of hosting.
+NATIVE_EQUIVALENTS = {"pyml_overlay": "analyticsoverlay"}
+
+#: What each native element calls the properties it shares with the one it
+#: replaces. A property missing here has no counterpart at all.
+NATIVE_PROPERTIES = {"pyml_overlay": {"tracking": "show-track"}}
+
+#: What separates two elements, on both spellings.
+SEPARATOR = "!"
+
+#: `(element name suffix, property)` a g2g pipeline does not spell, mapped to
+#: what to do when the behaviour it switches off is actually wanted.
+DEFAULTED_PROPERTIES = {
+ ("sink", "sync"): "a g2g sink does not wait on the clock, put a `clocksync` "
+ "element ahead of it instead",
+ ("sink", "async"): "a g2g sink does not wait on the clock, put a `clocksync` "
+ "element ahead of it instead",
+ ("textoverlay", "wait-text"): "g2g's textoverlay never holds video back for "
+ "the text pad",
+}
+
+FALSE_VALUES = ("false", "0", "no", "off")
+
+#: `pyelement` hosts an RGBA frame by default and `analyticsoverlay` draws on
+#: RGBA8 only, so a caps filter that names no format would fail to negotiate.
+G2G_RAW_VIDEO_FORMAT = "RGBA"
+
+#: The class constant an element states each of its pad caps in, and the host
+#: property it becomes. Each value is one `gst-launch` caps description, which
+#: g2g refuses unless it names exactly one concrete caps.
+G2G_CAPS_PROPERTIES = {
+ "INPUT_CAPS": "input-caps",
+ "OUTPUT_CAPS": "output-caps",
+}
+
+
+class ElementShell(NamedTuple):
+ """What hosting one gst-python-ml element on g2g takes: the module and class
+ to load, and the `(property, caps)` pairs it negotiates with, if it says."""
+
+ module: str
+ cls: str
+ caps: tuple = ()
+
+
+def plugin_dir():
+ """The directory holding the element modules: the checkout this module came
+ from, or the `python` subdirectory of a `GST_PLUGIN_PATH` entry."""
+ if CHECKOUT_PLUGINS:
+ return CHECKOUT_PLUGINS / "python"
+ for entry in os.environ.get("GST_PLUGIN_PATH", "").split(os.pathsep):
+ candidate = Path(entry) / "python"
+ if candidate.is_dir():
+ return candidate
+ raise SystemExit(
+ "pyml-launch: no gst-python-ml element modules found; point "
+ "GST_PLUGIN_PATH at the checkout's plugins directory"
+ )
+
+
+def element_shells(directory=None):
+ """Map each element name to the `ElementShell` hosting it takes.
+
+ Read out of the sources with `ast` rather than by importing: a plugin pulls
+ in torch and friends, and every one whose dependencies are missing would
+ drop out of the map.
+ """
+ shells = {}
+ registered = []
+ declarations = {}
+ for path in sorted((directory or plugin_dir()).glob("*.py")):
+ tree = ast.parse(path.read_text(), filename=str(path))
+ constants = string_constants(tree)
+ declarations.update(caps_declarations(tree))
+ for node in ast.walk(tree):
+ if not isinstance(node, ast.Call):
+ continue
+ called = node.func
+ name = (
+ called.attr
+ if isinstance(called, ast.Attribute)
+ else getattr(called, "id", None)
+ )
+ if name != "register_gst_element" or len(node.args) < 2:
+ continue
+ element, cls = node.args[0], node.args[1]
+ element = element_name(element, constants)
+ if element and isinstance(cls, ast.Name):
+ registered.append((element, path.stem, cls.id))
+ for element, module, cls in registered:
+ shells[element] = ElementShell(module, cls, declared_caps(cls, declarations))
+ return shells
+
+
+def caps_declarations(tree):
+ """Each class in the module: the classes it derives from, and the g2g caps it
+ states itself."""
+ declarations = {}
+ for node in ast.walk(tree):
+ if not isinstance(node, ast.ClassDef):
+ continue
+ constants = string_constants(node)
+ caps = {
+ property: constants[f"{node.name}.{constant}"]
+ for constant, property in G2G_CAPS_PROPERTIES.items()
+ if f"{node.name}.{constant}" in constants
+ }
+ bases = [base.id for base in node.bases if isinstance(base, ast.Name)]
+ declarations[node.name] = (bases, caps)
+ return declarations
+
+
+def declared_caps(cls, declarations):
+ """The `(property, caps)` pairs a class negotiates with, the ones it states
+ itself ahead of the ones it inherits.
+
+ A leaf that changes one pad restates only that pad, so a class taking text in
+ and giving audio out gets its input from the base and its output from itself.
+ """
+ caps = {}
+ pending, seen = [cls], set()
+ while pending:
+ name = pending.pop(0)
+ if name in seen or name not in declarations:
+ continue
+ seen.add(name)
+ bases, stated = declarations[name]
+ for property, value in stated.items():
+ caps.setdefault(property, value)
+ pending.extend(bases)
+ return tuple(
+ (property, caps[property])
+ for property in G2G_CAPS_PROPERTIES.values()
+ if property in caps
+ )
+
+
+def string_constants(tree):
+ """Every `NAME = "..."` under the node, class attributes keyed `Class.NAME`.
+
+ Some elements register under a class constant rather than a literal, so the
+ name has to be resolved before it can go in the map.
+ """
+ constants = {}
+ for node in ast.walk(tree):
+ if isinstance(node, ast.ClassDef):
+ prefix = f"{node.name}."
+ body = node.body
+ elif isinstance(node, ast.Module):
+ prefix = ""
+ body = node.body
+ else:
+ continue
+ for statement in body:
+ if not isinstance(statement, ast.Assign):
+ continue
+ if not isinstance(statement.value, ast.Constant):
+ continue
+ if not isinstance(statement.value.value, str):
+ continue
+ for target in statement.targets:
+ if isinstance(target, ast.Name):
+ constants[f"{prefix}{target.id}"] = statement.value.value
+ return constants
+
+
+def element_name(node, constants):
+ """The element name a `register_gst_element` first argument spells, whether
+ it is a literal or a constant defined in the same module."""
+ if isinstance(node, ast.Constant) and isinstance(node.value, str):
+ return node.value
+ if isinstance(node, ast.Attribute) and isinstance(node.value, ast.Name):
+ return constants.get(f"{node.value.id}.{node.attr}")
+ if isinstance(node, ast.Name):
+ return constants.get(node.id)
+ return None
+
+
+def rewrite_for_g2g(argv, shells):
+ """Translate `gst-launch` arguments into the ones `g2g-launch-py` takes.
+
+ Works on the argument list rather than one joined string so a property value
+ the shell already unquoted (`labels="person, bicycle"`) keeps its boundary,
+ and gets requoted on the way out.
+ """
+ rewritten = []
+ raw_format = None
+ parts = list(segments(argv))
+ fan_in = fan_in_names(parts)
+ for segment in parts:
+ # Separators and references are carried through where they were
+ # written: which chains link to which is the caller's, not ours.
+ if segment == [SEPARATOR] or is_reference(segment[0]):
+ rewritten.extend(segment)
+ continue
+ host = PY_AGGREGATOR if declared_name(segment) in fan_in else PY_ELEMENT
+ translated = rewrite_segment(segment, shells, raw_format, host)
+ rewritten.extend(translated)
+ raw_format = raw_video_format(translated[0]) or raw_format
+ return rewritten
+
+
+def declared_name(segment):
+ """The handle a segment gives itself with `name=`, if any."""
+ for token in segment[1:]:
+ key, separator, value = token.partition("=")
+ if separator and key == "name":
+ return value
+ return None
+
+
+def fan_in_names(parts):
+ """The element handles several chains feed into.
+
+ A chain *ending* in a bare `mux.` reference links into that element, so an
+ element named by more than one is taking more than its own chain's input.
+ That is the one shape g2g hosts on `pyaggregator` (its N-in batching host)
+ rather than the one-in `pyelement`. A reference that *starts* a chain
+ (`cap.text_src ! ...`) reads the other way, out of a second source pad, and
+ leaves the element a one-in transform.
+ """
+ feeding = [
+ segment[0].partition(".")[0]
+ for previous, segment in zip([None, *parts], parts)
+ if previous == [SEPARATOR] and len(segment) == 1 and is_reference(segment[0])
+ ]
+ return {name for name in feeding if feeding.count(name) > 1}
+
+
+def raw_video_format(caps):
+ """The format a raw-video caps filter pins, or `None` for anything else."""
+ if not caps.startswith("video/x-raw"):
+ return None
+ for field in caps.split(","):
+ key, separator, value = field.partition("=")
+ if separator and key == "format":
+ return value
+ return None
+
+
+def segments(argv):
+ """The argument list split into one token list per element, `!`, or
+ reference.
+
+ A `!` separates two elements, and a bare `name.` reference is a chain
+ boundary of its own: `mux. ! sink` starts a chain at that element and
+ `... ! mux.` ends one there. Both come back as their own single-token
+ segment, so the links the caller wrote survive the rewrite unchanged.
+ """
+ segment = []
+ for token in argv:
+ if token == SEPARATOR or is_reference(token):
+ if segment:
+ yield segment
+ yield [token]
+ segment = []
+ else:
+ segment.append(token)
+ if segment:
+ yield segment
+
+
+def is_reference(token):
+ """Whether the token is a `name.` / `name.pad` reference to an element."""
+ name, separator, pad = token.partition(".")
+ return bool(
+ separator
+ and name[:1].isalpha()
+ and name.replace("_", "").replace("-", "").isalnum()
+ and (not pad or pad.replace("_", "").isalnum())
+ )
+
+
+def rewrite_segment(segment, shells, raw_format=None, host=None):
+ if isinstance(segment, str):
+ segment = segment.split()
+ if not segment:
+ return []
+ head, properties = segment[0], segment[1:]
+
+ # A caps filter is the one segment whose first token is a media type; no
+ # element name contains a slash. Caps carry no spaces, so a caps written
+ # with one (`video/x-raw, width=320`) arrives as several tokens to rejoin.
+ if "/" in head:
+ return [with_pinned_format("".join(segment))]
+
+ native = NATIVE_EQUIVALENTS.get(head)
+ if native:
+ return [native, *[quoted(p) for p in renamed_properties(head, properties)]]
+
+ shell = shells.get(head)
+ if shell:
+ # Every property of a hosted element goes to the Python class, so the
+ # defaulted-property table must not touch one whose name happens to end
+ # like a g2g element's (`pyml_kafkasink` and its `sync`).
+ properties = [quoted(p) for p in properties]
+ hosted = [host or PY_ELEMENT, f"module={shell.module}", f"class={shell.cls}"]
+ # `pyelement` takes RGBA unless told otherwise, so an upstream caps
+ # filter naming another format has to reach the hosted element too.
+ # `pyaggregator` negotiates from its inputs and takes no format.
+ pin = raw_format not in (None, G2G_RAW_VIDEO_FORMAT)
+ if hosted[0] == PY_ELEMENT and pin:
+ if not any(p.startswith("format=") for p in properties):
+ hosted.append(f"format={raw_format}")
+ hosted.extend(quoted(f"{property}={value}") for property, value in shell.caps)
+ return [*hosted, *properties]
+
+ if head.startswith("pyml_"):
+ raise SystemExit(f"pyml-launch: no gst-python-ml element named {head!r}")
+
+ return [head, *[quoted(p) for p in drop_defaulted_properties(head, properties)]]
+
+
+def renamed_properties(head, properties):
+ """The properties under the names g2g's own element spells them.
+
+ A native equivalent is a different element, not a rename, so only the knobs
+ that mean the same thing carry over. One that does not is an error rather
+ than a guess, because dropping it would run a pipeline that quietly does
+ something else.
+ """
+ renames = NATIVE_PROPERTIES.get(head, {})
+ renamed = []
+ for property in properties:
+ key, _, value = property.partition("=")
+ if key not in renames:
+ raise SystemExit(
+ f"pyml-launch: g2g's {NATIVE_EQUIVALENTS[head]} has no counterpart "
+ f"for {head} {key}; of its properties only "
+ f"{', '.join(sorted(renames))} carries over"
+ )
+ renamed.append(f"{renames[key]}={value}")
+ return renamed
+
+
+def drop_defaulted_properties(head, properties):
+ """The properties minus the ones naming behaviour g2g already has.
+
+ Each of these switches off something g2g never does, so turning it off is a
+ no-op there. Asking for it back is refused rather than dropped quietly,
+ because that needs a pipeline change this cannot make on its own.
+ """
+ kept = []
+ for property in properties:
+ key, _, value = property.partition("=")
+ instead = next(
+ (
+ instead
+ for (suffix, name), instead in DEFAULTED_PROPERTIES.items()
+ if name == key and head.endswith(suffix)
+ ),
+ None,
+ )
+ if instead is None:
+ kept.append(property)
+ elif value.lower() not in FALSE_VALUES:
+ raise SystemExit(
+ f"pyml-launch: {head} {property} has no equivalent on g2g; {instead}"
+ )
+ return kept
+
+
+def quoted(token):
+ """A token spelled so g2g's launch parser reads the value back unchanged.
+
+ Its grammar treats a quote, a `\\`, a `!` or a `#` as syntax wherever it
+ appears, and the argument list is joined with spaces before parsing, so a
+ value carrying any of those has to say so.
+ """
+ key, separator, value = token.partition("=")
+ if not separator:
+ return token
+ for character in ("\\", '"', "'", "!", "#"):
+ value = value.replace(character, f"\\{character}")
+ if any(character.isspace() for character in value):
+ value = f'"{value}"'
+ return f"{key}={value}"
+
+
+def with_pinned_format(caps):
+ if not caps.startswith("video/x-raw") or "format=" in caps:
+ return caps
+ return f"{caps},format={G2G_RAW_VIDEO_FORMAT}"
+
+
+def _prepend_path(existing, *entries):
+ paths = [str(entry) for entry in entries]
+ if existing:
+ paths.append(existing)
+ return os.pathsep.join(paths)
+
+
+def launch_environment():
+ """The environment both launchers need.
+
+ Neither runs under this interpreter: GStreamer's Python loader embeds the
+ system one and g2g embeds its own, so torch and the rest are only reachable
+ if the venv's site directories are on `PYTHONPATH`.
+ """
+ env = os.environ.copy()
+ env["PYTHONPATH"] = _prepend_path(
+ env.get("PYTHONPATH"), plugin_dir(), *site.getsitepackages()
+ )
+ return env
+
+
+def gst_command(argv):
+ env = launch_environment()
+ if CHECKOUT_PLUGINS:
+ env["GST_PLUGIN_PATH"] = _prepend_path(
+ env.get("GST_PLUGIN_PATH"), CHECKOUT_PLUGINS
+ )
+ return ["gst-launch-1.0", *argv], env
+
+
+def g2g_binary():
+ """The `g2g-launch-py` to run: an explicit `G2G_LAUNCH`, else the release
+ build in the glass2glass checkout, else whatever is on `PATH`."""
+ explicit = os.environ.get("G2G_LAUNCH")
+ if explicit:
+ return explicit
+ checkout = os.environ.get("G2G_DIR")
+ if not checkout and CHECKOUT_PLUGINS:
+ checkout = CHECKOUT_PLUGINS.parent.parent / G2G_CHECKOUT_NAME
+ if checkout:
+ release = Path(checkout) / "target" / "release" / G2G_LAUNCH_NAME
+ if release.is_file():
+ return str(release)
+ return shutil.which(G2G_LAUNCH_NAME)
+
+
+def g2g_command(pipeline):
+ binary = g2g_binary()
+ if not binary:
+ raise SystemExit(
+ "pyml-launch: no g2g-launch-py release build found; point G2G_LAUNCH "
+ "at one, or G2G_DIR at a glass2glass checkout, and build it with: "
+ "PYO3_PYTHON=$(which python) cargo build --release -p g2g-python "
+ "--features ml --bin g2g-launch-py"
+ )
+ return [binary, *pipeline], launch_environment()
+
+
+def main(argv=None):
+ argv = sys.argv[1:] if argv is None else list(argv)
+ if not argv:
+ raise SystemExit(USAGE)
+
+ backend = os.environ.get("PYML_BACKEND", "gst").lower()
+ if backend == "gst":
+ command, env = gst_command(argv)
+ elif backend == "g2g":
+ command, env = g2g_command(rewrite_for_g2g(argv, element_shells()))
+ else:
+ raise SystemExit(
+ f"pyml-launch: unknown PYML_BACKEND={backend!r}; use 'gst' or 'g2g'"
+ )
+
+ # The line the backend actually runs, so it can be pasted back and extended.
+ print(f"pyml-launch: {shlex.join(command)}", file=sys.stderr)
+ return subprocess.call(command, env=env)
+
+
+if __name__ == "__main__":
+ sys.exit(main())
diff --git a/plugins/python/sam.py b/plugins/python/sam.py
index 81be32b..1d74bc6 100644
--- a/plugins/python/sam.py
+++ b/plugins/python/sam.py
@@ -17,24 +17,16 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import ctypes
- import json
-
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- from gi.repository import Gst, GObject
-
from video_transform import VideoTransform
from utils.format_converter import FormatConverter
- from utils.muxed_buffer_processor import MuxedBufferProcessor
- from engine.pytorch_engine import PyTorchEngine
+ from engine.sam_engine import SamEngine
from engine.engine_factory import EngineFactory
+ from backend import GObject
+ from tasks.sam import SamTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -43,109 +35,8 @@
# Header prefix for segmentation mask buffer metadata
SAM_META_HEADER = b"GST-SAM:"
-# Mask overlay colors (BGR) cycled across detected objects
-MASK_COLORS = [
- (255, 0, 0),
- (0, 255, 0),
- (0, 0, 255),
- (255, 255, 0),
- (0, 255, 255),
- (255, 0, 255),
- (128, 0, 255),
- (255, 128, 0),
- (0, 128, 255),
- (128, 255, 0),
-]
-
-
-class SamEngine(PyTorchEngine):
- """
- PyTorch engine for Segment Anything Model 2 (SAM2).
-
- Supports HuggingFace model IDs:
- facebook/sam2-hiera-large
- facebook/sam2-hiera-base-plus
- facebook/sam2-hiera-small
- facebook/sam2-hiera-tiny
- """
-
- def do_load_model(self, model_name, **kwargs):
- try:
- from transformers import Sam2Model, Sam2Processor
-
- self.processor = Sam2Processor.from_pretrained(model_name)
- self.model = Sam2Model.from_pretrained(model_name)
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.model.eval()
- self.logger.info(f"SAM2 model '{model_name}' loaded on {self.device}")
- except Exception as e:
- raise ValueError(f"Failed to load SAM2 model '{model_name}': {e}")
-
- def do_forward(self, frames, max_masks=10):
- import numpy as np
- import torch
- from PIL import Image
-
- is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
- if not is_batch:
- frames = frames[np.newaxis]
-
- results = []
- for frame in frames:
- try:
- pil_img = Image.fromarray(frame.astype(np.uint8))
- H, W = frame.shape[:2]
-
- # Automatic mask generation: grid of input points
- grid_size = int(np.ceil(np.sqrt(max_masks)))
- xs = np.linspace(0, W - 1, grid_size).astype(int)
- ys = np.linspace(0, H - 1, grid_size).astype(int)
- points = [[int(x), int(y)] for y in ys for x in xs][:max_masks]
- input_points = [points]
-
- inputs = self.processor(
- images=pil_img,
- input_points=input_points,
- return_tensors="pt",
- )
- inputs = {k: v.to(self.device) for k, v in inputs.items()}
-
- with torch.no_grad():
- outputs = self.model(**inputs)
-
- masks = self.processor.post_process_masks(
- outputs.pred_masks,
- inputs["original_sizes"],
- inputs["reshaped_input_sizes"],
- )
- scores = outputs.iou_scores
- mask_list = []
- if len(masks) > 0:
- frame_masks = masks[0].cpu().numpy()
- frame_scores = scores[0].cpu().numpy()
- for j in range(min(frame_masks.shape[0], max_masks)):
- best_idx = frame_scores[j].argmax()
- mask = frame_masks[j, best_idx]
- score = float(frame_scores[j, best_idx])
- mask_list.append(
- {"mask_idx": j, "score": score, "shape": list(mask.shape)}
- )
-
- results.append(
- {
- "masks": mask_list,
- "raw_masks": masks[0].cpu().numpy() if len(masks) > 0 else None,
- }
- )
- except Exception as e:
- self.logger.error(f"SAM inference error on frame: {e}")
- results.append({"masks": [], "raw_masks": None})
-
- return results[0] if not is_batch else results
-
-
-class SamTransform(VideoTransform):
+class SamTransform(VideoTransform, SamTask):
"""
GStreamer element for image segmentation using Segment Anything Model 2.
@@ -157,6 +48,8 @@ class SamTransform(VideoTransform):
GST-SAM: memory chunk (JSON with mask scores and shapes).
"""
+ META_HEADER = SAM_META_HEADER
+
__gstmetadata__ = (
"SAM Segmentation",
"Transform",
@@ -205,110 +98,10 @@ def engine_name(self):
def engine_name(self, value):
raise ValueError("'engine_name' is read-only for pyml_sam")
- def do_transform_ip(self, buf):
- try:
- processor = MuxedBufferProcessor(
- self.logger, self.width, self.height, 30, 1
- )
- frames, _, num_sources, fmt = processor.extract_frames(buf, self.sinkpad)
- if frames is None:
- return Gst.FlowReturn.ERROR
-
- result = self._do_forward(frames)
- if result is None:
- return Gst.FlowReturn.ERROR
-
- if num_sources == 1:
- self._apply_masks(buf, result, fmt, frames)
- else:
- if isinstance(result, list) and len(result) > 0:
- self._apply_masks(
- buf, result[0], fmt, frames[0] if frames.ndim == 4 else frames
- )
-
- return Gst.FlowReturn.OK
-
- except Exception as e:
- self.logger.error(f"SAM transform error: {e}")
- return Gst.FlowReturn.ERROR
-
- def _do_forward(self, frames):
- if self.engine:
- return self.engine.do_forward(frames, max_masks=self.max_masks)
- return None
-
- def _apply_masks(self, buf, result, fmt, frame):
- """Overlay masks on frame and append mask metadata."""
- import cv2
- import numpy as np
-
- raw_masks = result.get("raw_masks")
- mask_info = result.get("masks", [])
-
- # Draw mask overlays before appending read-only metadata memory
- if self.visualize and raw_masks is not None:
- overlay = frame.copy()
- for j in range(min(raw_masks.shape[0], self.max_masks)):
- best_idx = 0
- if raw_masks.ndim == 4 and raw_masks.shape[1] > 1:
- best_idx = raw_masks[j].sum(axis=(1, 2)).argmax()
- mask = raw_masks[j, best_idx]
- color = MASK_COLORS[j % len(MASK_COLORS)]
- colored = np.zeros_like(overlay)
- colored[:] = color
- mask_bool = mask > 0.5
- overlay[mask_bool] = cv2.addWeighted(
- overlay[mask_bool], 0.5, colored[mask_bool], 0.5, 0
- )
-
- output = self._convert_rgb_to_format(overlay, fmt)
- if output is not None:
- success, map_info = buf.map(Gst.MapFlags.WRITE)
- if success:
- try:
- frame_bytes = np.ascontiguousarray(output).tobytes()
- dst = (ctypes.c_char * map_info.size).from_buffer(map_info.data)
- ctypes.memmove(
- dst, frame_bytes, min(len(frame_bytes), map_info.size)
- )
- finally:
- buf.unmap(map_info)
-
- # Append mask metadata as a custom buffer memory chunk
- if mask_info:
- meta_bytes = SAM_META_HEADER + json.dumps(mask_info).encode("utf-8")
- tmp = Gst.Buffer.new_allocate(None, len(meta_bytes), None)
- tmp.fill(0, meta_bytes)
- buf.append_memory(tmp.get_memory(0))
-
- @staticmethod
- def _convert_rgb_to_format(rgb, fmt):
- """Convert an RGB numpy array to the target GStreamer video format."""
- import cv2
- import numpy as np
-
- if fmt == "RGB":
- return rgb
- elif fmt == "BGR":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
- elif fmt == "RGBA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- elif fmt == "BGRA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- elif fmt == "ARGB":
- rgba = cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- return np.roll(rgba, 1, axis=-1)
- elif fmt == "ABGR":
- bgra = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- return np.roll(bgra, 1, axis=-1)
- else:
- return rgb
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(SamTransform)
- __gstelementfactory__ = ("pyml_sam", Gst.Rank.NONE, SamTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_sam", SamTransform)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_sam' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/sepformer.py b/plugins/python/sepformer.py
index c63b57b..ccb20a6 100644
--- a/plugins/python/sepformer.py
+++ b/plugins/python/sepformer.py
@@ -17,19 +17,14 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GObject", "2.0")
- from gi.repository import Gst, GObject, GstBase # noqa: E402
+ from backend import GObject
from base_separate import BaseSeparate
- import os
- from engine.pytorch_engine import PyTorchEngine
+ from engine.sepformer_engine import SepformerEngine
from engine.engine_factory import EngineFactory
except ImportError as e:
@@ -38,85 +33,12 @@
f"The 'pyml_sepformer' element will not be available. Error: {e}"
)
+if backend.BACKEND == "gst":
+ import gi
-class SepformerEngine(PyTorchEngine):
- def __init__(self):
- super().__init__()
- self.sample_rate = 0
-
- def do_load_model(self, model_name, **kwargs):
- from speechbrain.pretrained import SepformerSeparation
- from huggingface_hub import snapshot_download
-
- if not model_name:
- return
- self.logger.info(f"Loading Sepformer-WhamR model on device: {self.device}")
- savedir = "pretrained_models/sepformer-whamr"
- repo_id = "speechbrain/sepformer-whamr"
- try:
- # Download the model files manually to avoid deprecated argument issues
- if not os.path.exists(savedir):
- snapshot_download(repo_id=repo_id, local_dir=savedir)
- # Load from local directory
- self.model = SepformerSeparation.from_hparams(
- source=savedir, savedir=savedir, run_opts={"device": self.device}
- )
- self.sample_rate = 8000 # Hz, as per SpeechBrain Sepformer models
- self.sources = ["source0", "source1"] # 2 sources for separation
- except Exception as e:
- self.logger.error(f"Failed to load Sepformer-WhamR model: {e}")
-
- def separate_sources(
- self,
- mix,
- segment=10.0,
- overlap=0.1,
- ):
- import torch
- from torchaudio.transforms import Fade
-
- device = mix.device
- batch, length = mix.shape # For SpeechBrain, input is (batch, time)
- chunk_len = int(self.sample_rate * segment * (1 + overlap))
- start = 0
- end = chunk_len
- overlap_frames = int(overlap * self.sample_rate)
- fade = Fade(fade_in_len=0, fade_out_len=overlap_frames, fade_shape="linear")
-
- final = torch.zeros(batch, len(self.sources), length, device=device)
-
- min_chunk_samples = int(self.sample_rate * 0.5) # Avoid tiny chunks
-
- while start < length - overlap_frames:
- actual_end = min(end, length)
- chunk_length = actual_end - start
- if chunk_length < min_chunk_samples:
- break
-
- chunk = mix[:, start:actual_end]
- if chunk_length < chunk_len:
- pad = chunk_len - chunk_length
- chunk = torch.nn.functional.pad(chunk, (0, pad))
-
- # Add small epsilon noise to avoid zero std
- chunk += 1e-8 * torch.randn_like(chunk)
-
- with torch.no_grad():
- out = self.model.separate_batch(chunk) # (batch, time, sources)
- out = out.permute(0, 2, 1) # (batch, sources, time)
-
- out = out[:, :, :chunk_length]
- out = fade(out)
- final[:, :, start:actual_end] += out
- if start == 0:
- fade.fade_in_len = overlap_frames
- start += int(chunk_len - overlap_frames)
- else:
- start += chunk_len
- end += chunk_len
- if end >= length:
- fade.fade_out_len = 0
- return final
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ from gi.repository import Gst, GstBase # noqa: E402
class Sepformer(BaseSeparate):
@@ -129,32 +51,27 @@ class Sepformer(BaseSeparate):
SAMPLE_RATE = 8000
- CAPS = Gst.Caps(
- Gst.Structure(
- "audio/x-raw",
- format="S16LE",
- layout="interleaved",
- rate=SAMPLE_RATE,
- channels=1,
+ INPUT_CAPS = "audio/x-raw,format=S16LE,layout=interleaved,rate=8000,channels=1"
+ OUTPUT_CAPS = "audio/x-raw,format=S16LE,layout=interleaved,rate=8000,channels=1"
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new_with_gtype(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.REQUEST,
+ Gst.Caps.from_string(INPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
+ Gst.PadTemplate.new_with_gtype(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string(OUTPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
)
- )
-
- __gsttemplates__ = (
- Gst.PadTemplate.new_with_gtype(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.REQUEST,
- CAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- Gst.PadTemplate.new_with_gtype(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- CAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- )
def __init__(self):
super().__init__()
@@ -208,8 +125,7 @@ def do_separate(self, audio_data):
return selected.cpu().numpy()
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(Sepformer)
- __gstelementfactory__ = ("pyml_sepformer", Gst.Rank.NONE, Sepformer)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_sepformer", Sepformer)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning("pyml_sepformer not registered")
diff --git a/plugins/python/stablediffusion.py b/plugins/python/stablediffusion.py
index afc2f3a..9c1017b 100644
--- a/plugins/python/stablediffusion.py
+++ b/plugins/python/stablediffusion.py
@@ -17,6 +17,7 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
@@ -28,14 +29,16 @@
gi.require_version("GObject", "2.0")
from gi.repository import Gst, GstBase # noqa: E402
from base_aggregator import BaseAggregator
+ from diffusers import StableDiffusionPipeline # noqa: F401
except ImportError as e:
CAN_REGISTER_ELEMENT = False
GlobalLogger().warning(
f"The 'pyml_stablediffusion' element will not be available. Error: {e}"
)
-# Set output caps to image format (e.g., PNG)
-ICAPS = Gst.Caps(Gst.Structure("text/plain", format="utf8"))
+# Building a Gst object needs Gst.init, which only the gst backend calls.
+if backend.BACKEND == "gst":
+ ICAPS = Gst.Caps(Gst.Structure("text/plain", format="utf8"))
class StableDiffusion(BaseAggregator):
@@ -46,23 +49,24 @@ class StableDiffusion(BaseAggregator):
"Aaron Boxer ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new_with_gtype(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.REQUEST,
- ICAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- Gst.PadTemplate.new(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- Gst.Caps.from_string(
- "video/x-raw, width=512, height=512, format=RGBA, framerate=0/1"
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new_with_gtype(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.REQUEST,
+ ICAPS,
+ GstBase.AggregatorPad.__gtype__,
),
- ),
- )
+ Gst.PadTemplate.new(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string(
+ "video/x-raw, width=512, height=512, format=RGBA, framerate=0/1"
+ ),
+ ),
+ )
def do_load_model(self):
"""
@@ -153,14 +157,11 @@ def push_image_to_pipeline(self, image_data):
self.logger.error(f"Error pushing image to pipeline: {e}")
-# if CAN_REGISTER_ELEMENT:
-# GObject.type_register(StableDiffusion)
-# __gstelementfactory__ = (
-# "pyml_stablediffusion",
-# Gst.Rank.NONE,
-# StableDiffusion,
-# )
-# else:
-# GlobalLogger().warning(
-# "The 'pyml_stablediffusion' element will not be registered because required modules are missing."
-# )
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_stablediffusion", StableDiffusion
+ )
+elif not CAN_REGISTER_ELEMENT:
+ GlobalLogger().warning(
+ "The 'pyml_stablediffusion' element will not be registered because required modules are missing."
+ )
diff --git a/plugins/python/streamdemux.py b/plugins/python/streamdemux.py
index e781e07..ca6f76d 100644
--- a/plugins/python/streamdemux.py
+++ b/plugins/python/streamdemux.py
@@ -16,13 +16,13 @@
# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
# Boston, MA 02110-1301, USA.
+import backend
import gi
gi.require_version("Gst", "1.0")
gi.require_version("GstBase", "1.0")
-gi.require_version("GstAnalytics", "1.0")
-gi.require_version("GLib", "2.0")
-from gi.repository import Gst, GObject, GstAnalytics, GLib # noqa: E402
+from gi.repository import Gst, GObject # noqa: E402
+from backend import analytics # noqa: E402
from log.logger_factory import LoggerFactory # noqa: E402
from utils.metadata import Metadata # noqa: E402
from collections import defaultdict # noqa: E402
@@ -121,25 +121,27 @@ def process_src_pad(self, buffer, memory_chunk, stream_idx):
out_buffer.duration = buffer.duration
out_buffer.dts = buffer.dts
out_buffer.offset = buffer.offset
- meta = GstAnalytics.buffer_get_analytics_relation_meta(buffer)
+ meta = analytics.get_relation_meta(buffer)
if meta:
- out_meta = GstAnalytics.buffer_add_analytics_relation_meta(out_buffer)
- count = GstAnalytics.relation_get_length(meta)
+ out_meta = analytics.add_relation_meta(out_buffer)
+ objects = analytics.read_objects(meta)
self.logger.info(
- f"Processing {count} analytics relations for stream_{stream_idx}"
+ f"Processing {len(objects)} analytics relations for stream_{stream_idx}"
)
- for i in range(count):
- ret, od_mtd = meta.get_od_mtd(i)
- if ret and od_mtd:
- label_quark = od_mtd.get_obj_type()
- label = GLib.quark_to_string(label_quark)
- if f"stream_{stream_idx}_" in label:
- presence, x, y, w, h, conf = od_mtd.get_location()
- if presence:
- qk = GLib.quark_from_string(label)
- ret, new_od_mtd = out_meta.add_od_mtd(qk, x, y, w, h, conf)
- if not ret:
- self.logger.error(f"Failed to attach metadata: {label}")
+ for obj in objects:
+ label = obj["label"]
+ if f"stream_{stream_idx}_" in label:
+ new_od_mtd = analytics.add_object(
+ out_meta,
+ label,
+ obj["x"],
+ obj["y"],
+ obj["w"],
+ obj["h"],
+ obj["score"],
+ )
+ if new_od_mtd is None:
+ self.logger.error(f"Failed to attach metadata: {label}")
return out_buffer
def push_buffer(self, pad, buffer, pad_name):
@@ -229,5 +231,7 @@ def event(self, pad, parent, event):
return Gst.PadProbeReturn.OK
-GObject.type_register(StreamDemux)
-__gstelementfactory__ = ("pyml_streamdemux", Gst.Rank.NONE, StreamDemux)
+if backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_streamdemux", StreamDemux
+ )
diff --git a/plugins/python/streammux.py b/plugins/python/streammux.py
index 5547285..29dea89 100644
--- a/plugins/python/streammux.py
+++ b/plugins/python/streammux.py
@@ -16,6 +16,7 @@
# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
# Boston, MA 02110-1301, USA.
+import backend
import gi
gi.require_version("Gst", "1.0")
@@ -210,5 +211,5 @@ def do_get_property(self, prop):
raise AttributeError(f"Unknown property: {prop.name}")
-GObject.type_register(StreamMux)
-__gstelementfactory__ = ("pyml_streammux", Gst.Rank.NONE, StreamMux)
+if backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_streammux", StreamMux)
diff --git a/plugins/python/superres.py b/plugins/python/superres.py
index 7ffed5d..341a3d9 100644
--- a/plugins/python/superres.py
+++ b/plugins/python/superres.py
@@ -17,103 +17,23 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import ctypes
-
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- from gi.repository import Gst, GObject
-
from video_transform import VideoTransform
from utils.format_converter import FormatConverter
- from utils.muxed_buffer_processor import MuxedBufferProcessor
- from engine.pytorch_engine import PyTorchEngine
+ from engine.super_res_engine import SuperResEngine
from engine.engine_factory import EngineFactory
+ from backend import frameio, GObject
+ from tasks.superres import SuperResTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
GlobalLogger().warning(f"The 'superres' element will not be available. Error {e}")
-class SuperResEngine(PyTorchEngine):
- """
- PyTorch engine for image super-resolution using Real-ESRGAN.
-
- Supports model variants:
- real-esrgan-x4 (4x upscale, general purpose)
- real-esrgan-x2 (2x upscale)
- """
-
- def do_load_model(self, model_name, **kwargs):
- try:
- from basicsr.archs.rrdbnet_arch import RRDBNet
- from realesrgan import RealESRGANer
-
- scale = 4
- if "x2" in model_name:
- scale = 2
-
- model = RRDBNet(
- num_in_ch=3,
- num_out_ch=3,
- num_feat=64,
- num_block=23,
- num_grow_ch=32,
- scale=scale,
- )
-
- model_url = (
- "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth"
- if scale == 4
- else "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth"
- )
-
- gpu_id = 0 if str(self.device) != "cpu" else None
- self.upsampler = RealESRGANer(
- scale=scale,
- model_path=model_url,
- model=model,
- tile=0,
- tile_pad=10,
- pre_pad=0,
- half=False,
- gpu_id=gpu_id,
- )
- self._scale = scale
- self.logger.info(f"Real-ESRGAN model '{model_name}' (scale={scale}) loaded")
- except Exception as e:
- raise ValueError(f"Failed to load Real-ESRGAN model '{model_name}': {e}")
-
- def do_forward(self, frames):
- import cv2
- import numpy as np
-
- is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
- if not is_batch:
- frames = frames[np.newaxis]
-
- results = []
- for frame in frames:
- try:
- # Real-ESRGAN expects BGR input
- bgr = cv2.cvtColor(frame.astype(np.uint8), cv2.COLOR_RGB2BGR)
- output, _ = self.upsampler.enhance(bgr, outscale=self._scale)
- # Convert back to RGB
- rgb_out = cv2.cvtColor(output, cv2.COLOR_BGR2RGB)
- results.append(rgb_out)
- except Exception as e:
- self.logger.error(f"Super-resolution inference error: {e}")
- results.append(None)
-
- return results[0] if not is_batch else results
-
-
-class SuperResTransform(VideoTransform):
+class SuperResTransform(VideoTransform, SuperResTask):
"""
GStreamer element for image super-resolution using Real-ESRGAN.
@@ -156,82 +76,22 @@ def engine_name(self):
def engine_name(self, value):
raise ValueError("'engine_name' is read-only for pyml_superres")
- def do_transform_ip(self, buf):
- try:
- processor = MuxedBufferProcessor(
- self.logger, self.width, self.height, 30, 1
- )
- frames, _, num_sources, fmt = processor.extract_frames(buf, self.sinkpad)
- if frames is None:
- return Gst.FlowReturn.ERROR
-
- frame = frames[0] if frames.ndim == 4 else frames
- upscaled = self._do_forward(frame)
- if upscaled is None:
- return Gst.FlowReturn.OK
-
- self._apply_superres(buf, upscaled, fmt)
- return Gst.FlowReturn.OK
-
- except Exception as e:
- self.logger.error(f"Super-resolution transform error: {e}")
- return Gst.FlowReturn.ERROR
-
- def _do_forward(self, frame):
- if self.engine:
- return self.engine.do_forward(frame)
- return None
-
- def _apply_superres(self, buf, upscaled, fmt):
- """Resize upscaled frame back to original dimensions and write to buffer."""
- import cv2
- import numpy as np
-
- # Resize back to original buffer dimensions for in-place compatibility
- resized = cv2.resize(
- upscaled, (self.width, self.height), interpolation=cv2.INTER_LANCZOS4
- )
- output = self._convert_rgb_to_format(resized, fmt)
- if output is not None:
- success, map_info = buf.map(Gst.MapFlags.WRITE)
- if success:
- try:
- frame_bytes = np.ascontiguousarray(output).tobytes()
- dst = (ctypes.c_char * map_info.size).from_buffer(map_info.data)
- ctypes.memmove(
- dst, frame_bytes, min(len(frame_bytes), map_info.size)
- )
- finally:
- buf.unmap(map_info)
-
- @staticmethod
- def _convert_rgb_to_format(rgb, fmt):
- """Convert an RGB numpy array to the target GStreamer video format."""
- import cv2
- import numpy as np
-
- if fmt == "RGB":
- return rgb
- elif fmt == "BGR":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
- elif fmt == "RGBA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- elif fmt == "BGRA":
- return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- elif fmt == "ARGB":
- rgba = cv2.cvtColor(rgb, cv2.COLOR_RGB2RGBA)
- return np.roll(rgba, 1, axis=-1)
- elif fmt == "ABGR":
- bgra = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGRA)
- return np.roll(bgra, 1, axis=-1)
- else:
- return rgb
-
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(SuperResTransform)
- __gstelementfactory__ = ("pyml_superres", Gst.Rank.NONE, SuperResTransform)
-else:
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Upscale the primary frame and write it back at the original size."""
+ frame = frames[0] if frames.ndim == 4 else frames
+ upscaled = self.forward(frame)
+ if upscaled is None:
+ return
+
+ output, _blob = self.decode(upscaled, fmt)
+ frameio.write_result(target, output)
+
+
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_superres", SuperResTransform
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_superres' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/tasks/__init__.py b/plugins/python/tasks/__init__.py
new file mode 100644
index 0000000..ddac92b
--- /dev/null
+++ b/plugins/python/tasks/__init__.py
@@ -0,0 +1,16 @@
+# Portable ML task logic
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+
+"""Backend-agnostic ML task logic.
+
+Each module here holds the inference and result-handling steps for one kind of
+element (object detection, ...), expressed only in terms of the ML engine,
+numpy results, and the analytics facade (`backend.analytics`). A backend
+combines a task mixin with a framework element shell, which supplies the
+per-buffer glue (do_transform_ip / process, pad templates, buffer I/O).
+"""
diff --git a/plugins/python/tasks/action.py b/plugins/python/tasks/action.py
new file mode 100644
index 0000000..ed1a196
--- /dev/null
+++ b/plugins/python/tasks/action.py
@@ -0,0 +1,79 @@
+# ActionTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic action-recognition task.
+
+`ActionTask` runs classification over a window of frames and turns the result
+into an optional output frame (a label overlay) plus a metadata dict, using
+only numpy/cv2 and the engine. It never touches the buffer and holds no
+temporal state; the backend element shell accumulates the frame window, does
+the frame read, the frame write, and the metadata attach through the frameio
+facade.
+
+Contract expected from the host element:
+ self.engine - the ML engine (provided by MLEngineMixin)
+ self.draw_label - bool overlay flag (a backend-declared property)
+"""
+
+from tasks.frame_format import to_format
+
+
+class ActionTask:
+ """Inference + frame/metadata production, independent of any framework."""
+
+ def forward(self, frame_buffer):
+ if self.engine:
+ return self.engine.do_forward(frame_buffer)
+ return None
+
+ def decode(self, frame, result, fmt):
+ """Turn a classification result into ``(output_frame_or_None, blob_bytes)``.
+
+ ``output_frame`` is the label-overlaid frame in pixel format ``fmt``
+ (or None when no overlay is drawn). ``blob_bytes`` is the serialized
+ action metadata, always returned, to be appended by the shell.
+ """
+ import json
+
+ import cv2
+
+ label = result.get("label", "")
+ score = result.get("score", 0.0)
+ output = None
+
+ # Draw label before appending read-only metadata memory
+ if self.draw_label and label:
+ overlay = frame.copy()
+ text = f"{label} ({score:.2f})"
+ (tw, th), _ = cv2.getTextSize(text, cv2.FONT_HERSHEY_SIMPLEX, 1.0, 2)
+ cv2.rectangle(overlay, (8, 8), (16 + tw, 16 + th + 8), (0, 0, 0), -1)
+ cv2.putText(
+ overlay,
+ text,
+ (12, 12 + th),
+ cv2.FONT_HERSHEY_SIMPLEX,
+ 1.0,
+ (0, 255, 0),
+ 2,
+ cv2.LINE_AA,
+ )
+
+ output = to_format(overlay, fmt)
+
+ # Serialize action metadata for the shell to append.
+ return output, json.dumps(result).encode("utf-8")
diff --git a/plugins/python/tasks/anomaly.py b/plugins/python/tasks/anomaly.py
new file mode 100644
index 0000000..8b619eb
--- /dev/null
+++ b/plugins/python/tasks/anomaly.py
@@ -0,0 +1,89 @@
+# AnomalyTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic anomaly-detection task.
+
+`AnomalyTask` runs inference and turns the result into an optional output frame
+(a heatmap overlay) plus a metadata dict, using only numpy/cv2 and the engine.
+It never touches the buffer; the backend element shell does the frame read,
+the frame write, and the metadata attach through the frameio facade.
+
+Contract expected from the host element:
+ self.engine - the ML engine (provided by MLEngineMixin)
+ self.threshold - anomaly threshold (a backend-declared property)
+ self.draw_heatmap - bool overlay flag (a backend-declared property)
+"""
+
+from tasks.frame_format import to_format
+
+
+class AnomalyTask:
+ """Inference + frame/metadata production, independent of any framework."""
+
+ def forward(self, frame):
+ if self.engine:
+ return self.engine.do_forward(frame, threshold=self.threshold)
+ return None
+
+ def decode(self, frame, result, fmt):
+ """Turn an inference result into ``(output_frame_or_None, blob_bytes)``.
+
+ ``output_frame`` is the heatmap-overlaid frame in pixel format ``fmt``
+ (or None when no overlay is drawn). ``blob_bytes`` is the serialized
+ anomaly metadata, always returned, to be appended by the shell.
+ """
+ import json
+
+ import cv2
+ import numpy as np
+
+ is_anomaly = result.get("is_anomaly", False)
+ heatmap = result.get("heatmap")
+ output = None
+
+ # Draw heatmap overlay when the frame is flagged.
+ if self.draw_heatmap and is_anomaly and heatmap is not None:
+ H, W = frame.shape[:2]
+ heatmap_resized = cv2.resize(heatmap, (W, H))
+ heatmap_uint8 = (heatmap_resized * 255).astype(np.uint8)
+ heatmap_color = cv2.applyColorMap(heatmap_uint8, cv2.COLORMAP_JET)
+ heatmap_rgb = cv2.cvtColor(heatmap_color, cv2.COLOR_BGR2RGB)
+
+ overlay = cv2.addWeighted(frame, 0.6, heatmap_rgb, 0.4, 0)
+
+ # Draw anomaly score text
+ score = result.get("score", 0.0)
+ text = f"ANOMALY: {score:.3f}"
+ cv2.putText(
+ overlay,
+ text,
+ (12, 36),
+ cv2.FONT_HERSHEY_SIMPLEX,
+ 1.0,
+ (255, 0, 0),
+ 2,
+ cv2.LINE_AA,
+ )
+
+ output = to_format(overlay, fmt)
+
+ meta = {
+ "score": result.get("score", 0.0),
+ "is_anomaly": is_anomaly,
+ }
+ return output, json.dumps(meta).encode("utf-8")
diff --git a/plugins/python/tasks/clip.py b/plugins/python/tasks/clip.py
new file mode 100644
index 0000000..7551794
--- /dev/null
+++ b/plugins/python/tasks/clip.py
@@ -0,0 +1,68 @@
+# ClipTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic CLIP classification task.
+
+`ClipTask` turns zero-shot classification results into GstAnalytics-style
+metadata via the analytics facade. It never touches the buffer; the backend
+element shell does the frame read, runs inference on a background thread, and
+passes the opaque metadata target to `decode`.
+
+Contract expected from the host element:
+ self.logger - logger (provided by MLEngineMixin)
+ self.top_k - number of top labels to attach (a backend-declared property)
+ self.threshold - minimum probability to include (a backend-declared property)
+ self.width - frame width (host geometry)
+ self.height - frame height (host geometry)
+
+`decode`'s `target` argument is the opaque metadata target (a Gst buffer on the
+gst backend); it is only passed through to the analytics facade.
+"""
+
+from backend import analytics
+
+
+class ClipTask:
+ """Result-to-metadata step, independent of any framework."""
+
+ def decode(self, target, results):
+ """Attach top-k classification results above threshold as GstAnalytics metadata."""
+ meta = analytics.add_relation_meta(target)
+ if not meta:
+ self.logger.error("Failed to add analytics relation metadata")
+ return
+
+ attached = 0
+ for label, prob in results:
+ if attached >= self.top_k:
+ break
+ if prob < self.threshold:
+ break
+
+ mtd = analytics.add_object(
+ meta,
+ f"clip_{label.replace(' ', '_')}",
+ 0,
+ 0,
+ self.width,
+ self.height,
+ prob,
+ )
+ if mtd is not None:
+ attached += 1
+ self.logger.info(f"CLIP: {label} = {prob:.3f}")
diff --git a/plugins/python/tasks/depth.py b/plugins/python/tasks/depth.py
new file mode 100644
index 0000000..b2627f6
--- /dev/null
+++ b/plugins/python/tasks/depth.py
@@ -0,0 +1,103 @@
+# DepthTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic monocular-depth task.
+
+`DepthTask` runs inference and turns a depth map into an optional output frame
+(a colorized depth visualization) plus a serialized metadata blob (the uint8
+normalized depth map), using only numpy/cv2 and the engine. It never touches
+the buffer; the backend element shell does the frame read, the frame write, and
+the metadata attach through the frameio facade.
+
+Contract expected from the host element:
+ self.engine - the ML engine (provided by MLEngineMixin)
+ self.logger - logger
+ self.visualize - bool, replace frame with colorized depth (a property)
+ self.colormap - str colormap name (a property)
+"""
+
+# cv2 colormap IDs for depth visualization
+COLORMAP_IDS = {
+ "inferno": 9,
+ "jet": 2,
+ "viridis": 16,
+ "plasma": 18,
+ "magma": 13,
+}
+
+
+class DepthTask:
+ """Inference + frame/metadata production, independent of any framework."""
+
+ def forward(self, frames):
+ if self.engine:
+ return self.engine.do_forward(frames)
+ return None
+
+ def decode(self, frame, depth_map, fmt):
+ """Normalize depth, optionally visualize, then build metadata.
+
+ Returns ``(output_frame_or_None, blob_bytes)``. ``output_frame`` is the
+ colorized depth visualization in pixel format ``fmt`` (or None when not
+ visualizing). ``blob_bytes`` is the uint8 normalized depth map, always
+ returned, to be appended by the shell.
+ """
+ import cv2
+ import numpy as np
+
+ d_min, d_max = depth_map.min(), depth_map.max()
+ if d_max > d_min:
+ depth_norm = ((depth_map - d_min) / (d_max - d_min) * 255).astype(np.uint8)
+ else:
+ depth_norm = np.zeros_like(depth_map, dtype=np.uint8)
+
+ output = None
+
+ # Visualize first, before appending any read-only metadata memory.
+ # (A READONLY chunk on the buffer would prevent buf.map(WRITE) from succeeding.)
+ if self.visualize:
+ cmap_id = COLORMAP_IDS.get(self.colormap, COLORMAP_IDS["inferno"])
+ depth_bgr = cv2.applyColorMap(depth_norm, cmap_id)
+ output = self._convert_bgr_to_format(depth_bgr, fmt)
+
+ # Build uint8 depth map metadata bytes (payload only; header added by shell).
+ depth_bytes = depth_norm.tobytes()
+ return output, depth_bytes
+
+ @staticmethod
+ def _convert_bgr_to_format(bgr, fmt):
+ """Convert a BGR numpy array to the target GStreamer video format."""
+ import cv2
+ import numpy as np
+
+ if fmt == "RGB":
+ return cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
+ elif fmt == "BGR":
+ return bgr
+ elif fmt == "RGBA":
+ return cv2.cvtColor(bgr, cv2.COLOR_BGR2RGBA)
+ elif fmt == "BGRA":
+ return cv2.cvtColor(bgr, cv2.COLOR_BGR2BGRA)
+ elif fmt == "ARGB":
+ rgba = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGBA)
+ return np.roll(rgba, 1, axis=-1) # RGBA -> ARGB
+ elif fmt == "ABGR":
+ bgra = cv2.cvtColor(bgr, cv2.COLOR_BGR2BGRA)
+ return np.roll(bgra, 1, axis=-1) # BGRA -> ABGR
+ else:
+ return cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
diff --git a/plugins/python/tasks/embedding.py b/plugins/python/tasks/embedding.py
new file mode 100644
index 0000000..ac81e95
--- /dev/null
+++ b/plugins/python/tasks/embedding.py
@@ -0,0 +1,67 @@
+# EmbeddingTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic embedding task.
+
+`EmbeddingTask` runs the embedding forward pass and serializes the result into
+the length-prefixed payload that follows the GST-EMBEDDING: header. It uses only
+numpy/json/struct and the engine; it never touches the buffer. The backend
+element shell does the frame read and appends the payload via the frameio
+facade, prepending the header.
+
+Contract expected from the host element:
+ self.engine - the ML engine (provided by MLEngineMixin)
+ self.normalize - L2-normalize flag (a backend-declared property)
+ self.text - optional text query (a backend-declared property)
+ self._text_embedding - cached text embedding (host-managed state)
+"""
+
+import json
+import struct
+
+
+class EmbeddingTask:
+ """Inference + payload serialization, independent of any framework."""
+
+ def forward(self, frame):
+ return self.engine.do_forward(frame, normalize=self.normalize)
+
+ def decode(self, emb):
+ """Serialize an embedding into ``(None, payload_bytes)``.
+
+ ``payload_bytes`` is everything AFTER the GST-EMBEDDING: header: the
+ 4-byte JSON header length, the JSON header, and the raw float32 bytes.
+ The shell prepends the header when appending the blob.
+ """
+ import numpy as np
+
+ # Build JSON header with dimension info and optional similarity score
+ header = {"dim": int(emb.shape[0]), "dtype": "float32"}
+ if self._text_embedding is not None:
+ similarity = float(np.dot(emb, self._text_embedding))
+ header["text"] = self.text
+ header["similarity"] = round(similarity, 6)
+
+ header_bytes = json.dumps(header).encode("utf-8")
+ header_len = struct.pack("= 6
+ ): # [x1, y1, x2, y2, score, label]
+ self.logger.info(f"Stream {stream_idx} - Processing list of detections")
+ boxes = [[det[0], det[1], det[2], det[3]] for det in output]
+ scores = [det[4] for det in output]
+ labels = [int(det[5]) for det in output]
+ else:
+ self.logger.error(
+ f"Stream {stream_idx} - Unrecognized format: {output} (type: {type(output)})"
+ )
+ return
+
+ meta = analytics.add_relation_meta(buf)
+ if not meta:
+ self.logger.error(
+ f"Stream {stream_idx} - Failed to add analytics relation metadata"
+ )
+ return
+
+ self.logger.info(f"Stream {stream_idx} - Adding {len(boxes)} detections")
+ for i, (box, label, score) in enumerate(zip(boxes, labels, scores)):
+ x1, y1, x2, y2 = box
+ qk_string = f"stream_{stream_idx}_label_{label}"
+ od_mtd = analytics.add_object(
+ meta, qk_string, x1, y1, x2 - x1, y2 - y1, score
+ )
+ if od_mtd is None:
+ self.logger.error(
+ f"Stream {stream_idx} - Failed to add od_mtd for detection {i}"
+ )
+ continue
+ self.logger.info(
+ f"Stream {stream_idx} - Added detection {i}: label={qk_string}, x1={x1}, y1={y1}, w={x2-x1}, h={y2-y1}, score={score}"
+ )
+
+ attached_meta = analytics.get_relation_meta(buf)
+ if attached_meta:
+ count = analytics.relation_length(attached_meta)
+ self.logger.info(
+ f"Stream {stream_idx} - Metadata relations after adding: {count}"
+ )
diff --git a/plugins/python/tasks/ocr.py b/plugins/python/tasks/ocr.py
new file mode 100644
index 0000000..b44707a
--- /dev/null
+++ b/plugins/python/tasks/ocr.py
@@ -0,0 +1,84 @@
+# OcrTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic OCR task.
+
+`OcrTask` runs text recognition and turns the result into an optional output
+frame (recognized text drawn on the frame) plus a serialized metadata blob
+(JSON with recognized text and regions), using only numpy/cv2 and the engine.
+It never touches the buffer; the backend element shell does the frame read, the
+frame write, and the metadata attach through the frameio facade.
+
+Contract expected from the host element:
+ self.engine - the ML engine (provided by MLEngineMixin)
+ self.logger - logger
+ self.draw_text - bool, draw recognized text on the frame (a property)
+"""
+
+from tasks.frame_format import to_format
+
+
+class OcrTask:
+ """Inference + frame/metadata production, independent of any framework."""
+
+ def forward(self, frames):
+ if self.engine:
+ return self.engine.do_forward(frames)
+ return None
+
+ def decode(self, frame, result, fmt):
+ """Draw recognized text on frame and build OCR metadata.
+
+ Returns ``(output_frame_or_None, blob_bytes_or_None)``. ``output_frame``
+ is the text-overlaid frame in pixel format ``fmt`` (or None when not
+ drawing). ``blob_bytes`` is the serialized OCR metadata (or None when
+ there are no regions), to be appended by the shell.
+ """
+ import json
+
+ import cv2
+
+ regions = result.get("regions", [])
+ output = None
+
+ # Draw text overlays before appending read-only metadata memory
+ if self.draw_text and regions:
+ overlay = frame.copy()
+ for region in regions:
+ x, y, w, h = region["x"], region["y"], region["w"], region["h"]
+ text = region["text"]
+ cv2.rectangle(overlay, (x, y), (x + w, y + h), (0, 255, 0), 2)
+ font_scale = max(0.4, min(w / 300.0, 1.0))
+ cv2.putText(
+ overlay,
+ text,
+ (x + 4, y + h - 8),
+ cv2.FONT_HERSHEY_SIMPLEX,
+ font_scale,
+ (0, 255, 0),
+ 1,
+ cv2.LINE_AA,
+ )
+
+ output = to_format(overlay, fmt)
+
+ # Build OCR results bytes (payload only; header added by shell).
+ blob = None
+ if regions:
+ blob = json.dumps(regions).encode("utf-8")
+ return output, blob
diff --git a/plugins/python/tasks/optical_flow.py b/plugins/python/tasks/optical_flow.py
new file mode 100644
index 0000000..9cfc119
--- /dev/null
+++ b/plugins/python/tasks/optical_flow.py
@@ -0,0 +1,70 @@
+# OpticalFlowTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic optical-flow task.
+
+`OpticalFlowTask` runs inference on a pair of frames and renders the resulting
+flow field as a color overlay, using only numpy/cv2 and the engine. It never
+touches the buffer; the backend element shell does the frame read, the frame
+write, and the temporal frame-pairing (the previous-frame state) itself.
+
+Contract expected from the host element:
+ self.engine - the ML engine (provided by MLEngineMixin)
+"""
+
+from tasks.frame_format import to_format
+
+
+class OpticalFlowTask:
+ """Inference + flow visualization, independent of any framework."""
+
+ def forward(self, prev_frame, curr_frame):
+ if self.engine:
+ return self.engine.do_forward(prev_frame, curr_frame)
+ return None
+
+ def decode(self, flow, frame, fmt):
+ """Render flow as a color overlay blended onto ``frame``.
+
+ Returns ``(output_frame_or_None, None)``: the blended frame in pixel
+ format ``fmt`` for the shell to write, and no metadata blob.
+ """
+ import cv2
+
+ flow_vis = self._flow_to_color(flow)
+ blended = cv2.addWeighted(frame, 0.5, flow_vis, 0.5, 0)
+ output = to_format(blended, fmt)
+ return output, None
+
+ @staticmethod
+ def _flow_to_color(flow):
+ """Convert optical flow (H, W, 2) to an RGB color image using HSV encoding."""
+ import cv2
+ import numpy as np
+
+ fx, fy = flow[..., 0], flow[..., 1]
+ mag = np.sqrt(fx**2 + fy**2)
+ ang = np.arctan2(fy, fx)
+
+ hsv = np.zeros((*flow.shape[:2], 3), dtype=np.uint8)
+ hsv[..., 0] = ((ang + np.pi) / (2 * np.pi) * 179).astype(np.uint8)
+ hsv[..., 1] = 255
+ mag_norm = mag / (mag.max() + 1e-8)
+ hsv[..., 2] = (mag_norm * 255).astype(np.uint8)
+
+ return cv2.cvtColor(hsv, cv2.COLOR_HSV2RGB)
diff --git a/plugins/python/tasks/sam.py b/plugins/python/tasks/sam.py
new file mode 100644
index 0000000..db77fb7
--- /dev/null
+++ b/plugins/python/tasks/sam.py
@@ -0,0 +1,98 @@
+# SamTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic segmentation task.
+
+`SamTask` runs Segment Anything inference and turns the result into an optional
+output frame (colored mask overlays) plus a serialized metadata blob (JSON with
+mask scores and shapes), using only numpy/cv2 and the engine. It never touches
+the buffer; the backend element shell does the frame read, the frame write, and
+the metadata attach through the frameio facade.
+
+Contract expected from the host element:
+ self.engine - the ML engine (provided by MLEngineMixin)
+ self.logger - logger
+ self.visualize - bool, overlay colored masks on the frame (a property)
+ self.max_masks - int, maximum number of masks (a property)
+"""
+
+from tasks.frame_format import to_format
+
+# Mask overlay colors (BGR) cycled across detected objects
+MASK_COLORS = [
+ (255, 0, 0),
+ (0, 255, 0),
+ (0, 0, 255),
+ (255, 255, 0),
+ (0, 255, 255),
+ (255, 0, 255),
+ (128, 0, 255),
+ (255, 128, 0),
+ (0, 128, 255),
+ (128, 255, 0),
+]
+
+
+class SamTask:
+ """Inference + frame/metadata production, independent of any framework."""
+
+ def forward(self, frames):
+ if self.engine:
+ return self.engine.do_forward(frames, max_masks=self.max_masks)
+ return None
+
+ def decode(self, frame, result, fmt):
+ """Overlay masks on frame and build mask metadata.
+
+ Returns ``(output_frame_or_None, blob_bytes_or_None)``. ``output_frame``
+ is the mask-overlaid frame in pixel format ``fmt`` (or None when not
+ visualizing). ``blob_bytes`` is the serialized mask metadata (or None
+ when there is none), to be appended by the shell.
+ """
+ import json
+
+ import cv2
+ import numpy as np
+
+ raw_masks = result.get("raw_masks")
+ mask_info = result.get("masks", [])
+ output = None
+
+ # Draw mask overlays before appending read-only metadata memory
+ if self.visualize and raw_masks is not None:
+ overlay = frame.copy()
+ for j in range(min(raw_masks.shape[0], self.max_masks)):
+ best_idx = 0
+ if raw_masks.ndim == 4 and raw_masks.shape[1] > 1:
+ best_idx = raw_masks[j].sum(axis=(1, 2)).argmax()
+ mask = raw_masks[j, best_idx]
+ color = MASK_COLORS[j % len(MASK_COLORS)]
+ colored = np.zeros_like(overlay)
+ colored[:] = color
+ mask_bool = mask > 0.5
+ overlay[mask_bool] = cv2.addWeighted(
+ overlay[mask_bool], 0.5, colored[mask_bool], 0.5, 0
+ )
+
+ output = to_format(overlay, fmt)
+
+ # Build mask metadata bytes (payload only; header added by shell).
+ blob = None
+ if mask_info:
+ blob = json.dumps(mask_info).encode("utf-8")
+ return output, blob
diff --git a/plugins/python/tasks/superres.py b/plugins/python/tasks/superres.py
new file mode 100644
index 0000000..a15fe9a
--- /dev/null
+++ b/plugins/python/tasks/superres.py
@@ -0,0 +1,56 @@
+# SuperResTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic super-resolution task.
+
+`SuperResTask` runs inference and resizes the upscaled output back to the
+original frame dimensions, using only numpy/cv2 and the engine. It never
+touches the buffer; the backend element shell does the frame read and the
+frame write through the frameio facade.
+
+Contract expected from the host element:
+ self.engine - the ML engine (provided by MLEngineMixin)
+ self.width - original buffer width (a backend-declared property)
+ self.height - original buffer height (a backend-declared property)
+"""
+
+from tasks.frame_format import to_format
+
+
+class SuperResTask:
+ """Inference + frame production, independent of any framework."""
+
+ def forward(self, frame):
+ if self.engine:
+ return self.engine.do_forward(frame)
+ return None
+
+ def decode(self, upscaled, fmt):
+ """Resize the upscaled frame back to original dimensions.
+
+ Returns ``(output_frame_or_None, None)``: the resized frame in pixel
+ format ``fmt`` for the shell to write, and no metadata blob.
+ """
+ import cv2
+
+ # Resize back to original buffer dimensions for in-place compatibility
+ resized = cv2.resize(
+ upscaled, (self.width, self.height), interpolation=cv2.INTER_LANCZOS4
+ )
+ output = to_format(resized, fmt)
+ return output, None
diff --git a/plugins/python/tasks/vlm.py b/plugins/python/tasks/vlm.py
new file mode 100644
index 0000000..e0f087f
--- /dev/null
+++ b/plugins/python/tasks/vlm.py
@@ -0,0 +1,57 @@
+# VlmTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic Vision-Language Model task.
+
+`VlmTask` runs the VLM forward pass and serializes the generated text into the
+payload that follows the GST-VLM: header. It uses only json and the engine; it
+never touches the buffer. The backend element shell does the frame read and
+appends the payload via the frameio facade, prepending the header.
+
+Contract expected from the host element:
+ self.engine - the ML engine (provided by MLEngineMixin)
+ self.prompt - user prompt (a backend-declared property)
+ self.system_prompt - optional system prompt (a backend-declared property)
+ self.max_tokens - max tokens to generate (a backend-declared property)
+ self.temperature - sampling temperature (a backend-declared property)
+"""
+
+import json
+
+
+class VlmTask:
+ """Inference + payload serialization, independent of any framework."""
+
+ def forward(self, frame):
+ return self.engine.do_forward(
+ frame,
+ prompt=self.prompt,
+ system_prompt=self.system_prompt,
+ max_tokens=self.max_tokens,
+ temperature=self.temperature,
+ )
+
+ def decode(self, text):
+ """Serialize a generated caption into ``(None, payload_bytes)``.
+
+ ``payload_bytes`` is the JSON result that follows the GST-VLM: header;
+ the shell prepends the header when appending the blob.
+ """
+ result = {"text": text}
+ payload = json.dumps(result).encode("utf-8")
+ return None, payload
diff --git a/plugins/python/tasks/yolo.py b/plugins/python/tasks/yolo.py
new file mode 100644
index 0000000..76d6fc8
--- /dev/null
+++ b/plugins/python/tasks/yolo.py
@@ -0,0 +1,220 @@
+# YoloTask
+# Copyright (C) 2024-2026 Collabora Ltd.
+#
+# This library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Library General Public
+# License as published by the Free Software Foundation; either
+# version 2 of the License, or (at your option) any later version.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Library General Public License for more details.
+#
+# You should have received a copy of the GNU Library General Public
+# License along with this library; if not, write to the
+# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+# Boston, MA 02110-1301, USA.
+
+"""Backend-agnostic YOLO task.
+
+`YoloTask` overrides `do_decode` with the YOLO-specific result handling
+(per-instance class names, tracking ids, segmentation) expressed only through
+the analytics facade and the engine result. It inherits `do_forward` from
+`ObjectDetectorTask`. A backend element shell combines this mixin with its
+framework element base (the gst backend uses `BaseObjectDetector`); the shell
+supplies the engine wiring, the read-only `engine_name` property, and element
+registration.
+
+See `ObjectDetectorTask` for the host-element contract (self.engine, self.logger,
+self.track). `do_decode`'s `buf` is the opaque metadata target.
+"""
+
+from backend import analytics
+from tasks.object_detector import ObjectDetectorTask
+
+COCO_CLASSES = {
+ 0: "person",
+ 1: "bicycle",
+ 2: "car",
+ 3: "motorcycle",
+ 4: "airplane",
+ 5: "bus",
+ 6: "train",
+ 7: "truck",
+ 8: "boat",
+ 9: "traffic light",
+ 10: "fire hydrant",
+ 11: "stop sign",
+ 12: "parking meter",
+ 13: "bench",
+ 14: "bird",
+ 15: "cat",
+ 16: "dog",
+ 17: "horse",
+ 18: "sheep",
+ 19: "cow",
+ 20: "elephant",
+ 21: "bear",
+ 22: "zebra",
+ 23: "giraffe",
+ 24: "backpack",
+ 25: "umbrella",
+ 26: "handbag",
+ 27: "tie",
+ 28: "suitcase",
+ 29: "frisbee",
+ 30: "skis",
+ 31: "snowboard",
+ 32: "sports ball",
+ 33: "kite",
+ 34: "baseball bat",
+ 35: "baseball glove",
+ 36: "skateboard",
+ 37: "surfboard",
+ 38: "tennis racket",
+ 39: "bottle",
+ 40: "wine glass",
+ 41: "cup",
+ 42: "fork",
+ 43: "knife",
+ 44: "spoon",
+ 45: "bowl",
+ 46: "banana",
+ 47: "apple",
+ 48: "sandwich",
+ 49: "orange",
+ 50: "broccoli",
+ 51: "carrot",
+ 52: "hot dog",
+ 53: "pizza",
+ 54: "donut",
+ 55: "cake",
+ 56: "chair",
+ 57: "couch",
+ 58: "potted plant",
+ 59: "bed",
+ 60: "dining table",
+ 61: "toilet",
+ 62: "TV",
+ 63: "laptop",
+ 64: "mouse",
+ 65: "remote",
+ 66: "keyboard",
+ 67: "cell phone",
+ 68: "microwave",
+ 69: "oven",
+ 70: "toaster",
+ 71: "sink",
+ 72: "refrigerator",
+ 73: "book",
+ 74: "clock",
+ 75: "vase",
+ 76: "scissors",
+ 77: "teddy bear",
+ 78: "hair drier",
+ 79: "toothbrush",
+}
+
+
+class YoloTask(ObjectDetectorTask):
+ """YOLO detection + tracking + segmentation, independent of any framework."""
+
+ def do_decode(self, buf, result, stream_idx=0):
+ self.logger.debug(
+ f"Decoding YOLO result for buffer {hex(id(buf))}, stream {stream_idx}: {result}"
+ )
+ boxes = result.boxes
+ masks = None
+ if not self.engine.track:
+ masks = result.masks
+
+ if boxes is None or len(boxes) == 0:
+ self.logger.info("No detections found.")
+ return
+
+ meta = analytics.add_relation_meta(buf)
+ if not meta:
+ self.logger.error(
+ f"Stream {stream_idx} - Failed to add analytics relation metadata"
+ )
+ return
+
+ self.logger.debug(
+ f"Stream {stream_idx} - Attaching metadata for {len(boxes)} detections"
+ )
+ for i in range(len(boxes)):
+ x1, y1, x2, y2 = boxes.xyxy[i]
+ score = boxes.conf[i]
+ label = boxes.cls[i]
+ label_num = label.item()
+ # Prefer the model's own class names; fall back to COCO for plain yolo.
+ names = getattr(result, "names", None) or COCO_CLASSES
+ class_name = names.get(label_num, f"unknown_{label_num}")
+
+ # Use class name for detection, track_id for tracking
+ if self.engine.track and hasattr(boxes, "id") and boxes.id is not None:
+ track_id = boxes.id[i]
+ track_id_int = int(track_id.item())
+ qk_string = f"stream_{stream_idx}_id_{track_id_int}"
+ else:
+ qk_string = (
+ f"stream_{stream_idx}_{class_name}" # No index, just class name
+ )
+
+ od_mtd = analytics.add_object(
+ meta,
+ qk_string,
+ x1.item(),
+ y1.item(),
+ x2.item() - x1.item(),
+ y2.item() - y1.item(),
+ score.item(),
+ )
+ if od_mtd is None:
+ self.logger.error(
+ f"Stream {stream_idx} - Failed to add object detection metadata"
+ )
+ continue
+ self.logger.debug(
+ f"Stream {stream_idx} - Added od_mtd: label={qk_string}, x1={x1.item()}, y1={y1.item()}, w={x2.item()-x1.item()}, h={y2.item()-y1.item()}, score={score.item()}"
+ )
+
+ # Tracking metadata only when track=True
+ if self.engine.track and hasattr(boxes, "id") and boxes.id is not None:
+ tracking_mtd = analytics.add_tracking(meta, track_id_int)
+ if tracking_mtd is None:
+ self.logger.error(
+ f"Stream {stream_idx} - Failed to add tracking metadata"
+ )
+ continue
+ ret = analytics.relate(meta, od_mtd, tracking_mtd)
+ if not ret:
+ self.logger.error(
+ f"Stream {stream_idx} - Failed to relate object detection and tracking metadata"
+ )
+ else:
+ self.logger.debug(
+ f"Stream {stream_idx} - Linked od_mtd {od_mtd} to tracking_mtd {tracking_mtd}"
+ )
+
+ if masks is not None:
+ self.add_segmentation_metadata(buf, masks[i], x1, y1, x2, y2)
+
+ attached_meta = analytics.get_relation_meta(buf)
+ if attached_meta:
+ count = analytics.relation_length(attached_meta)
+ self.logger.info(
+ f"Stream {stream_idx} - Metadata attached to buffer {hex(id(buf))}: {count} relations"
+ )
+ else:
+ self.logger.error(
+ f"Stream {stream_idx} - Metadata not attached to buffer after adding"
+ )
+
+ def add_segmentation_metadata(self, buf, mask, x1, y1, x2, y2):
+ """
+ Adds segmentation mask metadata to the buffer.
+ """
+ self.logger.info("Adding segmentation mask metadata")
+ pass
diff --git a/plugins/python/tracker.py b/plugins/python/tracker.py
index ad0db4d..3dba4df 100644
--- a/plugins/python/tracker.py
+++ b/plugins/python/tracker.py
@@ -17,6 +17,7 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
@@ -25,14 +26,15 @@
gi.require_version("Gst", "1.0")
gi.require_version("GstBase", "1.0")
gi.require_version("GstVideo", "1.0")
- gi.require_version("GstAnalytics", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GstBase, GstAnalytics, GObject, GLib # noqa: E402
+ from gi.repository import Gst, GstBase # noqa: E402
from log.logger_factory import LoggerFactory # noqa: E402
+ from backend import analytics, GObject # noqa: E402
- VIDEO_SRC_CAPS = Gst.Caps.from_string("video/x-raw")
- VIDEO_SINK_CAPS = Gst.Caps.from_string("video/x-raw")
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ VIDEO_SRC_CAPS = Gst.Caps.from_string("video/x-raw")
+ VIDEO_SINK_CAPS = Gst.Caps.from_string("video/x-raw")
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -130,12 +132,137 @@ def iou_batch(bb_det, bb_trk):
class SortTracker:
"""SORT/ByteTrack multi-object tracker using IoU + Kalman filtering."""
- def __init__(self, max_age=30, min_hits=3, iou_threshold=0.3):
+ def __init__(
+ self,
+ max_age=30,
+ min_hits=3,
+ iou_threshold=0.3,
+ keep_alive=2,
+ new_track_conf=0.25,
+ camera_motion=True,
+ dup_iou=0.8,
+ ):
self.max_age = max_age
self.min_hits = min_hits
self.iou_threshold = iou_threshold
+ # ByteTrack-style activation gate: a brand-new track is only started
+ # from a confident detection. Weak/ghost boxes can still *continue* an
+ # existing track (matched above) but won't spawn phantom circles.
+ self.new_track_conf = new_track_conf
+ # Keep emitting a confirmed track (with its Kalman-predicted box) for up
+ # to keep_alive frames after a missed detection — bridges flicker so the
+ # overlay doesn't blink when the detector drops a box for a frame or two.
+ self.keep_alive = keep_alive
+ # Camera-motion compensation: estimate the global image shift from the
+ # tracks that matched, then re-try matching the leftovers with their
+ # predictions shifted by it. Re-attaches players during a pan instead of
+ # leaving the old track behind and spawning a duplicate.
+ self.camera_motion = camera_motion
+ # Two confirmed tracks overlapping more than this IoU are duplicates;
+ # the weaker one is dropped (ByteTrack's remove_duplicate_stracks).
+ self.dup_iou = dup_iou
self.trackers = []
+ @staticmethod
+ def _center(bbox):
+ return (bbox[0] + bbox[2] / 2.0, bbox[1] + bbox[3] / 2.0)
+
+ def _estimate_motion(self, matches, predicted, det_bboxes):
+ """Fit a global 2D similarity transform (translation + uniform scale +
+ rotation) mapping each matched track's predicted centre to its observed
+ centre. Returns a callable box->warped-box, or None if it can't be
+ estimated. Uses RANSAC (via OpenCV) so players moving against the camera
+ consensus are rejected as outliers; falls back to a robust median
+ translation if OpenCV is unavailable or the fit is degenerate."""
+ import numpy as np
+
+ if len(matches) < 3:
+ return None
+ src = np.array(
+ [self._center(predicted[ti]) for _, ti in matches], dtype=np.float32
+ )
+ dst = np.array(
+ [self._center(det_bboxes[di]) for di, _ in matches], dtype=np.float32
+ )
+
+ M = None
+ try:
+ import cv2
+
+ M, _ = cv2.estimateAffinePartial2D(
+ src, dst, method=cv2.RANSAC, ransacReprojThreshold=5.0
+ )
+ except Exception:
+ M = None
+
+ if M is not None:
+ scale = float(np.hypot(M[0, 0], M[0, 1]))
+ # Reject implausible fits (e.g. from too few/noisy correspondences).
+ if 0.5 <= scale <= 2.0:
+
+ def warp(box):
+ cx, cy = self._center(box)
+ ncx = M[0, 0] * cx + M[0, 1] * cy + M[0, 2]
+ ncy = M[1, 0] * cx + M[1, 1] * cy + M[1, 2]
+ nw, nh = box[2] * scale, box[3] * scale
+ return np.array([ncx - nw / 2.0, ncy - nh / 2.0, nw, nh])
+
+ return warp
+
+ # Fallback: robust median translation (pan/tilt only).
+ delta = dst - src
+ tx, ty = float(np.median(delta[:, 0])), float(np.median(delta[:, 1]))
+ if abs(tx) < 1.0 and abs(ty) < 1.0:
+ return None
+ return lambda box: np.array([box[0] + tx, box[1] + ty, box[2], box[3]])
+
+ def _associate(self, det_bboxes, det_idxs, trk_idxs, trk_boxes, detections):
+ """Hungarian-match a subset of detections to a subset of trackers,
+ applying updates to matched trackers. Returns list of (det_i, trk_i)."""
+ from scipy.optimize import linear_sum_assignment
+
+ if not det_idxs or not trk_idxs:
+ return []
+ dets = [det_bboxes[d] for d in det_idxs]
+ trks = [trk_boxes[t] for t in trk_idxs]
+ iou_matrix = iou_batch(dets, trks)
+ if iou_matrix.size == 0:
+ return []
+ cost = 1.0 - iou_matrix
+ row_ind, col_ind = linear_sum_assignment(cost)
+ matches = []
+ for r, c in zip(row_ind, col_ind):
+ if iou_matrix[r, c] >= self.iou_threshold:
+ di, ti = det_idxs[r], trk_idxs[c]
+ self.trackers[ti].update(detections[di][:4])
+ self.trackers[ti].label_quark = detections[di][5]
+ matches.append((di, ti))
+ return matches
+
+ def _suppress_duplicates(self):
+ """Drop the weaker of any two confirmed tracks sitting on the same box."""
+ n = len(self.trackers)
+ if n < 2:
+ return
+ boxes = [t.get_bbox() for t in self.trackers]
+ iou_matrix = iou_batch(boxes, boxes)
+ remove = set()
+ for i in range(n):
+ if i in remove:
+ continue
+ for j in range(i + 1, n):
+ if j in remove:
+ continue
+ if iou_matrix[i, j] > self.dup_iou:
+ ti, tj = self.trackers[i], self.trackers[j]
+ # Keep the better track: matched more recently, then more
+ # hits; drop the other (usually the freshly-spawned dup).
+ ki = (ti.time_since_update, -ti.hits)
+ kj = (tj.time_since_update, -tj.hits)
+ remove.add(j if ki <= kj else i)
+ if remove:
+ self.trackers = [t for k, t in enumerate(self.trackers) if k not in remove]
+
def update(self, detections):
"""
Update tracks with new detections.
@@ -147,54 +274,66 @@ def update(self, detections):
list of (track_id, bbox, label_quark) for confirmed tracks
"""
import numpy as np
- from scipy.optimize import linear_sum_assignment
# Predict new locations for existing tracks
- predicted = []
to_remove = []
for i, trk in enumerate(self.trackers):
- pred = trk.predict()
- if np.any(np.isnan(pred)):
+ if np.any(np.isnan(trk.predict())):
to_remove.append(i)
- else:
- predicted.append(pred)
for i in reversed(to_remove):
self.trackers.pop(i)
- # Build cost matrix using IoU
det_bboxes = [d[:4] for d in detections] if len(detections) > 0 else []
- iou_matrix = iou_batch(det_bboxes, predicted)
- cost_matrix = 1.0 - iou_matrix
-
- # Hungarian assignment
- matched_det = set()
- matched_trk = set()
- if cost_matrix.size > 0:
- row_ind, col_ind = linear_sum_assignment(cost_matrix)
- for r, c in zip(row_ind, col_ind):
- if iou_matrix[r, c] >= self.iou_threshold:
- matched_det.add(r)
- matched_trk.add(c)
- self.trackers[c].update(detections[r][:4])
- # Store latest label quark on tracker
- self.trackers[c].label_quark = detections[r][5]
-
- # Create new tracks for unmatched detections
- for d_idx in range(len(detections)):
+ n_det = len(det_bboxes)
+ n_trk = len(self.trackers)
+ # Predicted box per tracker, captured before any update this frame.
+ predicted = [self.trackers[i].get_bbox() for i in range(n_trk)]
+
+ # 1) First association on the raw predictions.
+ matches = self._associate(
+ det_bboxes, list(range(n_det)), list(range(n_trk)), predicted, detections
+ )
+ matched_det = {di for di, _ in matches}
+ matched_trk = {ti for _, ti in matches}
+
+ # 2) Camera-motion compensation: fit a global image transform (pan, zoom
+ # and rotation) from the tracks that matched, apply it to the leftover
+ # predictions, and re-match. This recovers tracks during camera moves
+ # instead of leaving them behind and spawning duplicates.
+ if self.camera_motion:
+ warp = self._estimate_motion(matches, predicted, det_bboxes)
+ if warp is not None:
+ rem_trk = [i for i in range(n_trk) if i not in matched_trk]
+ rem_det = [d for d in range(n_det) if d not in matched_det]
+ if rem_trk and rem_det:
+ shifted = {i: warp(predicted[i]) for i in rem_trk}
+ m2 = self._associate(
+ det_bboxes, rem_det, rem_trk, shifted, detections
+ )
+ matched_det.update(di for di, _ in m2)
+
+ # Create new tracks for unmatched detections, but only from confident
+ # ones (ByteTrack activation gate) so weak/ghost boxes don't start a
+ # phantom track that gets drawn as a stray circle.
+ for d_idx in range(n_det):
if d_idx not in matched_det:
+ if detections[d_idx][4] < self.new_track_conf:
+ continue
trk = KalmanBoxTracker(detections[d_idx][:4])
trk.label_quark = detections[d_idx][5]
self.trackers.append(trk)
- # Remove dead tracks
+ # Remove dead tracks, then drop duplicate tracks sitting on one object.
self.trackers = [
t for t in self.trackers if t.time_since_update <= self.max_age
]
+ self._suppress_duplicates()
- # Return confirmed tracks
+ # Return confirmed tracks, including ones that missed a detection this
+ # frame (predicted box) for up to keep_alive frames — prevents flicker.
results = []
for trk in self.trackers:
- if trk.hits >= self.min_hits and trk.time_since_update == 0:
+ if trk.hits >= self.min_hits and trk.time_since_update <= self.keep_alive:
results.append((trk.id, trk.get_bbox(), trk.label_quark))
return results
@@ -215,20 +354,21 @@ class TrackerTransform(GstBase.BaseTransform):
"Aaron Boxer ",
)
- src_template = Gst.PadTemplate.new(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- VIDEO_SRC_CAPS.copy(),
- )
-
- sink_template = Gst.PadTemplate.new(
- "sink",
- Gst.PadDirection.SINK,
- Gst.PadPresence.ALWAYS,
- VIDEO_SINK_CAPS.copy(),
- )
- __gsttemplates__ = (src_template, sink_template)
+ if backend.BACKEND == "gst":
+ src_template = Gst.PadTemplate.new(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ VIDEO_SRC_CAPS.copy(),
+ )
+
+ sink_template = Gst.PadTemplate.new(
+ "sink",
+ Gst.PadDirection.SINK,
+ Gst.PadPresence.ALWAYS,
+ VIDEO_SINK_CAPS.copy(),
+ )
+ __gsttemplates__ = (src_template, sink_template)
tracker_type = GObject.Property(
type=str,
@@ -268,6 +408,50 @@ class TrackerTransform(GstBase.BaseTransform):
flags=GObject.ParamFlags.READWRITE,
)
+ keep_alive = GObject.Property(
+ type=int,
+ default=2,
+ minimum=0,
+ maximum=1000,
+ nick="Keep Alive",
+ blurb="Frames to keep emitting a confirmed track (Kalman-predicted box) "
+ "after a missed detection; bridges flicker (0 = only matched frames)",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ new_track_confidence = GObject.Property(
+ type=float,
+ default=0.25,
+ minimum=0.0,
+ maximum=1.0,
+ nick="New Track Confidence",
+ blurb="Minimum detection confidence to START a new track (ByteTrack "
+ "activation gate); weak boxes still continue existing tracks but "
+ "won't spawn phantom/duplicate circles",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ camera_motion = GObject.Property(
+ type=bool,
+ default=True,
+ nick="Camera Motion Compensation",
+ blurb="Estimate the global image shift from matched tracks and re-match "
+ "leftovers shifted by it, so a panning camera re-attaches players "
+ "instead of leaving the old track behind and spawning a duplicate",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
+ duplicate_iou = GObject.Property(
+ type=float,
+ default=0.8,
+ minimum=0.0,
+ maximum=1.0,
+ nick="Duplicate IoU",
+ blurb="Two confirmed tracks overlapping more than this are treated as "
+ "duplicates and the weaker one is dropped",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
def __init__(self):
super().__init__()
self.logger = LoggerFactory.get(LoggerFactory.LOGGER_TYPE_GST)
@@ -281,25 +465,24 @@ def _ensure_tracker(self):
max_age=self.max_age,
min_hits=self.min_hits,
iou_threshold=self.iou_threshold,
+ keep_alive=self.keep_alive,
+ new_track_conf=self.new_track_confidence,
+ camera_motion=self.camera_motion,
+ dup_iou=self.duplicate_iou,
)
return self._tracker
def _read_detections(self, buf):
- """Extract detections from upstream GstAnalytics od_mtd."""
+ """Extract detections from upstream object-detection metadata."""
detections = []
- meta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
+ meta = analytics.get_relation_meta(buf)
if not meta:
return detections
- count = GstAnalytics.relation_get_length(meta)
- for index in range(count):
- ret, od_mtd = meta.get_od_mtd(index)
- if not ret or od_mtd is None:
- continue
- label_quark = od_mtd.get_obj_type()
- presence, x, y, w, h, score = od_mtd.get_location()
- if presence:
- detections.append([x, y, w, h, score, label_quark])
+ for obj in analytics.read_objects(meta):
+ detections.append(
+ [obj["x"], obj["y"], obj["w"], obj["h"], obj["score"], obj["label"]]
+ )
return detections
def do_transform_ip(self, buf):
@@ -315,20 +498,20 @@ def do_transform_ip(self, buf):
tracked = tracker.update(detections)
# Attach tracking results as new analytics metadata
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
+ meta = analytics.add_relation_meta(buf)
if not meta:
self.logger.error(
"Failed to add analytics relation metadata for tracking"
)
return Gst.FlowReturn.ERROR
- for track_id, bbox, label_quark in tracked:
- label_str = GLib.quark_to_string(label_quark)
+ for track_id, bbox, label_str in tracked:
track_label = f"{label_str}_id_{track_id}"
- qk = GLib.quark_from_string(track_label)
x, y, w, h = bbox
- ret, od_mtd = meta.add_od_mtd(qk, int(x), int(y), int(w), int(h), 1.0)
- if not ret:
+ od_mtd = analytics.add_object(
+ meta, track_label, int(x), int(y), int(w), int(h), 1.0
+ )
+ if od_mtd is None:
self.logger.error(
f"Failed to add tracking od_mtd for track {track_id}"
)
@@ -351,6 +534,14 @@ def do_get_property(self, prop):
return self.min_hits
elif prop.name == "iou-threshold":
return self.iou_threshold
+ elif prop.name == "keep-alive":
+ return self.keep_alive
+ elif prop.name == "new-track-confidence":
+ return self.new_track_confidence
+ elif prop.name == "camera-motion":
+ return self.camera_motion
+ elif prop.name == "duplicate-iou":
+ return self.duplicate_iou
else:
raise AttributeError(f"Unknown property {prop.name}")
@@ -367,14 +558,27 @@ def do_set_property(self, prop, value):
elif prop.name == "iou-threshold":
self.iou_threshold = value
self._tracker = None
+ elif prop.name == "keep-alive":
+ self.keep_alive = value
+ self._tracker = None
+ elif prop.name == "new-track-confidence":
+ self.new_track_confidence = value
+ self._tracker = None
+ elif prop.name == "camera-motion":
+ self.camera_motion = value
+ self._tracker = None
+ elif prop.name == "duplicate-iou":
+ self.duplicate_iou = value
+ self._tracker = None
else:
raise AttributeError(f"Unknown property {prop.name}")
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(TrackerTransform)
- __gstelementfactory__ = ("pyml_tracker", Gst.Rank.NONE, TrackerTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_tracker", TrackerTransform
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_tracker' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/vad.py b/plugins/python/vad.py
index 1d9757c..f78ea97 100644
--- a/plugins/python/vad.py
+++ b/plugins/python/vad.py
@@ -17,6 +17,7 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
@@ -27,7 +28,8 @@
gi.require_version("Gst", "1.0")
gi.require_version("GstBase", "1.0")
- from gi.repository import Gst, GObject, GstBase
+ from gi.repository import Gst, GstBase
+ from backend import GObject
from log.logger_factory import LoggerFactory
@@ -72,17 +74,19 @@ class VoiceActivityDetector(GstBase.BaseTransform):
"Aaron Boxer ",
)
- AUDIO_CAPS = Gst.Caps.from_string(
- "audio/x-raw,format=S16LE,layout=interleaved,rate=16000,channels=1"
- )
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ AUDIO_CAPS = Gst.Caps.from_string(
+ "audio/x-raw,format=S16LE,layout=interleaved,rate=16000,channels=1"
+ )
- sink_template = Gst.PadTemplate.new(
- "sink", Gst.PadDirection.SINK, Gst.PadPresence.ALWAYS, AUDIO_CAPS
- )
- src_template = Gst.PadTemplate.new(
- "src", Gst.PadDirection.SRC, Gst.PadPresence.ALWAYS, AUDIO_CAPS
- )
- __gsttemplates__ = (sink_template, src_template)
+ sink_template = Gst.PadTemplate.new(
+ "sink", Gst.PadDirection.SINK, Gst.PadPresence.ALWAYS, AUDIO_CAPS
+ )
+ src_template = Gst.PadTemplate.new(
+ "src", Gst.PadDirection.SRC, Gst.PadPresence.ALWAYS, AUDIO_CAPS
+ )
+ __gsttemplates__ = (sink_template, src_template)
threshold = GObject.Property(
type=float,
@@ -187,10 +191,11 @@ def do_transform_ip(self, buf):
return Gst.FlowReturn.OK
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(VoiceActivityDetector)
- __gstelementfactory__ = ("pyml_vad", Gst.Rank.NONE, VoiceActivityDetector)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_vad", VoiceActivityDetector
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_vad' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/video_transform.py b/plugins/python/video_transform.py
index 0159a01..2c24ec7 100644
--- a/plugins/python/video_transform.py
+++ b/plugins/python/video_transform.py
@@ -16,38 +16,11 @@
# Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
# Boston, MA 02110-1301, USA.
-import gi
+"""Compatibility shim. `VideoTransform` now lives behind the pluggable backend
+in `backend/` (see `backend/__init__.py`). Leaf plugins keep importing it from
+here; the active backend is chosen by the PYML_BACKEND environment variable.
+"""
-gi.require_version("Gst", "1.0")
-gi.require_version("GstBase", "1.0")
-gi.require_version("GstVideo", "1.0")
-from gi.repository import Gst # noqa: E402
+from backend import VideoTransform
-from base_transform import BaseTransform # noqa: E402
-
-
-class VideoTransform(BaseTransform):
- """
- GStreamer element for video transformation using a PyTorch model.
- """
-
- # Define VIDEO_CAPS to support multiple formats
- VIDEO_CAPS = Gst.Caps.from_string(
- "video/x-raw,format=(string){ RGB, RGBA, ARGB, BGRA, ABGR },"
- "width=(int)[1,2147483647],height=(int)[1,2147483647]"
- )
- __gsttemplates__ = (
- Gst.PadTemplate.new(
- "src", Gst.PadDirection.SRC, Gst.PadPresence.ALWAYS, VIDEO_CAPS
- ),
- Gst.PadTemplate.new(
- "sink", Gst.PadDirection.SINK, Gst.PadPresence.ALWAYS, VIDEO_CAPS
- ),
- )
-
- def do_set_caps(self, incaps, outcaps):
- struct = incaps.get_structure(0)
- self.width = struct.get_int("width").value
- self.height = struct.get_int("height").value
-
- return True
+__all__ = ["VideoTransform"]
diff --git a/plugins/python/vlm.py b/plugins/python/vlm.py
index 1cac59b..bf71976 100644
--- a/plugins/python/vlm.py
+++ b/plugins/python/vlm.py
@@ -17,22 +17,15 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import json
-
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- from gi.repository import Gst, GObject
-
from video_transform import VideoTransform
- from utils.format_converter import FormatConverter
- from engine.pytorch_engine import PyTorchEngine
+ from engine.vlm_engine import VlmEngine
from engine.engine_factory import EngineFactory
+ from backend import frameio, GObject
+ from tasks.vlm import VlmTask
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -42,126 +35,7 @@
VLM_META_HEADER = b"GST-VLM:"
-class VlmEngine(PyTorchEngine):
- """
- PyTorch engine for Vision-Language Models.
-
- Supports HuggingFace VLM model IDs via AutoProcessor + AutoModelForVision2Seq:
- llava-hf/llava-1.5-7b-hf
- Qwen/Qwen2-VL-7B-Instruct
- OpenGVLab/InternVL2-8B
- """
-
- def __init__(self):
- super().__init__()
- self.processor = None
-
- def do_load_model(self, model_name, **kwargs):
- try:
- import torch
- from transformers import AutoProcessor, AutoModelForVision2Seq
-
- self.processor = AutoProcessor.from_pretrained(model_name)
- self.model = AutoModelForVision2Seq.from_pretrained(
- model_name,
- torch_dtype=torch.float16,
- device_map=self.device,
- )
- self.model.eval()
- self.logger.info(f"VLM model '{model_name}' loaded on {self.device}")
- except Exception as e:
- raise ValueError(f"Failed to load VLM model '{model_name}': {e}")
-
- def do_forward(
- self,
- frame,
- prompt="Describe this image in detail.",
- system_prompt=None,
- max_tokens=256,
- temperature=0.7,
- ):
- """
- Run VLM inference on a single video frame.
-
- Args:
- frame: numpy RGB array (H, W, 3).
- prompt: user prompt text.
- system_prompt: optional system prompt.
- max_tokens: maximum tokens to generate.
- temperature: sampling temperature.
-
- Returns:
- Generated text string, or None on failure.
- """
- import numpy as np
- from PIL import Image
-
- try:
- pil_img = Image.fromarray(frame.astype(np.uint8))
- text = self.do_generate(
- pil_img, prompt, system_prompt, max_tokens, temperature
- )
- return text
- except Exception as e:
- self.logger.error(f"VLM inference error: {e}")
- return None
-
- def do_generate(self, image, prompt, system_prompt, max_tokens, temperature):
- """
- Apply chat template, process image + text, and generate a response.
-
- Args:
- image: PIL Image.
- prompt: user prompt text.
- system_prompt: optional system prompt.
- max_tokens: maximum new tokens.
- temperature: sampling temperature.
-
- Returns:
- Generated text string.
- """
- import torch
-
- messages = []
- if system_prompt:
- messages.append({"role": "system", "content": system_prompt})
- messages.append(
- {
- "role": "user",
- "content": [
- {"type": "image"},
- {"type": "text", "text": prompt},
- ],
- }
- )
-
- # Apply chat template if the processor supports it
- if hasattr(self.processor, "apply_chat_template"):
- text_input = self.processor.apply_chat_template(
- messages, add_generation_prompt=True
- )
- else:
- text_input = prompt
-
- inputs = self.processor(text=text_input, images=image, return_tensors="pt")
- inputs = {k: v.to(self.model.device) for k, v in inputs.items()}
-
- with torch.no_grad():
- output_ids = self.model.generate(
- **inputs,
- max_new_tokens=max_tokens,
- temperature=temperature,
- do_sample=temperature > 0,
- )
-
- # Decode only the newly generated tokens
- input_len = inputs.get("input_ids", torch.tensor([])).shape[-1]
- generated = output_ids[0][input_len:]
- text = self.processor.decode(generated, skip_special_tokens=True)
- return text.strip()
-
-
-class VlmTransform(VideoTransform):
+class VlmTransform(VideoTransform, VlmTask):
"""
GStreamer element for Vision-Language Model inference on video frames.
@@ -227,7 +101,6 @@ def __init__(self):
super().__init__()
self.mgr.engine_name = "pyml_vlm_engine"
EngineFactory.register(self.mgr.engine_name, VlmEngine)
- self.format_converter = FormatConverter()
self._frame_count = 0
@GObject.Property(type=str)
@@ -239,68 +112,37 @@ def engine_name(self):
def engine_name(self, value):
raise ValueError("'engine_name' is read-only for pyml_vlm")
- def do_transform_ip(self, buf):
- try:
- self._frame_count += 1
- if self.frame_stride > 1 and (self._frame_count % self.frame_stride) != 1:
- return Gst.FlowReturn.OK
-
- if self.engine is None:
- return Gst.FlowReturn.OK
-
- success, map_info = buf.map(Gst.MapFlags.READ)
- if not success:
- self.logger.error("Failed to map video buffer for reading")
- return Gst.FlowReturn.ERROR
+ def process_frames(self, frames, num_sources, fmt, target):
+ """Caption the frame and append the payload, skipping strided-out frames."""
+ self._frame_count += 1
+ if self.frame_stride > 1 and (self._frame_count % self.frame_stride) != 1:
+ return
- try:
- frame = self.format_converter.to_rgb(
- map_info.data, self.width, self.height, buf, self.sinkpad
- )
- finally:
- buf.unmap(map_info)
+ if self.engine is None:
+ return
- if frame is None:
- return Gst.FlowReturn.ERROR
+ frame = frames[0] if num_sources > 1 else frames
- text = self.engine.do_forward(
- frame,
- prompt=self.prompt,
- system_prompt=self.system_prompt,
- max_tokens=self.max_tokens,
- temperature=self.temperature,
- )
+ text = self.forward(frame)
- if text is None:
- return Gst.FlowReturn.OK
+ if text is None:
+ return
- result = {"text": text}
- # Append VLM response as a JSON memory chunk.
- # Use new_allocate+fill: PyGI hides the maxsize arg in new_wrapped
- # (it derives it from data length), so passing it explicitly shifts
- # all subsequent args and causes a GI assertion crash.
- meta_bytes = VLM_META_HEADER + json.dumps(result).encode("utf-8")
- tmp = Gst.Buffer.new_allocate(None, len(meta_bytes), None)
- tmp.fill(0, meta_bytes)
- buf.append_memory(tmp.get_memory(0))
+ # Portable task: serialize the VLM response payload.
+ _, payload = self.decode(text)
+ # Append VLM response as a JSON memory chunk.
+ frameio.write_result(target, None, payload, VLM_META_HEADER)
- self.logger.debug(
- f"VLM response ({len(text)} chars): {text[:80]}..."
- if len(text) > 80
- else f"VLM response: {text}"
- )
-
- return Gst.FlowReturn.OK
-
- except Exception as e:
- self.logger.error(f"VLM transform error: {e}")
- return Gst.FlowReturn.ERROR
+ self.logger.debug(
+ f"VLM response ({len(text)} chars): {text[:80]}..."
+ if len(text) > 80
+ else f"VLM response: {text}"
+ )
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(VlmTransform)
- __gstelementfactory__ = ("pyml_vlm", Gst.Rank.NONE, VlmTransform)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_vlm", VlmTransform)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_vlm' element will not be registered because required modules are missing."
)
diff --git a/plugins/python/whisperlive.py b/plugins/python/whisperlive.py
index 9f55a16..c5f12a9 100644
--- a/plugins/python/whisperlive.py
+++ b/plugins/python/whisperlive.py
@@ -17,15 +17,11 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GObject", "2.0")
- from gi.repository import Gst, GObject, GstBase # noqa: E402
+ from backend import GObject
from base_transcribe import BaseTranscribe
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -33,19 +29,14 @@
f"The 'pyml_whisperlive' element will not be available. Error: {e}"
)
-TTS_SAMPLE_RATE = 24000
-model_ref = "collabora/whisperspeech:s2a-q4-base-en+pl.model"
+if backend.BACKEND == "gst":
+ import gi
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ from gi.repository import Gst, GstBase # noqa: E402
-OCAPS = Gst.Caps(
- Gst.Structure(
- "audio/x-raw",
- format="S16LE",
- layout="interleaved",
- rate=TTS_SAMPLE_RATE,
- channels=1,
- )
-)
+model_ref = "collabora/whisperspeech:s2a-q4-base-en+pl.model"
class WhisperLive(BaseTranscribe):
@@ -56,13 +47,17 @@ class WhisperLive(BaseTranscribe):
"Aaron Boxer ",
)
- __gsttemplates__ = Gst.PadTemplate.new_with_gtype(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- OCAPS,
- GstBase.AggregatorPad.__gtype__,
- )
+ OUTPUT_CAPS = "audio/x-raw,format=S16LE,layout=interleaved,rate=24000,channels=1"
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = Gst.PadTemplate.new_with_gtype(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string(OUTPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ )
llm_model_name = GObject.Property(
type=str,
@@ -206,19 +201,14 @@ def do_process_text(self, transcript):
else:
audio_np = audio_np.T # Transpose the numpy array if it's not 1D
- duration = len(audio_np) / TTS_SAMPLE_RATE * Gst.SECOND
- buffer = Gst.Buffer.new_wrapped(audio_np.tobytes())
-
- buffer.pts = Gst.CLOCK_TIME_NONE
- buffer.duration = duration
+ return audio_np.tobytes()
- return buffer
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(WhisperLive)
- __gstelementfactory__ = ("pyml_whisperlive", Gst.Rank.NONE, WhisperLive)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_whisperlive", WhisperLive
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_whisperlive' element will not be registered because required modules were missing."
)
diff --git a/plugins/python/whisperspeechtts.py b/plugins/python/whisperspeechtts.py
index ca9c929..3ba5004 100644
--- a/plugins/python/whisperspeechtts.py
+++ b/plugins/python/whisperspeechtts.py
@@ -17,14 +17,10 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- from gi.repository import Gst, GObject, GstBase # noqa: E402
from base_tts import BaseTts
except ImportError as e:
CAN_REGISTER_ELEMENT = False
@@ -32,20 +28,17 @@
f"The 'pyml_whisperspeechtts' element will not be available. Error: {e}"
)
+if backend.BACKEND == "gst":
+ import gi
+
+ gi.require_version("Gst", "1.0")
+ gi.require_version("GstBase", "1.0")
+ from gi.repository import Gst, GstBase # noqa: E402
+
TTS_SAMPLE_RATE = 24000
model_ref = "collabora/whisperspeech:s2a-q4-base-en+pl.model"
-OCAPS = Gst.Caps(
- Gst.Structure(
- "audio/x-raw",
- format="S16LE",
- layout="interleaved",
- rate=TTS_SAMPLE_RATE,
- channels=1,
- )
-)
-
class WhisperSpeechTTS(BaseTts):
__gstmetadata__ = (
@@ -55,15 +48,20 @@ class WhisperSpeechTTS(BaseTts):
"Aaron Boxer ",
)
- __gsttemplates__ = (
- Gst.PadTemplate.new_with_gtype(
- "src",
- Gst.PadDirection.SRC,
- Gst.PadPresence.ALWAYS,
- OCAPS,
- GstBase.AggregatorPad.__gtype__,
- ),
- )
+ # the rate has to match TTS_SAMPLE_RATE, which the element reports downstream
+ OUTPUT_CAPS = "audio/x-raw,format=S16LE,layout=interleaved,rate=24000,channels=1"
+
+ # Building a Gst object needs Gst.init, which only the gst backend calls.
+ if backend.BACKEND == "gst":
+ __gsttemplates__ = (
+ Gst.PadTemplate.new_with_gtype(
+ "src",
+ Gst.PadDirection.SRC,
+ Gst.PadPresence.ALWAYS,
+ Gst.Caps.from_string(OUTPUT_CAPS),
+ GstBase.AggregatorPad.__gtype__,
+ ),
+ )
def do_load_model(self):
from whisperspeech.pipeline import Pipeline
@@ -103,10 +101,11 @@ def do_get_sample_rate(self):
return TTS_SAMPLE_RATE
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(WhisperSpeechTTS)
- __gstelementfactory__ = ("pyml_whisperspeechtts", Gst.Rank.NONE, WhisperSpeechTTS)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_whisperspeechtts", WhisperSpeechTTS
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_whisperspeechtts' element will not be registered because required modules were missing."
)
diff --git a/plugins/python/whispertranscribe.py b/plugins/python/whispertranscribe.py
index c70a94c..2d6f551 100644
--- a/plugins/python/whispertranscribe.py
+++ b/plugins/python/whispertranscribe.py
@@ -17,17 +17,13 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GObject", "2.0")
- from gi.repository import Gst, GObject # noqa: E402
+ from backend import GObject
from base_transcribe import BaseTranscribe
- from engine.pytorch_engine import PyTorchEngine
+ from engine.whisper_engine import WhisperEngine
from engine.engine_factory import EngineFactory
except ImportError as e:
@@ -37,21 +33,6 @@
)
-class WhisperEngine(PyTorchEngine):
- def do_load_model(self, model_name, **kwargs):
- from faster_whisper import WhisperModel
-
- if not model_name:
- return
- compute_type = "float16" if self.device.startswith("cuda") else "int8"
- self.logger.info(
- f"Loading Whisper model on device: {self.device} with compute_type: {compute_type}"
- )
- self.model = WhisperModel(
- model_name, device=self.device, compute_type=compute_type
- )
-
-
class WhisperTranscribe(BaseTranscribe):
__gstmetadata__ = (
"WhisperTranscribe",
@@ -89,10 +70,11 @@ def do_transcribe(self, audio_data, task):
return result
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(WhisperTranscribe)
- __gstelementfactory__ = ("pyml_whispertranscribe", Gst.Rank.NONE, WhisperTranscribe)
-else:
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element(
+ "pyml_whispertranscribe", WhisperTranscribe
+ )
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_whispertranscribe' element will not be registered because base_transcribe module is missing."
)
diff --git a/plugins/python/yolo.py b/plugins/python/yolo.py
index a34deb6..517b158 100644
--- a/plugins/python/yolo.py
+++ b/plugins/python/yolo.py
@@ -17,185 +17,28 @@
# Boston, MA 02110-1301, USA.
from log.global_logger import GlobalLogger
+from backend import GObject
+import backend
CAN_REGISTER_ELEMENT = True
try:
- import gi
-
- gi.require_version("Gst", "1.0")
- gi.require_version("GstBase", "1.0")
- gi.require_version("GstVideo", "1.0")
- gi.require_version("GstAnalytics", "1.0")
- gi.require_version("GLib", "2.0")
- from gi.repository import Gst, GObject, GstAnalytics, GLib # noqa: E402
from base_objectdetector import BaseObjectDetector
+ from tasks.yolo import YoloTask
- import time
- from engine.pytorch_engine import PyTorchEngine
+ from engine.yolo_engine import YoloEngine
from engine.engine_factory import EngineFactory
except ImportError as e:
CAN_REGISTER_ELEMENT = False
GlobalLogger().warning(f"The 'yolo' element will not be available. Error {e}")
-COCO_CLASSES = {
- 0: "person",
- 1: "bicycle",
- 2: "car",
- 3: "motorcycle",
- 4: "airplane",
- 5: "bus",
- 6: "train",
- 7: "truck",
- 8: "boat",
- 9: "traffic light",
- 10: "fire hydrant",
- 11: "stop sign",
- 12: "parking meter",
- 13: "bench",
- 14: "bird",
- 15: "cat",
- 16: "dog",
- 17: "horse",
- 18: "sheep",
- 19: "cow",
- 20: "elephant",
- 21: "bear",
- 22: "zebra",
- 23: "giraffe",
- 24: "backpack",
- 25: "umbrella",
- 26: "handbag",
- 27: "tie",
- 28: "suitcase",
- 29: "frisbee",
- 30: "skis",
- 31: "snowboard",
- 32: "sports ball",
- 33: "kite",
- 34: "baseball bat",
- 35: "baseball glove",
- 36: "skateboard",
- 37: "surfboard",
- 38: "tennis racket",
- 39: "bottle",
- 40: "wine glass",
- 41: "cup",
- 42: "fork",
- 43: "knife",
- 44: "spoon",
- 45: "bowl",
- 46: "banana",
- 47: "apple",
- 48: "sandwich",
- 49: "orange",
- 50: "broccoli",
- 51: "carrot",
- 52: "hot dog",
- 53: "pizza",
- 54: "donut",
- 55: "cake",
- 56: "chair",
- 57: "couch",
- 58: "potted plant",
- 59: "bed",
- 60: "dining table",
- 61: "toilet",
- 62: "TV",
- 63: "laptop",
- 64: "mouse",
- 65: "remote",
- 66: "keyboard",
- 67: "cell phone",
- 68: "microwave",
- 69: "oven",
- 70: "toaster",
- 71: "sink",
- 72: "refrigerator",
- 73: "book",
- 74: "clock",
- 75: "vase",
- 76: "scissors",
- 77: "teddy bear",
- 78: "hair drier",
- 79: "toothbrush",
-}
-
-
-class YoloEngine(PyTorchEngine):
- def do_load_model(self, model_name, **kwargs):
- try:
- from ultralytics import YOLO
-
- self.model = YOLO(f"{model_name}.pt")
- self.execute_with_stream(lambda: self.model.to(self.device))
- self.logger.info(f"YOLO model '{model_name}' loaded on {self.device}")
- except Exception as e:
- raise ValueError(f"Failed to load YOLO model '{model_name}'. Error: {e}")
-
- def do_forward(self, frames):
- import numpy as np
-
- is_batch = isinstance(frames, np.ndarray) and frames.ndim == 4
- writable_frames = np.array(frames, copy=True)
- batch_size = writable_frames.shape[0] if is_batch else 1
-
- model = self.get_model()
- if model is None:
- self.logger.error("Model is not loaded.")
- return None if not is_batch else [None] * batch_size
- try:
- start_pre = time.time()
- img_list = (
- [
- writable_frames[i] if is_batch else writable_frames
- for i in range(batch_size)
- ]
- if is_batch
- else [writable_frames]
- )
- self.logger.debug(
- f"Input shape: {writable_frames.shape}, min={writable_frames.min()}, max={writable_frames.max()}"
- )
- end_pre = time.time()
-
- if self.track:
- # Ensure tracker persists across batches
- results = self.execute_with_stream(
- lambda: model.track(
- source=img_list,
- persist=True,
- imgsz=640,
- conf=0.1,
- verbose=True,
- tracker="botsort.yaml",
- )
- )
- else:
- results = self.execute_with_stream(
- lambda: model(img_list, imgsz=640, conf=0.1, verbose=True)
- )
- end_inf = time.time()
-
- if results is None or (isinstance(results, list) and not results):
- self.logger.warning("Inference returned None or empty list.")
- return None if not is_batch else [None] * batch_size
-
- self.logger.info(
- f"Preprocessing: {(end_pre - start_pre)*1000:.2f} ms, Inference: {(end_inf - end_pre)*1000:.2f} ms for {batch_size} frames"
- )
- return results[0] if not is_batch else results
-
- except Exception as e:
- self.logger.error(f"Error during inference: {e}")
- return None if not is_batch else [None] * batch_size
-
-
-class YOLOTransform(BaseObjectDetector):
+class YOLOTransform(BaseObjectDetector, YoloTask):
"""
- GStreamer element for YOLO model inference on video frames
- (detection, segmentation, and tracking).
+ GStreamer element shell for YOLO model inference on video frames
+ (detection, segmentation, and tracking). The result handling (do_decode)
+ is inherited from the backend-agnostic YoloTask; this class supplies the
+ engine wiring, the read-only engine_name property, and registration.
"""
__gstmetadata__ = (
@@ -205,6 +48,36 @@ class YOLOTransform(BaseObjectDetector):
"Aaron Boxer ",
)
+ confidence = GObject.Property(
+ type=float,
+ default=0.1,
+ minimum=0.0,
+ maximum=1.0,
+ nick="Confidence Threshold",
+ blurb="Minimum detection confidence (matches football_analyzer); kept "
+ "low on purpose so the tracker can use weak boxes to continue tracks "
+ "-- the tracker's new-track-confidence gates phantom tracks",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ nms_iou = GObject.Property(
+ type=float,
+ default=0.7,
+ minimum=0.0,
+ maximum=1.0,
+ nick="NMS IoU",
+ blurb="NMS IoU threshold (matches football_analyzer's default); lower "
+ "suppresses more overlap but can also drop genuinely close players",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+ agnostic_nms = GObject.Property(
+ type=bool,
+ default=False,
+ nick="Class-Agnostic NMS",
+ blurb="Suppress overlapping boxes across classes too; off by default "
+ "(like football_analyzer) so two close players aren't merged",
+ flags=GObject.ParamFlags.READWRITE,
+ )
+
def __init__(self):
super().__init__()
self.mgr.engine_name = "pyml_yolo_engine"
@@ -222,112 +95,21 @@ def engine_name(self, value):
"The 'engine_name' property cannot be set in this derived class."
)
- def do_decode(self, buf, result, stream_idx=0):
- self.logger.debug(
- f"Decoding YOLO result for buffer {hex(id(buf))}, stream {stream_idx}: {result}"
- )
- boxes = result.boxes
- masks = None
- if not self.engine.track:
- masks = result.masks
-
- if boxes is None or len(boxes) == 0:
- self.logger.info("No detections found.")
- return
-
- meta = GstAnalytics.buffer_add_analytics_relation_meta(buf)
- if not meta:
- self.logger.error(
- f"Stream {stream_idx} - Failed to add analytics relation metadata"
- )
- return
-
- self.logger.debug(
- f"Stream {stream_idx} - Attaching metadata for {len(boxes)} detections"
- )
- for i in range(len(boxes)):
- x1, y1, x2, y2 = boxes.xyxy[i]
- score = boxes.conf[i]
- label = boxes.cls[i]
- label_num = label.item()
- class_name = COCO_CLASSES.get(label_num, f"unknown_{label_num}")
-
- # Use class name for detection, track_id for tracking
- if self.engine.track and hasattr(boxes, "id") and boxes.id is not None:
- track_id = boxes.id[i]
- track_id_int = int(track_id.item())
- qk_string = f"stream_{stream_idx}_id_{track_id_int}"
- else:
- qk_string = (
- f"stream_{stream_idx}_{class_name}" # No index, just class name
- )
-
- qk = GLib.quark_from_string(qk_string)
- ret, od_mtd = meta.add_od_mtd(
- qk,
- x1.item(),
- y1.item(),
- x2.item() - x1.item(),
- y2.item() - y1.item(),
- score.item(),
- )
- if not ret:
- self.logger.error(
- f"Stream {stream_idx} - Failed to add object detection metadata"
- )
- continue
- self.logger.debug(
- f"Stream {stream_idx} - Added od_mtd: label={qk_string}, x1={x1.item()}, y1={y1.item()}, w={x2.item()-x1.item()}, h={y2.item()-y1.item()}, score={score.item()}"
- )
-
- # Tracking metadata only when track=True
- if self.engine.track and hasattr(boxes, "id") and boxes.id is not None:
- ret, tracking_mtd = meta.add_tracking_mtd(
- track_id_int, Gst.util_get_timestamp()
- )
- if not ret:
- self.logger.error(
- f"Stream {stream_idx} - Failed to add tracking metadata"
- )
- continue
- ret = GstAnalytics.RelationMeta.set_relation(
- meta, GstAnalytics.RelTypes.RELATE_TO, od_mtd.id, tracking_mtd.id
- )
- if not ret:
- self.logger.error(
- f"Stream {stream_idx} - Failed to relate object detection and tracking metadata"
- )
- else:
- self.logger.debug(
- f"Stream {stream_idx} - Linked od_mtd {od_mtd.id} to tracking_mtd {tracking_mtd.id}"
- )
-
- if masks is not None:
- self.add_segmentation_metadata(buf, masks[i], x1, y1, x2, y2)
-
- attached_meta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
- if attached_meta:
- count = GstAnalytics.relation_get_length(attached_meta)
- self.logger.info(
- f"Stream {stream_idx} - Metadata attached to buffer {hex(id(buf))}: {count} relations"
- )
- else:
- self.logger.error(
- f"Stream {stream_idx} - Metadata not attached to buffer after adding"
- )
-
- def add_segmentation_metadata(self, buf, mask, x1, y1, x2, y2):
- """
- Adds segmentation mask metadata to the buffer.
- """
- self.logger.info("Adding segmentation mask metadata")
- pass
-
-
-if CAN_REGISTER_ELEMENT:
- GObject.type_register(YOLOTransform)
- __gstelementfactory__ = ("pyml_yolo", Gst.Rank.NONE, YOLOTransform)
-else:
+ def do_forward(self, frames):
+ # Push NMS/confidence knobs to the engine before it runs the model.
+ if self.engine:
+ self.engine.conf = self.confidence
+ self.engine.iou = self.nms_iou
+ self.engine.agnostic_nms = self.agnostic_nms
+ return super().do_forward(frames)
+
+
+# The class is backend-agnostic: under g2g the host imports this module and
+# instantiates YOLOTransform directly, so no GObject registration applies.
+# GStreamer factory registration runs only under the gst backend.
+if CAN_REGISTER_ELEMENT and backend.BACKEND == "gst":
+ __gstelementfactory__ = backend.register_gst_element("pyml_yolo", YOLOTransform)
+elif not CAN_REGISTER_ELEMENT:
GlobalLogger().warning(
"The 'pyml_yolo' element will not be registered because required modules are missing."
)
diff --git a/pyml-launch.py b/pyml-launch.py
new file mode 100755
index 0000000..efc9bb2
--- /dev/null
+++ b/pyml-launch.py
@@ -0,0 +1,34 @@
+#!/usr/bin/env python3
+"""Run a pipeline on whichever backend `PYML_BACKEND` selects, from the checkout.
+
+See `plugins/python/pyml_launch.py` for what the two backends spell differently.
+Re-runs itself under the repo's venv when started from another interpreter: the
+launcher hands its own site directories to GStreamer and to g2g, neither of
+which has a venv of its own.
+"""
+
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent
+VENV = ROOT / ".venv"
+VENV_PYTHON = (
+ VENV / "Scripts" / "python.exe" if os.name == "nt" else VENV / "bin" / "python"
+)
+
+
+def main():
+ # sys.prefix, not the executable path: a venv's `python` is a symlink to the
+ # base interpreter, so resolving it would say we are already inside.
+ if VENV_PYTHON.exists() and Path(sys.prefix) != VENV:
+ return subprocess.call([str(VENV_PYTHON), __file__, *sys.argv[1:]])
+ sys.path.insert(0, str(ROOT / "plugins" / "python"))
+ import pyml_launch
+
+ return pyml_launch.main()
+
+
+if __name__ == "__main__":
+ sys.exit(main())
diff --git a/pyproject.toml b/pyproject.toml
index fc4cb7c..c0d80fa 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -105,7 +105,7 @@ all = [
# Optional
[dependency-groups]
test = [
- "pytest>=7.0",
+ "pytest>=9.0.3",
# Add other test tools
]
dev = [
@@ -113,8 +113,18 @@ dev = [
# Add linting, etc., e.g., "ruff"
]
+[tool.pytest.ini_options]
+markers = [
+ "serial: run this test on its own, it needs the GPU or a capture device to itself",
+]
+
+[project.scripts]
+pyml-launch = "pyml_launch:main"
+
[tool.setuptools]
include-package-data = true
+py-modules = ["pyml_launch"]
+package-dir = {"" = "plugins/python"}
[tool.setuptools.packages.find]
where = ["plugins/python"]
\ No newline at end of file
diff --git a/renovate.json b/renovate.json
new file mode 100644
index 0000000..5d853ec
--- /dev/null
+++ b/renovate.json
@@ -0,0 +1,25 @@
+{
+ "$schema": "https://docs.renovatebot.com/renovate-schema.json",
+ "extends": [
+ "config:recommended",
+ ":enableVulnerabilityAlerts",
+ "security:minimumReleaseAgePypi"
+ ],
+ "timezone": "Etc/UTC",
+ "schedule": ["* 0-3 * * 1"],
+ "osvVulnerabilityAlerts": true,
+ "vulnerabilityAlerts": {
+ "enabled": true,
+ "labels": ["security"]
+ },
+ "lockFileMaintenance": {
+ "enabled": true,
+ "schedule": ["* 0-3 * * 1"]
+ },
+ "packageRules": [
+ {
+ "matchPackageNames": ["torch", "torchvision", "torchaudio"],
+ "groupName": "pytorch"
+ }
+ ]
+}
diff --git a/tests/test_g2g_backend.py b/tests/test_g2g_backend.py
new file mode 100644
index 0000000..850d424
--- /dev/null
+++ b/tests/test_g2g_backend.py
@@ -0,0 +1,943 @@
+"""Unit tests for the g2g element backend (`PYML_BACKEND=g2g`).
+
+These exercise the backend with no GStreamer present: the backend selection, the
+`GObject` / `FlowReturn` shims, the `G2gFrameIO` buffer round-trip, the
+`G2gAnalyticsBackend` mapping onto a flat sink, and an end-to-end `g2g_process`
+on a `VideoTransform` subclass. The g2g host's `FrameBuffer` (a writable
+buffer-protocol view) and `MetaSink` (write-only staging) are stubbed with a
+`bytearray` and a recording object, so no Rust host is needed.
+
+The payload tests at the end drive one text element through both backend
+drivers, so they need GStreamer for the gst half and skip without it.
+"""
+
+import importlib
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+PLUGIN_DIR = Path(__file__).resolve().parent.parent / "plugins" / "python"
+
+# Select the g2g backend before importing `backend`, and make the plugin package
+# importable.
+os.environ["PYML_BACKEND"] = "g2g"
+sys.path.insert(0, str(PLUGIN_DIR))
+
+import backend # noqa: E402
+from backend import (
+ GObject,
+ FlowReturn,
+ frameio,
+ analytics,
+ BaseAggregator,
+ VideoTransform,
+) # noqa: E402
+
+
+class StubMetaSink:
+ """Stand-in for the host's write-only `g2g.MetaSink`.
+
+ Like the real sink, every `add_*` stages one record into a single list and
+ returns its index, which is the handle `relate` takes.
+ """
+
+ def __init__(self):
+ self.staged = []
+ self.relations = []
+ self.class_names = None
+ self.emitted = []
+ self.emitted_durations = []
+
+ def emit(self, payload, duration_ns=None):
+ self.emitted.append(payload)
+ self.emitted_durations.append(duration_ns)
+
+ def set_class_names(self, names):
+ self.class_names = list(names)
+
+ def _stage(self, record):
+ self.staged.append(record)
+ return len(self.staged) - 1
+
+ def add_object(self, label, x, y, w, h, score):
+ return self._stage(("object", label, x, y, w, h, score))
+
+ def add_classification(self, label, score):
+ return self._stage(("classification", label, score))
+
+ def add_blob(self, header, payload):
+ return self._stage(("blob", header, payload))
+
+ def add_tracking(self, object_id):
+ return self._stage(("tracking", object_id))
+
+ def relate(self, src, dst):
+ self.relations.append((src, dst))
+
+ def _of_kind(self, kind):
+ return [record[1:] for record in self.staged if record[0] == kind]
+
+ @property
+ def objects(self):
+ return self._of_kind("object")
+
+ @property
+ def blobs(self):
+ return self._of_kind("blob")
+
+
+#: What the shadowed `gi` raises, so a warning naming it means something wanted it.
+NO_PYGOBJECT = "no pygobject"
+
+#: The families that run hosted on g2g, so none of them may need GStreamer.
+HOSTED_ELEMENT_MODULES = [
+ "base_translate",
+ "base_transcribe",
+ "base_llm",
+ "base_separate",
+ "base_tts",
+ "base_caption",
+ "mariantranslate",
+ "whispertranscribe",
+ "whisperlive",
+ "llm",
+ "demucs",
+ "sepformer",
+ "coquitts",
+ "whisperspeechtts",
+ "caption_phi",
+ "caption_qwen",
+]
+
+
+def test_hosted_elements_import_under_g2g_with_no_pygobject(tmp_path):
+ """A fresh interpreter, with `gi` shadowed by one that refuses to import.
+
+ In-process the check cannot be honest: once any test here has initialised
+ Gst, these imports succeed whether or not the element guards its Gst
+ construction. Shadowing `gi` rather than just checking stderr also makes an
+ element that reaches for GStreamer fail here, instead of quietly depending
+ on a pygobject that happens to be installed.
+ """
+ shadow = tmp_path / "gi"
+ shadow.mkdir()
+ (shadow / "__init__.py").write_text(f'raise ImportError("{NO_PYGOBJECT}")\n')
+
+ result = subprocess.run(
+ [sys.executable, "-c", "import " + ", ".join(HOSTED_ELEMENT_MODULES)],
+ env={
+ **os.environ,
+ "PYML_BACKEND": "g2g",
+ "PYTHONPATH": os.pathsep.join([str(tmp_path), str(PLUGIN_DIR)]),
+ },
+ capture_output=True,
+ text=True,
+ )
+
+ assert result.returncode == 0, result.stderr
+ # The shadow raises with this text, so it appears only if something imported gi.
+ assert (
+ NO_PYGOBJECT not in result.stderr
+ ), "an element reached for GStreamer at import time"
+
+
+def test_every_element_module_imports_under_g2g_without_gst_init():
+ """The same check over the whole plugin directory, so a new element cannot
+ quietly reintroduce the crash.
+
+ Kept apart from the payload check above because a module here may warn about
+ a dependency it cannot find, which is not what this is looking for. A pad
+ template built without `Gst.init` segfaults, so an element that forgets the
+ backend guard takes down whatever process imports it.
+ """
+ pytest.importorskip("gi", reason="the element modules import it at module scope")
+
+ modules = sorted(p.stem for p in PLUGIN_DIR.glob("*.py"))
+
+ result = subprocess.run(
+ [sys.executable, "-c", "import " + ", ".join(modules)],
+ env={**os.environ, "PYML_BACKEND": "g2g", "PYTHONPATH": str(PLUGIN_DIR)},
+ capture_output=True,
+ text=True,
+ )
+
+ assert result.returncode == 0, result.stderr
+
+
+def test_backend_selected_is_g2g():
+ assert backend.BACKEND == "g2g"
+ assert FlowReturn.OK == 0
+ assert FlowReturn.ERROR != FlowReturn.OK
+
+
+def test_gobject_property_shim_decorator_and_attribute_forms():
+ class Widget:
+ @GObject.Property(type=str)
+ def name(self):
+ return getattr(self, "_n", "default")
+
+ @name.setter
+ def name(self, value):
+ self._n = value.upper()
+
+ size = GObject.Property(type=int, default=7, nick="Size", blurb="px")
+
+ w = Widget()
+ assert w.name == "default" # getter default
+ w.name = "yolo"
+ assert w.name == "YOLO" # setter ran
+ assert w.size == 7 # attribute-form default
+ w.size = 42
+ assert w.size == 42
+
+
+def test_a_property_set_from_a_pipeline_line_arrives_as_its_declared_type():
+ """The host cannot know a hosted class's property types, so it forwards the
+ text a pipeline line carries and the declaration here converts it."""
+
+ class Widget:
+ count = GObject.Property(type=int, default=1)
+ enabled = GObject.Property(type=bool, default=False)
+ ratio = GObject.Property(type=float, default=0.0)
+ name = GObject.Property(type=str, default="")
+
+ w = Widget()
+ w.count, w.enabled, w.ratio, w.name = "4", "TRUE", "0.5", "3"
+ assert (w.count, w.enabled, w.ratio, w.name) == (4, True, 0.5, "3")
+
+ w.enabled = "no"
+ assert w.enabled is False
+ w.count = 9 # already typed, set from Python
+ assert w.count == 9
+
+ with pytest.raises(ValueError, match="enabled"):
+ w.enabled = "maybe"
+
+
+def test_an_element_lists_the_properties_it_declares():
+ """The host checks a pipeline against this, so a knob the element has must be
+ in it and a name it does not have must not."""
+ pytest.importorskip("gi", reason="the video transform needs it on the gst backend")
+ pytest.importorskip("torch", reason="the depth engine imports it")
+ from depth import DepthTransform
+
+ declared = DepthTransform().g2g_properties()
+
+ assert "colormap" in declared, "the element's own knob"
+ assert "batch_size" in declared, "one it inherits from the shared tunables"
+ assert "speaker" not in declared, "a detector has no speaker"
+ assert len(declared) == len(set(declared)), "an overridden property listed twice"
+
+
+def test_frameio_read_write_round_trip():
+ width, height = 4, 3
+ buf = bytearray(width * height * 3) # RGB, writable buffer-protocol object
+ sink = StubMetaSink()
+ frameio.bind(sink, "RGB")
+
+ frame, num_sources, fmt = frameio.read_frames(buf, None, width, height)
+ assert num_sources == 1 and fmt == "RGB"
+ assert frame.shape == (height, width, 3)
+
+ frameio.write_frame(buf, np.full((height, width, 3), 200, dtype=np.uint8))
+ assert all(b == 200 for b in buf), "write_frame must update the buffer in place"
+
+ frameio.append_blob(buf, "tag", b"\x01\x02")
+ assert sink.blobs == [("tag", b"\x01\x02")]
+
+
+def test_analytics_maps_onto_flat_sink():
+ sink = StubMetaSink()
+ analytics.bind(sink)
+
+ meta = analytics.add_relation_meta(buf=None)
+ assert meta is not None
+ assert analytics.get_relation_meta(None) is meta
+
+ # String labels intern to stable u32 ids (quark), reused across calls.
+ analytics.add_object(meta, "person", 1, 2, 3, 4, 0.9)
+ analytics.add_object(meta, "person", 5, 6, 7, 8, 0.8)
+ analytics.add_object(meta, "handbag", 0, 0, 1, 1, 0.5)
+
+ assert analytics.relation_length(meta) == 3
+ labels = [o[0] for o in sink.objects]
+ assert labels[0] == labels[1], "same string -> same id"
+ assert labels[2] != labels[0], "different string -> different id"
+ assert sink.objects[0][5] == 0.9
+
+
+def test_class_names_are_published_so_a_consumer_can_name_a_label():
+ sink = StubMetaSink()
+ analytics.bind(sink)
+ meta = analytics.add_relation_meta(buf=None)
+
+ person = analytics.add_object(meta, "person", 1, 2, 3, 4, 0.9)
+ handbag = analytics.add_object(meta, "handbag", 0, 0, 1, 1, 0.5)
+
+ # The table is indexed by the label id staged on the detection, so a
+ # consumer holding only the id can look the name up.
+ names = sink.class_names
+ assert names is not None, "the sink was never sent a name table"
+ assert names[sink.staged[person][1]] == "person"
+ assert names[sink.staged[handbag][1]] == "handbag"
+
+
+def test_class_names_are_resent_for_each_frames_sink():
+ """Each frame gets a fresh sink, so a name interned on an earlier frame has
+ to be published again rather than assumed already known."""
+ first = StubMetaSink()
+ analytics.bind(first)
+ staged = analytics.add_object(
+ analytics.add_relation_meta(None), "person", 1, 2, 3, 4, 0.9
+ )
+ assert first.class_names[first.staged[staged][1]] == "person"
+
+ second = StubMetaSink()
+ analytics.bind(second)
+ staged = analytics.add_object(
+ analytics.add_relation_meta(None), "person", 5, 6, 7, 8, 0.8
+ )
+ assert second.class_names is not None, "the second sink was sent no table"
+ assert second.class_names[second.staged[staged][1]] == "person"
+
+
+def test_tracking_relates_to_its_detection():
+ sink = StubMetaSink()
+ analytics.bind(sink)
+ meta = analytics.add_relation_meta(buf=None)
+
+ od = analytics.add_object(meta, "person", 1, 2, 3, 4, 0.9)
+ track = analytics.add_tracking(meta, 77)
+ assert analytics.relate(meta, od, track) is True
+
+ # Handles are the sink's own staging indices, so the relation names the
+ # detection and the tracking record that were actually staged.
+ assert sink.staged[od][0] == "object"
+ assert sink.staged[track] == ("tracking", 77)
+ assert sink.relations == [(od, track)]
+
+
+def test_video_transform_g2g_process_end_to_end():
+ """A VideoTransform subclass inverts the frame and stages one detection,
+ driven exactly as the host drives it: instance.g2g_process(buf, w, h, fmt, sink)."""
+
+ class InvertAndDetect(VideoTransform):
+ def process_frames(self, frames, num_sources, fmt, target):
+ assert num_sources == 1
+ inverted = 255 - frames
+ frameio.write_frame(target, inverted)
+ meta = analytics.add_relation_meta(target)
+ analytics.add_object(meta, "person", 0, 0, 10, 10, 0.99)
+
+ width, height = 8, 8
+ buf = bytearray([10] * (width * height * 3))
+ sink = StubMetaSink()
+
+ elem = InvertAndDetect()
+ # EngineManager defaults to the pytorch engine, so the first frame would try
+ # to load a model. This test covers the frame plumbing, not inference.
+ elem.engine_name = None
+ ret = elem.g2g_process(buf, width, height, "RGB", sink)
+
+ assert ret is None
+ assert all(b == 245 for b in buf), "frame inverted in place (255 - 10)"
+ assert len(sink.objects) == 1
+ assert sink.objects[0][5] == 0.99
+ assert elem.width == width and elem.height == height
+
+
+def test_aggregator_drives_the_same_hook_a_transform_fills_in():
+ """The N-source case spells the same on both backends: one `process_frames`
+ taking (H, W, C) for a single source and (N, H, W, C) for several."""
+ seen = []
+
+ class Batching(BaseAggregator):
+ def process_frames(self, frames, num_sources, fmt, target):
+ seen.append((frames.shape, num_sources, fmt))
+
+ width, height = 4, 3
+ elem = Batching()
+ elem.engine_name = None
+
+ buffers = [bytearray([n] * (width * height * 3)) for n in (1, 2)]
+ elem.g2g_process_batch(buffers, width, height, "RGB", StubMetaSink())
+ elem.g2g_process_batch(buffers[:1], width, height, "RGB", StubMetaSink())
+
+ assert seen == [
+ ((2, height, width, 3), 2, "RGB"),
+ ((height, width, 3), 1, "RGB"),
+ ]
+
+
+def test_g2g_caption_stages_the_caption_on_the_frame():
+ """The caption family runs on the shared per-frame seam, so it works with no
+ text pad and no GStreamer: the caption is staged as a classification."""
+ from base_caption import BaseCaption
+
+ class FakeCaption(BaseCaption):
+ def forward(self, frames):
+ return "a cat on a mat"
+
+ leaf = FakeCaption()
+ leaf.mgr.engine_name = None # the fake above is the model
+ sink = StubMetaSink()
+ width, height = 4, 3
+
+ leaf.g2g_process(bytearray(width * height * 4), width, height, "RGBA", sink)
+
+ captions = [record for record in sink.staged if record[0] == "classification"]
+ assert len(captions) == 1, "the caption was not staged"
+ assert sink.class_names[captions[0][1]] == "a cat on a mat"
+
+
+def test_a_packed_frame_is_reduced_to_rgb_in_channel_order():
+ from backend.g2g.frameio import as_rgb
+
+ pixel = np.array([[[10, 20, 30, 40]]], dtype=np.uint8)
+
+ assert as_rgb(pixel, "RGBA").tolist() == [[[10, 20, 30]]]
+ assert as_rgb(pixel, "BGRA").tolist() == [[[30, 20, 10]]]
+ assert as_rgb(pixel, "ARGB").tolist() == [[[20, 30, 40]]]
+ assert as_rgb(pixel, "ABGR").tolist() == [[[40, 30, 20]]]
+
+ rgb = np.array([[[10, 20, 30]]], dtype=np.uint8)
+ assert as_rgb(rgb, "RGB") is rgb, "an RGB frame is handed over untouched"
+ assert as_rgb(rgb, "BGR").tolist() == [[[30, 20, 10]]]
+
+
+def test_two_elements_on_their_own_threads_keep_their_own_sinks():
+ """The host runs one thread per element, so a binding made on one thread must
+ not redirect what another thread stages."""
+ import threading
+
+ sinks = {}
+ staged = {}
+ start = threading.Barrier(2)
+ bound = threading.Barrier(2)
+
+ def stage(name, label):
+ sinks[name] = StubMetaSink()
+ start.wait()
+ analytics.bind(sinks[name])
+ frameio.bind(sinks[name], "RGB")
+ # Both threads have bound before either stages anything, so a shared
+ # binding would send both records to whichever bound last.
+ bound.wait()
+ analytics.add_object(analytics.add_relation_meta(None), label, 0, 0, 1, 1, 1.0)
+ frameio.append_blob(bytearray(4), "tag", label.encode())
+ staged[name] = sinks[name].staged
+
+ threads = [
+ threading.Thread(target=stage, args=(name, label))
+ for name, label in (("first", "person"), ("second", "handbag"))
+ ]
+ for thread in threads:
+ thread.start()
+ for thread in threads:
+ thread.join()
+
+ assert len(staged["first"]) == 2, "the first element lost a record to the second"
+ assert len(staged["second"]) == 2
+ assert sinks["first"].blobs == [("tag", b"person")]
+ assert sinks["second"].blobs == [("tag", b"handbag")]
+
+
+class RecordingLogger:
+ """Stands in for the element's logger, keeping what it was told."""
+
+ def __init__(self):
+ self.warnings = []
+
+ def warning(self, message):
+ self.warnings.append(message)
+
+ def info(self, message):
+ pass
+
+ def error(self, message):
+ pass
+
+
+def gst():
+ pytest.importorskip("gi", reason="the gst driver and the pad templates need it")
+ import gi
+
+ gi.require_version("Gst", "1.0")
+ from gi.repository import Gst
+
+ Gst.init(None)
+ return Gst
+
+
+def gst_payload_driver():
+ """The gst backend's driver half, to mix into a leaf.
+
+ This process selected the g2g backend, so a leaf's element base is the g2g
+ one. Adding the driver gives a single instance both backends' entry points,
+ which is what the seam claims: the same element runs under either.
+ """
+ gst()
+ from backend.gst.aggregator import PayloadDriver
+
+ return PayloadDriver
+
+
+CHUNKED_OUTPUTS = [b"one", b"two", b"three"]
+
+
+def chunking_leaf():
+ """A payload element that answers one input buffer with several outputs."""
+ driver = gst_payload_driver()
+
+ class ChunkingLeaf(BaseAggregator, driver):
+ def process_payload(self, payload):
+ return list(CHUNKED_OUTPUTS)
+
+ leaf = ChunkingLeaf()
+ leaf.engine_name = None
+ leaf.logger = RecordingLogger()
+ return leaf
+
+
+def translate_leaf(translate_text):
+ """A real `BaseTranslate` whose model is the given text -> text function.
+
+ Importing the element needs GStreamer even under the g2g backend, since the
+ family still declares its pad templates with Gst types.
+ """
+ driver = gst_payload_driver()
+ from base_translate import BaseTranslate
+
+ class FakeTranslate(BaseTranslate, driver):
+ def do_translate_text(self, text):
+ return translate_text(text)
+
+ leaf = FakeTranslate()
+ leaf.engine_name = None # nothing to load: the fake above is the model
+ return leaf
+
+
+class StubSrcPad:
+ """Stands in for the element's src pad, for the family that pushes straight
+ out of it instead of through the aggregator."""
+
+ def __init__(self, pushed):
+ self._pushed = pushed
+ self.push_count = 0
+
+ def push(self, buf):
+ self._pushed.append(buf)
+ self.push_count += 1
+ return 0
+
+
+def drive_gst_payload(leaf, payload, pts=1000, duration=500):
+ """Run the gst driver over one input buffer, returning what it sent.
+
+ Both send routes are stubbed, the aggregator's `finish_buffer` and the src
+ pad, so `leaf.srcpad.push_count` says which one the element took.
+ """
+ Gst = gst()
+ from backend.gst.aggregator import BaseAggregator as GstBaseAggregator
+
+ pushed = []
+ leaf.finish_buffer = pushed.append
+ leaf.srcpad = StubSrcPad(pushed)
+
+ inbuf = Gst.Buffer.new_allocate(None, len(payload), None)
+ inbuf.fill(0, payload)
+ inbuf.pts = pts
+ inbuf.duration = duration
+
+ ret = GstBaseAggregator.do_process(leaf, inbuf)
+ return ret, pushed
+
+
+def buffer_bytes(buf):
+ Gst = gst()
+ success, map_info = buf.map(Gst.MapFlags.READ)
+ assert success
+ data = bytes(map_info.data)
+ buf.unmap(map_info)
+ return data
+
+
+def test_g2g_payload_driver_emits_the_translated_bytes():
+ leaf = translate_leaf(lambda text: "hello" if text == "hola" else "")
+ sink = StubMetaSink()
+
+ leaf.g2g_process_payload([bytearray(b"hola")], "text/x-raw,format=utf8", sink)
+
+ assert sink.emitted == [b"hello"]
+ assert sink.emitted_durations == [None], "text keeps the input buffer's timing"
+
+
+def test_g2g_payload_driver_emits_nothing_when_the_element_has_no_output():
+ leaf = translate_leaf(lambda text: "")
+ sink = StubMetaSink()
+
+ leaf.g2g_process_payload([bytearray(b"hola")], "text/x-raw,format=utf8", sink)
+
+ assert sink.emitted == [], "an empty result must not reach the host"
+
+
+def test_g2g_payload_driver_emits_every_payload():
+ leaf = chunking_leaf()
+ sink = StubMetaSink()
+
+ leaf.g2g_process_payload([bytearray(b"in")], "audio/x-raw", sink)
+
+ assert sink.emitted == CHUNKED_OUTPUTS
+ assert leaf.logger.warnings == []
+
+
+def test_gst_payload_driver_pushes_the_translated_buffer():
+ Gst = gst()
+ leaf = translate_leaf(lambda text: "hello" if text == "hola" else "")
+
+ ret, pushed = drive_gst_payload(leaf, b"hola")
+
+ assert ret == Gst.FlowReturn.OK
+ assert len(pushed) == 1
+ assert buffer_bytes(pushed[0]) == b"hello"
+ assert pushed[0].pts == 1000 and pushed[0].duration == 500
+
+
+def test_gst_payload_driver_pushes_nothing_when_the_element_has_no_output():
+ Gst = gst()
+ leaf = translate_leaf(lambda text: "")
+
+ ret, pushed = drive_gst_payload(leaf, b"hola")
+
+ assert ret == Gst.FlowReturn.OK
+ assert pushed == []
+
+
+def test_gst_payload_driver_pushes_every_payload_as_its_own_buffer():
+ leaf = chunking_leaf()
+
+ _, pushed = drive_gst_payload(leaf, b"in")
+
+ assert [buffer_bytes(buf) for buf in pushed] == CHUNKED_OUTPUTS
+ assert leaf.logger.warnings == []
+
+
+VAD_CHUNK_SAMPLES = 2400 # 150 ms at 16 kHz, so two silent chunks end a clip
+
+
+class Segment:
+ """One piece of a transcript, as the Whisper models hand it back."""
+
+ def __init__(self, text):
+ self.text = text
+
+
+def stub_vad(monkeypatch):
+ """Stand in for the optional VAD package the transcribe family builds in its
+ constructor. This one calls any non-zero sample speech, which lets a test
+ write silence and speech as buffer contents."""
+ import types
+
+ class SpeechIsNonZero:
+ def chunk_samples(self):
+ return VAD_CHUNK_SAMPLES
+
+ def process_chunk(self, chunk):
+ return 1.0 if any(chunk) else 0.0
+
+ module = types.ModuleType("pysilero_vad")
+ module.SileroVoiceActivityDetector = SpeechIsNonZero
+ monkeypatch.setitem(sys.modules, "pysilero_vad", module)
+
+
+def transcribe_leaf(monkeypatch, transcript):
+ """A real `BaseTranscribe` with a scripted VAD and transcriber."""
+ driver = gst_payload_driver()
+ stub_vad(monkeypatch)
+
+ from base_transcribe import BaseTranscribe
+
+ class FakeTranscribe(BaseTranscribe, driver):
+ def do_transcribe(self, audio_data, task):
+ return [Segment(word) for word in transcript.split()]
+
+ leaf = FakeTranscribe()
+ leaf.engine_name = None
+ return leaf
+
+
+def speech(chunks=1):
+ return np.full(VAD_CHUNK_SAMPLES * chunks, 1000, dtype=np.int16).tobytes()
+
+
+def silence(chunks=1):
+ return np.zeros(VAD_CHUNK_SAMPLES * chunks, dtype=np.int16).tobytes()
+
+
+def separate_leaf(sample_rate=4):
+ """A real `BaseSeparate` whose model returns the audio it was given."""
+ driver = gst_payload_driver()
+ from base_separate import BaseSeparate
+
+ class PassThroughSeparate(BaseSeparate, driver):
+ SAMPLE_RATE = sample_rate
+
+ def do_separate(self, audio_data):
+ return audio_data
+
+ leaf = PassThroughSeparate()
+ leaf.engine_name = None
+ leaf.streaming = True # a one second chunk, so four samples at this rate
+ return leaf
+
+
+class FakeEngine:
+ def do_generate(self, text, system_prompt=None):
+ return "answered " + text
+
+
+def llm_leaf():
+ """A real `BaseLlm` whose engine is a canned generator."""
+ driver = gst_payload_driver()
+ from base_llm import BaseLlm
+
+ class FakeLlm(BaseLlm, driver):
+ @property
+ def engine(self):
+ return self._fake_engine
+
+ def get_tokenizer(self):
+ return "tokenizer"
+
+ def get_model(self):
+ return "model"
+
+ leaf = FakeLlm()
+ leaf._fake_engine = FakeEngine()
+ leaf.engine_name = None
+ return leaf
+
+
+PAYLOAD_FAMILY_BASES = [
+ ("base_translate", "BaseTranslate"),
+ ("base_transcribe", "BaseTranscribe"),
+ ("base_llm", "BaseLlm"),
+ ("base_separate", "BaseSeparate"),
+ ("base_tts", "BaseTts"),
+]
+
+
+def declared_properties(cls):
+ """Every property the class declares or inherits, mapped to its owner."""
+ return {
+ name: klass.__name__
+ for klass in reversed(cls.__mro__)
+ for name, value in vars(klass).items()
+ if isinstance(value, GObject.Property)
+ }
+
+
+@pytest.mark.parametrize("module_name,class_name", PAYLOAD_FAMILY_BASES)
+def test_every_declared_property_reads_before_anything_sets_it(
+ monkeypatch, module_name, class_name
+):
+ """A freshly built element has to answer every property it declares.
+
+ Nothing applies a declared default: GObject keeps it in the pspec and never
+ routes it through the setter, so the constructor is the only thing that can
+ create the backing attribute a custom getter reads.
+ """
+ gst()
+ stub_vad(monkeypatch)
+ element = getattr(importlib.import_module(module_name), class_name)()
+
+ unreadable = []
+ for name, owner in sorted(declared_properties(type(element)).items()):
+ try:
+ getattr(element, name)
+ except AttributeError as exception:
+ unreadable.append(f"{owner}.{name}: {exception}")
+
+ assert unreadable == []
+
+
+def test_g2g_transcribe_emits_only_once_the_clip_ends(monkeypatch):
+ leaf = transcribe_leaf(monkeypatch, "hello world")
+ sink = StubMetaSink()
+ caps = "audio/x-raw,format=S16LE,rate=16000,channels=1"
+
+ leaf.g2g_process_payload([bytearray(speech())], caps, sink)
+ assert sink.emitted == [], "speech is still being accumulated"
+
+ leaf.g2g_process_payload([bytearray(silence(3))], caps, sink)
+ assert sink.emitted == [b"hello world"], "the silence should end the clip"
+
+
+def test_gst_transcribe_pushes_only_once_the_clip_ends(monkeypatch):
+ leaf = transcribe_leaf(monkeypatch, "hello world")
+
+ _, pushed = drive_gst_payload(leaf, speech())
+ assert pushed == [], "speech is still being accumulated"
+
+ _, pushed = drive_gst_payload(leaf, silence(3))
+ assert [buffer_bytes(buf) for buf in pushed] == [b"hello world"]
+ assert pushed[0].pts == 1000 and pushed[0].duration == 500
+
+
+def test_gst_transcribe_pushes_nothing_for_a_buffer_below_one_vad_chunk(monkeypatch):
+ leaf = transcribe_leaf(monkeypatch, "hello world")
+
+ _, pushed = drive_gst_payload(leaf, np.zeros(8, dtype=np.int16).tobytes())
+
+ assert pushed == []
+
+
+def test_gst_separate_pushes_one_buffer_per_whole_chunk():
+ leaf = separate_leaf()
+ samples = np.arange(1, 11, dtype=np.int16) # ten samples, so two chunks of four
+
+ _, pushed = drive_gst_payload(leaf, samples.tobytes())
+
+ assert [buffer_bytes(buf) for buf in pushed] == [
+ samples[:4].tobytes(),
+ samples[4:8].tobytes(),
+ ]
+ assert len(leaf.clip_buffer) == 2, "the remainder waits for the next buffer"
+
+
+def test_g2g_separate_emits_one_buffer_per_whole_chunk():
+ leaf = separate_leaf()
+ leaf.logger = RecordingLogger()
+ sink = StubMetaSink()
+ samples = np.arange(1, 11, dtype=np.int16)
+
+ leaf.g2g_process_payload([bytearray(samples.tobytes())], "audio/x-raw", sink)
+
+ assert sink.emitted == [samples[:4].tobytes(), samples[4:8].tobytes()]
+ assert leaf.logger.warnings == []
+
+
+def test_g2g_llm_emits_the_generated_text():
+ leaf = llm_leaf()
+ sink = StubMetaSink()
+
+ leaf.g2g_process_payload([bytearray(b"question")], "text/x-raw,format=utf8", sink)
+
+ assert sink.emitted == [b"answered question"]
+
+
+def test_gst_llm_pushes_out_of_the_src_pad_rather_than_the_aggregator():
+ Gst = gst()
+ leaf = llm_leaf()
+
+ ret, pushed = drive_gst_payload(leaf, b"question")
+
+ assert ret == Gst.FlowReturn.OK
+ assert [buffer_bytes(buf) for buf in pushed] == [b"answered question"]
+ assert leaf.srcpad.push_count == 1, "this family has never used finish_buffer"
+ assert pushed[0].pts == 1000 and pushed[0].duration == 500
+
+
+TTS_SAMPLE_RATE = 22050
+TTS_SAMPLES = 8000
+
+
+def tts_leaf():
+ """A real `BaseTts` whose voice is a ramp of the right sample count."""
+ driver = gst_payload_driver()
+ from base_tts import BaseTts
+
+ class FakeTts(BaseTts, driver):
+ def do_load_model(self):
+ pass
+
+ def do_generate_speech(self, transcript):
+ return np.linspace(-0.5, 0.5, TTS_SAMPLES, dtype=np.float32)
+
+ def do_get_sample_rate(self):
+ return TTS_SAMPLE_RATE
+
+ leaf = FakeTts()
+ leaf.engine_name = None
+ return leaf
+
+
+def expected_tts_duration_ns():
+ return int(TTS_SAMPLES / TTS_SAMPLE_RATE * 1_000_000_000)
+
+
+def test_gst_tts_stamps_its_own_duration_and_no_presentation_time():
+ Gst = gst()
+ leaf = tts_leaf()
+
+ ret, pushed = drive_gst_payload(leaf, b"speak this")
+
+ assert ret == Gst.FlowReturn.OK
+ assert len(pushed) == 1
+ assert len(buffer_bytes(pushed[0])) == TTS_SAMPLES * 2, "S16LE, one channel"
+ assert (
+ pushed[0].pts == Gst.CLOCK_TIME_NONE
+ ), "the text buffer's pts is not the audio's"
+ assert pushed[0].dts == Gst.CLOCK_TIME_NONE
+ assert pushed[0].duration == expected_tts_duration_ns()
+ assert leaf.srcpad.push_count == 1, "this family has never used finish_buffer"
+
+
+def test_g2g_tts_emits_the_duration_the_audio_actually_runs_for():
+ leaf = tts_leaf()
+ sink = StubMetaSink()
+
+ leaf.g2g_process_payload([bytearray(b"speak this")], "text/x-raw,format=utf8", sink)
+
+ assert len(sink.emitted) == 1
+ assert len(sink.emitted[0]) == TTS_SAMPLES * 2
+ assert sink.emitted_durations == [expected_tts_duration_ns()]
+
+
+def test_tts_streaming_speaks_each_chunk_of_text_separately():
+ """Streaming splits the text into 20 character chunks, one payload each."""
+ gst()
+ leaf = tts_leaf()
+ leaf.streaming = True
+
+ _, pushed = drive_gst_payload(leaf, b"x" * 45)
+
+ assert len(pushed) == 3, "45 characters is three chunks"
+ assert {buf.duration for buf in pushed} == {expected_tts_duration_ns()}
+
+
+def test_g2g_tts_streaming_emits_each_chunk_of_text_separately():
+ leaf = tts_leaf()
+ leaf.streaming = True
+ sink = StubMetaSink()
+
+ leaf.g2g_process_payload([bytearray(b"x" * 45)], "text/x-raw,format=utf8", sink)
+
+ assert len(sink.emitted) == 3, "45 characters is three chunks"
+ assert sink.emitted_durations == [expected_tts_duration_ns()] * 3
+
+
+def test_gst_driver_keeps_the_input_timing_including_dts():
+ """base_llm timestamped its output with the input's dts; the driver, which
+ now builds that buffer, has to keep doing it."""
+ Gst = gst()
+ leaf = llm_leaf()
+
+ pushed = []
+ leaf.finish_buffer = pushed.append
+ leaf.srcpad = StubSrcPad(pushed)
+ inbuf = Gst.Buffer.new_allocate(None, len(b"question"), None)
+ inbuf.fill(0, b"question")
+ inbuf.pts = 90
+ inbuf.dts = 80
+ inbuf.duration = 70
+
+ from backend.gst.aggregator import BaseAggregator as GstBaseAggregator
+
+ GstBaseAggregator.do_process(leaf, inbuf)
+
+ assert (pushed[0].pts, pushed[0].dts, pushed[0].duration) == (90, 80, 70)
diff --git a/tests/test_pipelines.py b/tests/test_pipelines.py
index ef13558..29e0be7 100644
--- a/tests/test_pipelines.py
+++ b/tests/test_pipelines.py
@@ -1,5 +1,6 @@
import subprocess
import os
+import signal
import re
import pytest
from pathlib import Path
@@ -12,8 +13,28 @@
BASE_DIR = Path(__file__).resolve().parent.parent
LOG_DIR = BASE_DIR / "tests" / "logs"
-# Check if gst-launch-1.0 is available
-if not shutil.which("gst-launch-1.0"):
+# Seconds to let a pipeline run. Overridable because a local LLM generating a
+# few hundred tokens takes minutes, not seconds.
+PIPELINE_TIMEOUT = int(os.environ.get("PIPELINE_TIMEOUT", "30"))
+
+# Only these mean the pipeline broke. Warnings are not fatal: plugins unrelated
+# to the pipeline warn during setup and would fail a run that went fine.
+FATAL_LOG_PATTERNS = (
+ re.compile(r"^ERROR:.*", re.MULTILINE),
+ re.compile(r"^WARNING: erroneous pipeline.*", re.MULTILINE),
+ re.compile(r"^\S+ +\S+ +\S+ +ERROR +python .*", re.MULTILINE),
+ # g2g's own failures: the launcher's parse / run errors and the rewrite's.
+ re.compile(r"^(?:parse|pipeline) error:.*", re.MULTILINE),
+ re.compile(r"^pyml-launch: (?:no|unknown) .*", re.MULTILINE),
+)
+
+BACKEND = os.environ.get("PYML_BACKEND", "gst").lower()
+
+# The launcher the README examples name. Runs from a tmp dir, so it is spelled
+# absolute here.
+LAUNCHER = f"python {BASE_DIR / 'pyml-launch.py'}"
+
+if BACKEND == "gst" and not shutil.which("gst-launch-1.0"):
raise RuntimeError("gst-launch-1.0 not found in PATH. Please install GStreamer.")
@@ -26,9 +47,9 @@ def get_pipelines_from_readme():
with open(readme_path, "r") as f:
content = f.read()
- # Match gst-launch-1.0 commands, accounting for Markdown backticks
+ # Match pyml-launch commands, accounting for Markdown backticks
pipeline_pattern = (
- r"(?:`)?\s*(GST_DEBUG=\d+\s+gst-launch-1\.0\s+.*?)(?:`)?(?=\n\n|\n\s*\n|$)"
+ r"(?:`)?\s*(python pyml-launch\.py\s+.*?)(?:`)?(?=\n\n|\n\s*\n|$)"
)
pipelines = re.findall(pipeline_pattern, content, re.DOTALL)
@@ -37,13 +58,16 @@ def get_pipelines_from_readme():
pipeline = pipeline.strip().strip("`")
print(f"Raw pipeline after stripping: {pipeline}")
- if not pipeline.startswith(("GST_DEBUG=", "gst-launch-1.0")):
+ if not pipeline.startswith("python pyml-launch.py"):
print(f"Skipping invalid pipeline: {pipeline}")
continue
+ pipeline = pipeline.replace("python pyml-launch.py", LAUNCHER, 1)
parts = pipeline.split("!")
- modified = False
+ # A `filesrc` run has no equivalent cap: these mp4s carry `moov` at the
+ # end, so bounding the source by bytes leaves the decoder with no index.
+ # Those pipelines run until PIPELINE_TIMEOUT instead.
for i, part in enumerate(parts):
part_clean = part.strip()
if "videotestsrc" in part_clean:
@@ -52,20 +76,8 @@ def get_pipelines_from_readme():
parts[i] = f"{part_clean} num-buffers=100"
else:
parts[i] = re.sub(r"num-buffers=\d+", "num-buffers=100", part_clean)
- modified = True
break
- if not modified:
- for i, part in enumerate(parts):
- part_clean = part.strip()
- if "filesrc" in part_clean:
- parts.insert(i + 1, "queue max-size-buffers=100 leaky=upstream")
- modified = True
- break
-
- if not modified:
- print(f"Warning: No filesrc or videotestsrc found in pipeline: {pipeline}")
-
modified_pipeline = " ! ".join(parts).strip()
print(f"Modified pipeline: {modified_pipeline}")
modified_pipelines.append(modified_pipeline)
@@ -75,14 +87,51 @@ def get_pipelines_from_readme():
PIPELINES = get_pipelines_from_readme()
+def end_process_group(process):
+ """Stop the pipeline and everything it started.
+
+ `shell=True` makes the shell the direct child, so signalling the process
+ alone leaves the launcher and its window behind.
+ """
+ try:
+ group = os.getpgid(process.pid)
+ except ProcessLookupError:
+ return
+ os.killpg(group, signal.SIGTERM)
+ try:
+ process.wait(timeout=5)
+ except subprocess.TimeoutExpired:
+ os.killpg(group, signal.SIGKILL)
+
+
+def absolutize_project_inputs(pipeline):
+ """Point a pipeline's relative input paths at the project directory.
+
+ Lets the pipeline run from the test's tmp dir so its output lands there.
+ Only values that already name a file are rewritten, so output paths and
+ caps strings are left alone.
+ """
+
+ def rewrite(match):
+ key, value = match.group(1), match.group(2)
+ candidate = BASE_DIR / value.strip('"')
+ return f"{key}={candidate}" if candidate.is_file() else match.group(0)
+
+ return re.sub(r"([\w-]+)=([^\s!]+)", rewrite, pipeline)
+
+
@pytest.mark.serial
@pytest.mark.parametrize("pipeline", PIPELINES, ids=lambda p: p)
def test_pipeline(pipeline, tmp_path):
"""
- Test a GStreamer pipeline for 100 frames, checking for errors, with latency tracing.
+ Run a README pipeline and check its log for errors.
+
+ A pipeline still running at `PIPELINE_TIMEOUT` passes: only `videotestsrc`
+ takes a frame cap, so a file-backed one runs as long as its media lasts.
"""
LOG_DIR.mkdir(parents=True, exist_ok=True)
os.sync()
+ pipeline = absolutize_project_inputs(pipeline)
unique_id = uuid.uuid4().hex[:8]
log_file = LOG_DIR / f"test_{unique_id}.log"
@@ -124,30 +173,37 @@ def test_pipeline(pipeline, tmp_path):
# Set up environment with latency tracer
env = os.environ.copy()
env["GST_TRACERS"] = "latency"
+ # Colour escapes land in the log file and break matching on the level field.
+ env["GST_DEBUG_NO_COLOR"] = "1"
# Run the pipeline
+ ran_to_the_cap = False
try:
with open(log_file, "w") as log:
+ # Own process group: the shell is not the pipeline, it is the
+ # launcher's parent, so killing the group is what stops the run.
+ # Terminating the shell alone leaves the launcher holding a window
+ # and the GPU until the machine is rebooted.
process = subprocess.Popen(
pipeline,
shell=True,
stdout=log,
stderr=subprocess.STDOUT,
- cwd=BASE_DIR,
+ cwd=tmp_path,
env=env,
+ start_new_session=True,
)
- process.wait(timeout=30)
+ process.wait(timeout=PIPELINE_TIMEOUT)
return_code = process.returncode
except subprocess.TimeoutExpired:
- process.terminate()
- try:
- process.wait(timeout=2)
- except subprocess.TimeoutExpired:
- process.kill()
- pytest.fail(
- f"Pipeline timed out after 30s. Full pipeline: {pipeline}. See {log_file}"
- )
+ # Still running at the cap, which is what a healthy uncapped pipeline
+ # does: the media outlasts any timeout worth waiting. The log below says
+ # whether it was working, so the run is judged on that, not on exiting.
+ end_process_group(process)
+ ran_to_the_cap = True
+ return_code = None
except Exception as e:
+ end_process_group(process)
pytest.fail(
f"Failed to execute pipeline: {e}. Full pipeline: {pipeline}. See {log_file}"
)
@@ -157,18 +213,21 @@ def test_pipeline(pipeline, tmp_path):
pytest.fail(f"Log file {log_file} was not created. Full pipeline: {pipeline}")
with open(log_file, "r") as log:
log_content = log.read()
- error_lines = [
- line
- for line in log_content.splitlines()
- if "ERROR" in line or "WARN" in line
- ]
- if error_lines:
- pytest.fail(
- f"Errors/Warnings found in pipeline:\n{''.join(error_lines)}\nFull pipeline: {pipeline}\nSee {log_file}"
- )
+ failures = [m.group(0) for p in FATAL_LOG_PATTERNS for m in p.finditer(log_content)]
+ if failures:
+ reported = "\n".join(failures)
+ pytest.fail(
+ f"Errors found in pipeline:\n{reported}\nFull pipeline: {pipeline}\nSee {log_file}"
+ )
+
+ # Without this a pipeline that never left PAUSED passes on an empty log.
+ if "Setting pipeline to PLAYING" not in log_content:
+ pytest.fail(
+ f"Pipeline never reached PLAYING. Full pipeline: {pipeline}. See {log_file}"
+ )
# Check exit code
- if return_code != 0:
+ if not ran_to_the_cap and return_code != 0:
if (
"End-Of-Stream" not in log_content
and "reached end of stream" not in log_content
@@ -177,13 +236,14 @@ def test_pipeline(pipeline, tmp_path):
f"Pipeline failed with exit code {return_code}. Full pipeline: {pipeline}. See {log_file}"
)
- print(f"Pipeline processed 100 frames successfully: {pipeline}")
+ ending = f"ran the full {PIPELINE_TIMEOUT}s" if ran_to_the_cap else "ran to the end"
+ print(f"Pipeline {ending} with no errors: {pipeline}")
def test_pipelines_found():
"""Ensure at least one pipeline was found in README."""
if not PIPELINES:
- pytest.fail("No gst-launch-1.0 pipelines found in README.md")
+ pytest.fail("No pyml-launch pipelines found in README.md")
print(f"Found {len(PIPELINES)} pipelines to test")
diff --git a/tests/test_pyml_launch.py b/tests/test_pyml_launch.py
new file mode 100644
index 0000000..39d5453
--- /dev/null
+++ b/tests/test_pyml_launch.py
@@ -0,0 +1,365 @@
+"""Unit tests for `pyml-launch`'s gst -> g2g pipeline rewrite.
+
+These run the real `element_shells` scan over the real plugin directory, so a
+plugin that stops declaring `register_gst_element` fails a test rather than
+silently dropping out of the map.
+"""
+
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "plugins" / "python"))
+
+import pyml_launch # noqa: E402
+
+
+@pytest.fixture(scope="module")
+def shells():
+ return pyml_launch.element_shells()
+
+
+def test_element_shells_finds_the_hosted_elements(shells):
+ assert shells["pyml_yolo"] == pyml_launch.ElementShell("yolo", "YOLOTransform")
+ assert shells["pyml_objectdetector"].module == "objectdetector"
+ assert shells["pyml_overlay"] == pyml_launch.ElementShell("overlay", "Overlay")
+
+
+def test_an_element_several_chains_feed_becomes_the_batching_host(shells):
+ pipeline = (
+ "videotestsrc ! pyml_streammux name=mux "
+ "videotestsrc ! mux. videotestsrc ! mux. "
+ "mux. ! fakesink"
+ ).split()
+ rewritten = pyml_launch.rewrite_for_g2g(pipeline, shells)
+ assert "pyaggregator" in rewritten
+ assert "pyelement" not in rewritten
+
+
+def test_a_single_input_element_stays_on_the_one_in_host(shells):
+ # Whisper derives from GstBase.Aggregator but takes one chain, which g2g
+ # hosts as a transform.
+ assert (
+ pyml_launch.rewrite_segment("pyml_whispertranscribe language=ko", shells)[0]
+ == "pyelement"
+ )
+
+
+def test_element_shells_resolves_a_name_given_as_a_class_constant(shells):
+ # kafkasink registers under `KafkaSink.GST_PLUGIN_NAME`, not a literal.
+ assert shells["pyml_kafkasink"] == pyml_launch.ElementShell(
+ "kafkasink", "KafkaSink"
+ )
+
+
+def test_hosted_element_becomes_pyelement_keeping_its_properties(shells):
+ assert pyml_launch.rewrite_segment(
+ "pyml_yolo model-name=yolo11m device=cuda:0", shells
+ ) == [
+ "pyelement",
+ "module=yolo",
+ "class=YOLOTransform",
+ "model-name=yolo11m",
+ "device=cuda:0",
+ ]
+
+
+def test_overlay_becomes_the_native_element(shells):
+ assert pyml_launch.rewrite_segment("pyml_overlay", shells) == ["analyticsoverlay"]
+
+
+def test_overlay_properties_carry_over_under_the_native_name(shells):
+ assert pyml_launch.rewrite_segment("pyml_overlay tracking=True", shells) == [
+ "analyticsoverlay",
+ "show-track=True",
+ ]
+
+
+def test_an_overlay_property_with_no_counterpart_is_refused(shells):
+ with pytest.raises(SystemExit, match="meta-path"):
+ pyml_launch.rewrite_segment(
+ "pyml_overlay meta-path=data/sample_metadata.json", shells
+ )
+
+
+def test_unknown_pyml_element_is_refused(shells):
+ with pytest.raises(SystemExit, match="pyml_nonesuch"):
+ pyml_launch.rewrite_segment("pyml_nonesuch", shells)
+
+
+def test_shared_elements_pass_through_untouched(shells):
+ for segment in ("filesrc location=data/people.mp4", "decodebin", "videoscale"):
+ assert pyml_launch.rewrite_segment(segment, shells) == segment.split()
+
+
+def test_raw_video_caps_gain_a_format_only_when_they_lack_one(shells):
+ assert pyml_launch.rewrite_segment("video/x-raw,width=640,height=480", shells) == [
+ "video/x-raw,width=640,height=480,format=RGBA"
+ ]
+ pinned = "video/x-raw,format=NV12,width=640,height=480"
+ assert pyml_launch.rewrite_segment(pinned, shells) == [pinned]
+
+
+def test_caps_written_with_spaces_become_one_token(shells):
+ assert pyml_launch.rewrite_segment(
+ ["video/x-raw,", "width=320,", "height=240"], shells
+ ) == ["video/x-raw,width=320,height=240,format=RGBA"]
+
+
+def test_a_property_value_with_spaces_keeps_them(shells):
+ assert pyml_launch.rewrite_segment(
+ ["pyml_clip", "labels=person, bicycle, car", "top-k=3"], shells
+ ) == [
+ "pyelement",
+ "module=clip",
+ "class=CLIPTransform",
+ 'labels="person, bicycle, car"',
+ "top-k=3",
+ ]
+
+
+def test_a_property_value_keeps_the_quotes_inside_it(shells):
+ assert (
+ pyml_launch.rewrite_segment(["pyml_alert", 'rules={"class":"person"}'], shells)[
+ -1
+ ]
+ == r"rules={\"class\":\"person\"}"
+ )
+
+
+def test_a_sink_drops_the_clock_properties_g2g_has_no_counterpart_for(shells):
+ assert pyml_launch.rewrite_segment("fakesink async=0 sync=0", shells) == [
+ "fakesink"
+ ]
+
+
+def test_a_sink_asked_to_wait_on_the_clock_is_refused(shells):
+ with pytest.raises(SystemExit, match="clocksync"):
+ pyml_launch.rewrite_segment("autovideosink sync=true", shells)
+
+
+def test_a_hosted_element_named_like_a_sink_keeps_its_own_properties(shells):
+ # `sync` on a g2g sink names a clock wait that does not exist there, but on
+ # `pyml_kafkasink` it is the Kafka producer's own knob.
+ assert pyml_launch.rewrite_segment("pyml_kafkasink sync=true", shells) == [
+ "pyelement",
+ "module=kafkasink",
+ "class=KafkaSink",
+ "sync=true",
+ ]
+
+
+def test_textoverlay_drops_the_wait_it_never_does(shells):
+ assert pyml_launch.rewrite_segment(
+ "textoverlay name=overlay wait-text=false", shells
+ ) == ["textoverlay", "name=overlay"]
+
+
+def test_whole_readme_pipeline_rewrites(shells):
+ pipeline = (
+ "filesrc location=data/soccer_tracking.mp4 ! decodebin ! videoconvert "
+ "! videoscale ! video/x-raw,width=640,height=480 "
+ "! pyml_yolo model-name=yolo11m device=cuda:0 track=True "
+ "! pyml_overlay ! videoconvert ! autovideosink"
+ )
+ assert (
+ pyml_launch.rewrite_for_g2g(pipeline.split(), shells)
+ == (
+ "filesrc location=data/soccer_tracking.mp4 ! decodebin ! videoconvert "
+ "! videoscale ! video/x-raw,width=640,height=480,format=RGBA "
+ "! pyelement module=yolo class=YOLOTransform model-name=yolo11m "
+ "device=cuda:0 track=True "
+ "! analyticsoverlay ! videoconvert ! autovideosink"
+ ).split()
+ )
+
+
+def test_hosted_element_takes_the_format_the_caps_ahead_of_it_pin(shells):
+ pipeline = (
+ "videoconvert ! video/x-raw,format=RGB,width=640,height=640 "
+ "! pyml_inference engine-name=onnx"
+ ).split()
+ assert pyml_launch.rewrite_for_g2g(pipeline, shells)[-4:] == [
+ "module=inference",
+ "class=GenericInferenceTransform",
+ "format=RGB",
+ "engine-name=onnx",
+ ]
+
+
+def test_the_batching_host_takes_the_caps_of_an_audio_in_text_out_element(shells):
+ assert pyml_launch.rewrite_segment(
+ "pyml_whispertranscribe", shells, host=pyml_launch.PY_AGGREGATOR
+ ) == [
+ "pyaggregator",
+ "module=whispertranscribe",
+ "class=WhisperTranscribe",
+ "input-caps=audio/x-raw,format=S16LE,layout=interleaved,rate=16000,channels=1",
+ "output-caps=text/x-raw,format=utf8",
+ ]
+
+
+def test_the_batching_host_takes_the_caps_of_a_text_in_audio_out_element(shells):
+ assert pyml_launch.rewrite_segment(
+ "pyml_coquitts", shells, host=pyml_launch.PY_AGGREGATOR
+ ) == [
+ "pyaggregator",
+ "module=coquitts",
+ "class=CoquiTTS",
+ "input-caps=text/x-raw,format=utf8",
+ "output-caps=audio/x-raw,format=S16LE,layout=interleaved,rate=22050,channels=1",
+ ]
+
+
+def test_the_one_in_host_takes_the_caps_of_an_audio_in_text_out_element(shells):
+ pipeline = (
+ "filesrc location=data/audio_sample.wav ! decodebin ! audioconvert "
+ "! pyml_whispertranscribe language=ko ! fakesink"
+ ).split()
+ assert (
+ pyml_launch.rewrite_for_g2g(pipeline, shells)
+ == (
+ "filesrc location=data/audio_sample.wav ! decodebin ! audioconvert "
+ "! pyelement module=whispertranscribe class=WhisperTranscribe "
+ "input-caps=audio/x-raw,format=S16LE,layout=interleaved,rate=16000,channels=1 "
+ "output-caps=text/x-raw,format=utf8 language=ko "
+ "! fakesink"
+ ).split()
+ )
+
+
+def test_the_one_in_host_takes_the_caps_of_a_text_in_audio_out_element(shells):
+ pipeline = "filesrc location=data/lines.txt ! pyml_coquitts ! fakesink".split()
+ assert (
+ pyml_launch.rewrite_for_g2g(pipeline, shells)
+ == (
+ "filesrc location=data/lines.txt "
+ "! pyelement module=coquitts class=CoquiTTS "
+ "input-caps=text/x-raw,format=utf8 "
+ "output-caps=audio/x-raw,format=S16LE,layout=interleaved,rate=22050,channels=1 "
+ "! fakesink"
+ ).split()
+ )
+
+
+def test_a_leaf_that_restates_a_rate_wins_over_the_base_it_inherits_from(shells):
+ # Demucs runs at the base class's 44100, Sepformer restates both pads at 8000.
+ assert shells["pyml_demucs"].caps == (
+ (
+ "input-caps",
+ "audio/x-raw,format=S16LE,layout=interleaved,rate=44100,channels=1",
+ ),
+ (
+ "output-caps",
+ "audio/x-raw,format=S16LE,layout=interleaved,rate=44100,channels=1",
+ ),
+ )
+ assert shells["pyml_sepformer"].caps == (
+ (
+ "input-caps",
+ "audio/x-raw,format=S16LE,layout=interleaved,rate=8000,channels=1",
+ ),
+ (
+ "output-caps",
+ "audio/x-raw,format=S16LE,layout=interleaved,rate=8000,channels=1",
+ ),
+ )
+
+
+def test_a_leaf_inherits_the_pad_it_does_not_restate(shells):
+ # WhisperLive transcribes into speech, so only its src pad differs.
+ assert shells["pyml_whisperlive"].caps == (
+ (
+ "input-caps",
+ "audio/x-raw,format=S16LE,layout=interleaved,rate=16000,channels=1",
+ ),
+ (
+ "output-caps",
+ "audio/x-raw,format=S16LE,layout=interleaved,rate=24000,channels=1",
+ ),
+ )
+
+
+def test_an_element_declaring_no_caps_gets_no_caps_properties(shells):
+ pipeline = (
+ "videotestsrc ! pyml_streammux name=mux "
+ "videotestsrc ! mux. videotestsrc ! mux. "
+ "mux. ! fakesink"
+ ).split()
+ rewritten = pyml_launch.rewrite_for_g2g(pipeline, shells)
+ assert "pyaggregator" in rewritten
+ assert not [token for token in rewritten if token.startswith(("input-", "output-"))]
+
+
+def test_the_caps_survive_the_rewrite_of_a_whole_pipeline(shells):
+ pipeline = (
+ "filesrc location=data/audio_sample.wav ! decodebin ! audioconvert "
+ "! pyml_whispertranscribe name=stt language=ko "
+ "audiotestsrc ! stt. audiotestsrc ! stt. stt. ! fakesink"
+ ).split()
+ assert (
+ pyml_launch.rewrite_for_g2g(pipeline, shells)
+ == (
+ "filesrc location=data/audio_sample.wav ! decodebin ! audioconvert "
+ "! pyaggregator module=whispertranscribe class=WhisperTranscribe "
+ "input-caps=audio/x-raw,format=S16LE,layout=interleaved,rate=16000,channels=1 "
+ "output-caps=text/x-raw,format=utf8 name=stt language=ko "
+ "audiotestsrc ! stt. audiotestsrc ! stt. stt. ! fakesink"
+ ).split()
+ )
+
+
+def g2g_pipeline(*segments):
+ """Run a pipeline on the g2g launcher, returning what it printed.
+
+ Skips where there is no build to check against, so the drift check runs for
+ whoever has both repos and never blocks whoever has one.
+ """
+ binary = pyml_launch.g2g_binary()
+ if not binary:
+ pytest.skip("needs a g2g-launch-py build to check the names against")
+ return subprocess.run(
+ [binary, *segments],
+ capture_output=True,
+ text=True,
+ timeout=120,
+ )
+
+
+#: Enough of a pipeline to negotiate raw video into the element under test.
+G2G_VIDEO_SOURCE = (
+ "videotestsrc",
+ "num-buffers=1",
+ "!",
+ "videoconvert",
+ "!",
+ f"video/x-raw,format={pyml_launch.G2G_RAW_VIDEO_FORMAT}",
+ "!",
+)
+
+
+def test_the_native_element_and_the_names_it_renames_to_still_exist():
+ """`NATIVE_EQUIVALENTS` and `NATIVE_PROPERTIES` copy names out of glass2glass,
+ and `G2G_RAW_VIDEO_FORMAT` states what that element negotiates. Nothing here
+ notices when any of the three changes there, so run one and see."""
+ for element, native in pyml_launch.NATIVE_EQUIVALENTS.items():
+ renames = pyml_launch.NATIVE_PROPERTIES.get(element, {})
+ properties = [f"{name}=true" for name in renames.values()]
+ result = g2g_pipeline(*G2G_VIDEO_SOURCE, native, *properties, "!", "fakesink")
+ assert result.returncode == 0, (
+ f"{element} rewrites to `{native} {' '.join(properties)}`, which g2g "
+ f"no longer runs:\n{result.stdout}{result.stderr}"
+ )
+
+
+def test_an_element_read_from_twice_is_not_a_muxer(shells):
+ # Two chains start at `cap.`, reading two source pads, which is a fan-out.
+ pipeline = (
+ "videotestsrc ! pyml_caption_qwen name=cap "
+ "cap.src ! fakesink cap.text_src ! fakesink"
+ ).split()
+ rewritten = pyml_launch.rewrite_for_g2g(pipeline, shells)
+ assert "pyelement" in rewritten
+ assert "pyaggregator" not in rewritten
diff --git a/uv.lock b/uv.lock
index 4050a24..171a7d4 100644
--- a/uv.lock
+++ b/uv.lock
@@ -10,6 +10,15 @@ resolution-markers = [
"python_full_version < '3.13' and sys_platform != 'win32'",
]
+[[package]]
+name = "absl-py"
+version = "2.4.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/64/c7/8de93764ad66968d19329a7e0c147a2bb3c7054c554d4a119111b8f9440f/absl_py-2.4.0.tar.gz", hash = "sha256:8c6af82722b35cf71e0f4d1d47dcaebfff286e27110a99fc359349b247dfb5d4", size = 116543, upload-time = "2026-01-28T10:17:05.322Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/18/a6/907a406bb7d359e6a63f99c313846d9eec4f7e6f7437809e03aa00fa3074/absl_py-2.4.0-py3-none-any.whl", hash = "sha256:88476fd881ca8aab94ffa78b7b6c632a782ab3ba1cd19c9bd423abc4fb4cd28d", size = 135750, upload-time = "2026-01-28T10:17:04.19Z" },
+]
+
[[package]]
name = "accelerate"
version = "1.13.0"
@@ -28,6 +37,33 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/7e/46/02ac5e262d4af18054b3e922b2baedbb2a03289ee792162de60a865defc5/accelerate-1.13.0-py3-none-any.whl", hash = "sha256:cf1a3efb96c18f7b152eb0fa7490f3710b19c3f395699358f08decca2b8b62e0", size = 383744, upload-time = "2026-03-04T19:34:10.313Z" },
]
+[[package]]
+name = "ai-edge-litert"
+version = "2.1.5"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "backports-strenum" },
+ { name = "flatbuffers" },
+ { name = "numpy" },
+ { name = "protobuf" },
+ { name = "tqdm" },
+ { name = "typing-extensions" },
+]
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/d1/37/cf525a4ed6aff573b10a162b7bff19673e50ce6f45daac0a095dcd099a28/ai_edge_litert-2.1.5-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:b62fd3d90e643bcc3e3a31885a7636b5662527d48c27154cf07793f26896f018", size = 9711976, upload-time = "2026-05-15T23:34:16.988Z" },
+ { url = "https://files.pythonhosted.org/packages/4a/6f/b43fc56831ecc439a6e6eb0e1bfdffbfe6213e8c706aa421017dce1ac56b/ai_edge_litert-2.1.5-cp312-cp312-manylinux_2_27_aarch64.whl", hash = "sha256:1d282722cdfb70bb42d457f4bfe7789f88187ccadd73fea9f0107dce6e98fe45", size = 12867617, upload-time = "2026-05-15T23:11:22.759Z" },
+ { url = "https://files.pythonhosted.org/packages/c9/20/a76ba29b1c4c3009b5c64f4ec7f5fe5165104fe15481bd79b55927948541/ai_edge_litert-2.1.5-cp312-cp312-manylinux_2_27_x86_64.whl", hash = "sha256:f1c6d8db4382890881baeb8ed13c0802ada022e0b104b0db8fccf31353899ee0", size = 17570644, upload-time = "2026-05-15T23:05:33.664Z" },
+ { url = "https://files.pythonhosted.org/packages/82/fa/5399085daffd9949668d13a83a48db9f69af535ef9f9d06fdf71f0205321/ai_edge_litert-2.1.5-cp312-cp312-win_amd64.whl", hash = "sha256:6d2db6759188b12be9fd095468c3c7069c1bc7c128de78f42f862fb654d247ea", size = 17755006, upload-time = "2026-05-15T23:39:24.555Z" },
+ { url = "https://files.pythonhosted.org/packages/3d/fd/28f24305904954f31ee1a979eea39b670c6a1ab13cf6e5e4fe5f128e9eff/ai_edge_litert-2.1.5-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:2af16685a4cc8923d9a89ee5c732b8c66ae2c4d0cb538dbcebba3b93afb662ca", size = 9713185, upload-time = "2026-05-15T23:34:19.431Z" },
+ { url = "https://files.pythonhosted.org/packages/d3/88/e7e32b37f972b4877df788d1ad0995388f823d74dc2743b200bf71a25a53/ai_edge_litert-2.1.5-cp313-cp313-manylinux_2_27_aarch64.whl", hash = "sha256:57506e9e9b91ee04e4cf3e986d184b0c92f1aa9f510c32841835c0d719a7fbb5", size = 12867247, upload-time = "2026-05-15T23:11:24.667Z" },
+ { url = "https://files.pythonhosted.org/packages/7d/c1/5192c360a467dd6e09fa28388022c37bca5d9777e616da40dc5d517186a9/ai_edge_litert-2.1.5-cp313-cp313-manylinux_2_27_x86_64.whl", hash = "sha256:b7825e13454a90e7a4782ecdb6f6d987a2bcc9147aa90eb181193867bf696f24", size = 17570488, upload-time = "2026-05-15T23:05:36.078Z" },
+ { url = "https://files.pythonhosted.org/packages/d5/ba/154599984a8eccbdd14bcd89fabc2e9cbdfcbf537a3c1ab08712c7ce7d02/ai_edge_litert-2.1.5-cp313-cp313-win_amd64.whl", hash = "sha256:602fc90f6baf396df0fe4b0317b35627305c2233beb48db194b79e7e420bafa7", size = 17755085, upload-time = "2026-05-15T23:39:26.901Z" },
+ { url = "https://files.pythonhosted.org/packages/5e/51/34f0ad0def8b15093d602ca6904b129d3ab242ae4711c872190d5c95c010/ai_edge_litert-2.1.5-cp314-cp314-macosx_12_0_arm64.whl", hash = "sha256:597ea79f947452c79ca7e9dd7abfedad6179302851df4a4bb4c755069a2f5ffe", size = 9712496, upload-time = "2026-05-15T23:34:21.36Z" },
+ { url = "https://files.pythonhosted.org/packages/42/43/2b8ef317dca3096372cbf820aa5a3a6c8857dad67f31840a1e17a951e2d3/ai_edge_litert-2.1.5-cp314-cp314-manylinux_2_27_aarch64.whl", hash = "sha256:41d2e8267f047ed14f14a5ff4e0f677a9c5143dd17a36991513f8cc5ef580a3b", size = 12867899, upload-time = "2026-05-15T23:11:26.584Z" },
+ { url = "https://files.pythonhosted.org/packages/c2/d5/9640952de755409c84e09eec5d107675c59fe4f4859391464275eb2c3db6/ai_edge_litert-2.1.5-cp314-cp314-manylinux_2_27_x86_64.whl", hash = "sha256:4fa10d99b2f8647678850684d31075fbb304f50535d5788c2ca93e273ce727a8", size = 17570849, upload-time = "2026-05-15T23:05:37.981Z" },
+ { url = "https://files.pythonhosted.org/packages/0d/d4/0843e5a41bf71eb99ccd1e54714485df527bec19e8c8e679e1395176345d/ai_edge_litert-2.1.5-cp314-cp314-win_amd64.whl", hash = "sha256:14b35558bfce76146046cc2fc647f3fde97146e8b7622ea7fb02e85ac16cd906", size = 18380722, upload-time = "2026-05-15T23:39:29.365Z" },
+]
+
[[package]]
name = "aiohappyeyeballs"
version = "2.6.1"
@@ -39,7 +75,7 @@ wheels = [
[[package]]
name = "aiohttp"
-version = "3.13.5"
+version = "3.14.3"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "aiohappyeyeballs" },
@@ -48,78 +84,93 @@ dependencies = [
{ name = "frozenlist" },
{ name = "multidict" },
{ name = "propcache" },
+ { name = "typing-extensions", marker = "python_full_version < '3.13'" },
{ name = "yarl" },
]
-sdist = { url = "https://files.pythonhosted.org/packages/77/9a/152096d4808df8e4268befa55fba462f440f14beab85e8ad9bf990516918/aiohttp-3.13.5.tar.gz", hash = "sha256:9d98cc980ecc96be6eb4c1994ce35d28d8b1f5e5208a23b421187d1209dbb7d1", size = 7858271, upload-time = "2026-03-31T22:01:03.343Z" }
-wheels = [
- { url = "https://files.pythonhosted.org/packages/be/6f/353954c29e7dcce7cf00280a02c75f30e133c00793c7a2ed3776d7b2f426/aiohttp-3.13.5-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:023ecba036ddd840b0b19bf195bfae970083fd7024ce1ac22e9bba90464620e9", size = 748876, upload-time = "2026-03-31T21:57:36.319Z" },
- { url = "https://files.pythonhosted.org/packages/f5/1b/428a7c64687b3b2e9cd293186695affc0e1e54a445d0361743b231f11066/aiohttp-3.13.5-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:15c933ad7920b7d9a20de151efcd05a6e38302cbf0e10c9b2acb9a42210a2416", size = 499557, upload-time = "2026-03-31T21:57:38.236Z" },
- { url = "https://files.pythonhosted.org/packages/29/47/7be41556bfbb6917069d6a6634bb7dd5e163ba445b783a90d40f5ac7e3a7/aiohttp-3.13.5-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:ab2899f9fa2f9f741896ebb6fa07c4c883bfa5c7f2ddd8cf2aafa86fa981b2d2", size = 500258, upload-time = "2026-03-31T21:57:39.923Z" },
- { url = "https://files.pythonhosted.org/packages/67/84/c9ecc5828cb0b3695856c07c0a6817a99d51e2473400f705275a2b3d9239/aiohttp-3.13.5-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a60eaa2d440cd4707696b52e40ed3e2b0f73f65be07fd0ef23b6b539c9c0b0b4", size = 1749199, upload-time = "2026-03-31T21:57:41.938Z" },
- { url = "https://files.pythonhosted.org/packages/f0/d3/3c6d610e66b495657622edb6ae7c7fd31b2e9086b4ec50b47897ad6042a9/aiohttp-3.13.5-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:55b3bdd3292283295774ab585160c4004f4f2f203946997f49aac032c84649e9", size = 1721013, upload-time = "2026-03-31T21:57:43.904Z" },
- { url = "https://files.pythonhosted.org/packages/49/a0/24409c12217456df0bae7babe3b014e460b0b38a8e60753d6cb339f6556d/aiohttp-3.13.5-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c2b2355dc094e5f7d45a7bb262fe7207aa0460b37a0d87027dcf21b5d890e7d5", size = 1781501, upload-time = "2026-03-31T21:57:46.285Z" },
- { url = "https://files.pythonhosted.org/packages/98/9d/b65ec649adc5bccc008b0957a9a9c691070aeac4e41cea18559fef49958b/aiohttp-3.13.5-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:b38765950832f7d728297689ad78f5f2cf79ff82487131c4d26fe6ceecdc5f8e", size = 1878981, upload-time = "2026-03-31T21:57:48.734Z" },
- { url = "https://files.pythonhosted.org/packages/57/d8/8d44036d7eb7b6a8ec4c5494ea0c8c8b94fbc0ed3991c1a7adf230df03bf/aiohttp-3.13.5-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b18f31b80d5a33661e08c89e202edabf1986e9b49c42b4504371daeaa11b47c1", size = 1767934, upload-time = "2026-03-31T21:57:51.171Z" },
- { url = "https://files.pythonhosted.org/packages/31/04/d3f8211f273356f158e3464e9e45484d3fb8c4ce5eb2f6fe9405c3273983/aiohttp-3.13.5-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:33add2463dde55c4f2d9635c6ab33ce154e5ecf322bd26d09af95c5f81cfa286", size = 1566671, upload-time = "2026-03-31T21:57:53.326Z" },
- { url = "https://files.pythonhosted.org/packages/41/db/073e4ebe00b78e2dfcacff734291651729a62953b48933d765dc513bf798/aiohttp-3.13.5-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:327cc432fdf1356fb4fbc6fe833ad4e9f6aacb71a8acaa5f1855e4b25910e4a9", size = 1705219, upload-time = "2026-03-31T21:57:55.385Z" },
- { url = "https://files.pythonhosted.org/packages/48/45/7dfba71a2f9fd97b15c95c06819de7eb38113d2cdb6319669195a7d64270/aiohttp-3.13.5-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:7c35b0bf0b48a70b4cb4fc5d7bed9b932532728e124874355de1a0af8ec4bc88", size = 1743049, upload-time = "2026-03-31T21:57:57.341Z" },
- { url = "https://files.pythonhosted.org/packages/18/71/901db0061e0f717d226386a7f471bb59b19566f2cae5f0d93874b017271f/aiohttp-3.13.5-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:df23d57718f24badef8656c49743e11a89fd6f5358fa8a7b96e728fda2abf7d3", size = 1749557, upload-time = "2026-03-31T21:57:59.626Z" },
- { url = "https://files.pythonhosted.org/packages/08/d5/41eebd16066e59cd43728fe74bce953d7402f2b4ddfdfef2c0e9f17ca274/aiohttp-3.13.5-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:02e048037a6501a5ec1f6fc9736135aec6eb8a004ce48838cb951c515f32c80b", size = 1558931, upload-time = "2026-03-31T21:58:01.972Z" },
- { url = "https://files.pythonhosted.org/packages/30/e6/4a799798bf05740e66c3a1161079bda7a3dd8e22ca392481d7a7f9af82a6/aiohttp-3.13.5-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:31cebae8b26f8a615d2b546fee45d5ffb76852ae6450e2a03f42c9102260d6fe", size = 1774125, upload-time = "2026-03-31T21:58:04.007Z" },
- { url = "https://files.pythonhosted.org/packages/84/63/7749337c90f92bc2cb18f9560d67aa6258c7060d1397d21529b8004fcf6f/aiohttp-3.13.5-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:888e78eb5ca55a615d285c3c09a7a91b42e9dd6fc699b166ebd5dee87c9ccf14", size = 1732427, upload-time = "2026-03-31T21:58:06.337Z" },
- { url = "https://files.pythonhosted.org/packages/98/de/cf2f44ff98d307e72fb97d5f5bbae3bfcb442f0ea9790c0bf5c5c2331404/aiohttp-3.13.5-cp312-cp312-win32.whl", hash = "sha256:8bd3ec6376e68a41f9f95f5ed170e2fcf22d4eb27a1f8cb361d0508f6e0557f3", size = 433534, upload-time = "2026-03-31T21:58:08.712Z" },
- { url = "https://files.pythonhosted.org/packages/aa/ca/eadf6f9c8fa5e31d40993e3db153fb5ed0b11008ad5d9de98a95045bed84/aiohttp-3.13.5-cp312-cp312-win_amd64.whl", hash = "sha256:110e448e02c729bcebb18c60b9214a87ba33bac4a9fa5e9a5f139938b56c6cb1", size = 460446, upload-time = "2026-03-31T21:58:10.945Z" },
- { url = "https://files.pythonhosted.org/packages/78/e9/d76bf503005709e390122d34e15256b88f7008e246c4bdbe915cd4f1adce/aiohttp-3.13.5-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:a5029cc80718bbd545123cd8fe5d15025eccaaaace5d0eeec6bd556ad6163d61", size = 742930, upload-time = "2026-03-31T21:58:13.155Z" },
- { url = "https://files.pythonhosted.org/packages/57/00/4b7b70223deaebd9bb85984d01a764b0d7bd6526fcdc73cca83bcbe7243e/aiohttp-3.13.5-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:4bb6bf5811620003614076bdc807ef3b5e38244f9d25ca5fe888eaccea2a9832", size = 496927, upload-time = "2026-03-31T21:58:15.073Z" },
- { url = "https://files.pythonhosted.org/packages/9c/f5/0fb20fb49f8efdcdce6cd8127604ad2c503e754a8f139f5e02b01626523f/aiohttp-3.13.5-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:a84792f8631bf5a94e52d9cc881c0b824ab42717165a5579c760b830d9392ac9", size = 497141, upload-time = "2026-03-31T21:58:17.009Z" },
- { url = "https://files.pythonhosted.org/packages/3b/86/b7c870053e36a94e8951b803cb5b909bfbc9b90ca941527f5fcafbf6b0fa/aiohttp-3.13.5-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:57653eac22c6a4c13eb22ecf4d673d64a12f266e72785ab1c8b8e5940d0e8090", size = 1732476, upload-time = "2026-03-31T21:58:18.925Z" },
- { url = "https://files.pythonhosted.org/packages/b5/e5/4e161f84f98d80c03a238671b4136e6530453d65262867d989bbe78244d0/aiohttp-3.13.5-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:e5e5f7debc7a57af53fdf5c5009f9391d9f4c12867049d509bf7bb164a6e295b", size = 1706507, upload-time = "2026-03-31T21:58:21.094Z" },
- { url = "https://files.pythonhosted.org/packages/d4/56/ea11a9f01518bd5a2a2fcee869d248c4b8a0cfa0bb13401574fa31adf4d4/aiohttp-3.13.5-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c719f65bebcdf6716f10e9eff80d27567f7892d8988c06de12bbbd39307c6e3a", size = 1773465, upload-time = "2026-03-31T21:58:23.159Z" },
- { url = "https://files.pythonhosted.org/packages/eb/40/333ca27fb74b0383f17c90570c748f7582501507307350a79d9f9f3c6eb1/aiohttp-3.13.5-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d97f93fdae594d886c5a866636397e2bcab146fd7a132fd6bb9ce182224452f8", size = 1873523, upload-time = "2026-03-31T21:58:25.59Z" },
- { url = "https://files.pythonhosted.org/packages/f0/d2/e2f77eef1acb7111405433c707dc735e63f67a56e176e72e9e7a2cd3f493/aiohttp-3.13.5-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3df334e39d4c2f899a914f1dba283c1aadc311790733f705182998c6f7cae665", size = 1754113, upload-time = "2026-03-31T21:58:27.624Z" },
- { url = "https://files.pythonhosted.org/packages/fb/56/3f653d7f53c89669301ec9e42c95233e2a0c0a6dd051269e6e678db4fdb0/aiohttp-3.13.5-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:fe6970addfea9e5e081401bcbadf865d2b6da045472f58af08427e108d618540", size = 1562351, upload-time = "2026-03-31T21:58:29.918Z" },
- { url = "https://files.pythonhosted.org/packages/ec/a6/9b3e91eb8ae791cce4ee736da02211c85c6f835f1bdfac0594a8a3b7018c/aiohttp-3.13.5-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:7becdf835feff2f4f335d7477f121af787e3504b48b449ff737afb35869ba7bb", size = 1693205, upload-time = "2026-03-31T21:58:32.214Z" },
- { url = "https://files.pythonhosted.org/packages/98/fc/bfb437a99a2fcebd6b6eaec609571954de2ed424f01c352f4b5504371dd3/aiohttp-3.13.5-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:676e5651705ad5d8a70aeb8eb6936c436d8ebbd56e63436cb7dd9bb36d2a9a46", size = 1730618, upload-time = "2026-03-31T21:58:34.728Z" },
- { url = "https://files.pythonhosted.org/packages/e4/b6/c8534862126191a034f68153194c389addc285a0f1347d85096d349bbc15/aiohttp-3.13.5-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:9b16c653d38eb1a611cc898c41e76859ca27f119d25b53c12875fd0474ae31a8", size = 1745185, upload-time = "2026-03-31T21:58:36.909Z" },
- { url = "https://files.pythonhosted.org/packages/0b/93/4ca8ee2ef5236e2707e0fd5fecb10ce214aee1ff4ab307af9c558bda3b37/aiohttp-3.13.5-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:999802d5fa0389f58decd24b537c54aa63c01c3219ce17d1214cbda3c2b22d2d", size = 1557311, upload-time = "2026-03-31T21:58:39.38Z" },
- { url = "https://files.pythonhosted.org/packages/57/ae/76177b15f18c5f5d094f19901d284025db28eccc5ae374d1d254181d33f4/aiohttp-3.13.5-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:ec707059ee75732b1ba130ed5f9580fe10ff75180c812bc267ded039db5128c6", size = 1773147, upload-time = "2026-03-31T21:58:41.476Z" },
- { url = "https://files.pythonhosted.org/packages/01/a4/62f05a0a98d88af59d93b7fcac564e5f18f513cb7471696ac286db970d6a/aiohttp-3.13.5-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:2d6d44a5b48132053c2f6cd5c8cb14bc67e99a63594e336b0f2af81e94d5530c", size = 1730356, upload-time = "2026-03-31T21:58:44.049Z" },
- { url = "https://files.pythonhosted.org/packages/e4/85/fc8601f59dfa8c9523808281f2da571f8b4699685f9809a228adcc90838d/aiohttp-3.13.5-cp313-cp313-win32.whl", hash = "sha256:329f292ed14d38a6c4c435e465f48bebb47479fd676a0411936cc371643225cc", size = 432637, upload-time = "2026-03-31T21:58:46.167Z" },
- { url = "https://files.pythonhosted.org/packages/c0/1b/ac685a8882896acf0f6b31d689e3792199cfe7aba37969fa91da63a7fa27/aiohttp-3.13.5-cp313-cp313-win_amd64.whl", hash = "sha256:69f571de7500e0557801c0b51f4780482c0ec5fe2ac851af5a92cfce1af1cb83", size = 458896, upload-time = "2026-03-31T21:58:48.119Z" },
- { url = "https://files.pythonhosted.org/packages/5d/ce/46572759afc859e867a5bc8ec3487315869013f59281ce61764f76d879de/aiohttp-3.13.5-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:eb4639f32fd4a9904ab8fb45bf3383ba71137f3d9d4ba25b3b3f3109977c5b8c", size = 745721, upload-time = "2026-03-31T21:58:50.229Z" },
- { url = "https://files.pythonhosted.org/packages/13/fe/8a2efd7626dbe6049b2ef8ace18ffda8a4dfcbe1bcff3ac30c0c7575c20b/aiohttp-3.13.5-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:7e5dc4311bd5ac493886c63cbf76ab579dbe4641268e7c74e48e774c74b6f2be", size = 497663, upload-time = "2026-03-31T21:58:52.232Z" },
- { url = "https://files.pythonhosted.org/packages/9b/91/cc8cc78a111826c54743d88651e1687008133c37e5ee615fee9b57990fac/aiohttp-3.13.5-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:756c3c304d394977519824449600adaf2be0ccee76d206ee339c5e76b70ded25", size = 499094, upload-time = "2026-03-31T21:58:54.566Z" },
- { url = "https://files.pythonhosted.org/packages/0a/33/a8362cb15cf16a3af7e86ed11962d5cd7d59b449202dc576cdc731310bde/aiohttp-3.13.5-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ecc26751323224cf8186efcf7fbcbc30f4e1d8c7970659daf25ad995e4032a56", size = 1726701, upload-time = "2026-03-31T21:58:56.864Z" },
- { url = "https://files.pythonhosted.org/packages/45/0c/c091ac5c3a17114bd76cbf85d674650969ddf93387876cf67f754204bd77/aiohttp-3.13.5-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:10a75acfcf794edf9d8db50e5a7ec5fc818b2a8d3f591ce93bc7b1210df016d2", size = 1683360, upload-time = "2026-03-31T21:58:59.072Z" },
- { url = "https://files.pythonhosted.org/packages/23/73/bcee1c2b79bc275e964d1446c55c54441a461938e70267c86afaae6fba27/aiohttp-3.13.5-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:0f7a18f258d124cd678c5fe072fe4432a4d5232b0657fca7c1847f599233c83a", size = 1773023, upload-time = "2026-03-31T21:59:01.776Z" },
- { url = "https://files.pythonhosted.org/packages/c7/ef/720e639df03004fee2d869f771799d8c23046dec47d5b81e396c7cda583a/aiohttp-3.13.5-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:df6104c009713d3a89621096f3e3e88cc323fd269dbd7c20afe18535094320be", size = 1853795, upload-time = "2026-03-31T21:59:04.568Z" },
- { url = "https://files.pythonhosted.org/packages/bd/c9/989f4034fb46841208de7aeeac2c6d8300745ab4f28c42f629ba77c2d916/aiohttp-3.13.5-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:241a94f7de7c0c3b616627aaad530fe2cb620084a8b144d3be7b6ecfe95bae3b", size = 1730405, upload-time = "2026-03-31T21:59:07.221Z" },
- { url = "https://files.pythonhosted.org/packages/ce/75/ee1fd286ca7dc599d824b5651dad7b3be7ff8d9a7e7b3fe9820d9180f7db/aiohttp-3.13.5-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c974fb66180e58709b6fc402846f13791240d180b74de81d23913abe48e96d94", size = 1558082, upload-time = "2026-03-31T21:59:09.484Z" },
- { url = "https://files.pythonhosted.org/packages/c3/20/1e9e6650dfc436340116b7aa89ff8cb2bbdf0abc11dfaceaad8f74273a10/aiohttp-3.13.5-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:6e27ea05d184afac78aabbac667450c75e54e35f62238d44463131bd3f96753d", size = 1692346, upload-time = "2026-03-31T21:59:12.068Z" },
- { url = "https://files.pythonhosted.org/packages/d8/40/8ebc6658d48ea630ac7903912fe0dd4e262f0e16825aa4c833c56c9f1f56/aiohttp-3.13.5-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:a79a6d399cef33a11b6f004c67bb07741d91f2be01b8d712d52c75711b1e07c7", size = 1698891, upload-time = "2026-03-31T21:59:14.552Z" },
- { url = "https://files.pythonhosted.org/packages/d8/78/ea0ae5ec8ba7a5c10bdd6e318f1ba5e76fcde17db8275188772afc7917a4/aiohttp-3.13.5-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:c632ce9c0b534fbe25b52c974515ed674937c5b99f549a92127c85f771a78772", size = 1742113, upload-time = "2026-03-31T21:59:17.068Z" },
- { url = "https://files.pythonhosted.org/packages/8a/66/9d308ed71e3f2491be1acb8769d96c6f0c47d92099f3bc9119cada27b357/aiohttp-3.13.5-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:fceedde51fbd67ee2bcc8c0b33d0126cc8b51ef3bbde2f86662bd6d5a6f10ec5", size = 1553088, upload-time = "2026-03-31T21:59:19.541Z" },
- { url = "https://files.pythonhosted.org/packages/da/a6/6cc25ed8dfc6e00c90f5c6d126a98e2cf28957ad06fa1036bd34b6f24a2c/aiohttp-3.13.5-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:f92995dfec9420bb69ae629abf422e516923ba79ba4403bc750d94fb4a6c68c1", size = 1757976, upload-time = "2026-03-31T21:59:22.311Z" },
- { url = "https://files.pythonhosted.org/packages/c1/2b/cce5b0ffe0de99c83e5e36d8f828e4161e415660a9f3e58339d07cce3006/aiohttp-3.13.5-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:20ae0ff08b1f2c8788d6fb85afcb798654ae6ba0b747575f8562de738078457b", size = 1712444, upload-time = "2026-03-31T21:59:24.635Z" },
- { url = "https://files.pythonhosted.org/packages/6c/cf/9e1795b4160c58d29421eafd1a69c6ce351e2f7c8d3c6b7e4ca44aea1a5b/aiohttp-3.13.5-cp314-cp314-win32.whl", hash = "sha256:b20df693de16f42b2472a9c485e1c948ee55524786a0a34345511afdd22246f3", size = 438128, upload-time = "2026-03-31T21:59:27.291Z" },
- { url = "https://files.pythonhosted.org/packages/22/4d/eaedff67fc805aeba4ba746aec891b4b24cebb1a7d078084b6300f79d063/aiohttp-3.13.5-cp314-cp314-win_amd64.whl", hash = "sha256:f85c6f327bf0b8c29da7d93b1cabb6363fb5e4e160a32fa241ed2dce21b73162", size = 464029, upload-time = "2026-03-31T21:59:29.429Z" },
- { url = "https://files.pythonhosted.org/packages/79/11/c27d9332ee20d68dd164dc12a6ecdef2e2e35ecc97ed6cf0d2442844624b/aiohttp-3.13.5-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:1efb06900858bb618ff5cee184ae2de5828896c448403d51fb633f09e109be0a", size = 778758, upload-time = "2026-03-31T21:59:31.547Z" },
- { url = "https://files.pythonhosted.org/packages/04/fb/377aead2e0a3ba5f09b7624f702a964bdf4f08b5b6728a9799830c80041e/aiohttp-3.13.5-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:fee86b7c4bd29bdaf0d53d14739b08a106fdda809ca5fe032a15f52fae5fe254", size = 512883, upload-time = "2026-03-31T21:59:34.098Z" },
- { url = "https://files.pythonhosted.org/packages/bb/a6/aa109a33671f7a5d3bd78b46da9d852797c5e665bfda7d6b373f56bff2ec/aiohttp-3.13.5-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:20058e23909b9e65f9da62b396b77dfa95965cbe840f8def6e572538b1d32e36", size = 516668, upload-time = "2026-03-31T21:59:36.497Z" },
- { url = "https://files.pythonhosted.org/packages/79/b3/ca078f9f2fa9563c36fb8ef89053ea2bb146d6f792c5104574d49d8acb63/aiohttp-3.13.5-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8cf20a8d6868cb15a73cab329ffc07291ba8c22b1b88176026106ae39aa6df0f", size = 1883461, upload-time = "2026-03-31T21:59:38.723Z" },
- { url = "https://files.pythonhosted.org/packages/b7/e3/a7ad633ca1ca497b852233a3cce6906a56c3225fb6d9217b5e5e60b7419d/aiohttp-3.13.5-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:330f5da04c987f1d5bdb8ae189137c77139f36bd1cb23779ca1a354a4b027800", size = 1747661, upload-time = "2026-03-31T21:59:41.187Z" },
- { url = "https://files.pythonhosted.org/packages/33/b9/cd6fe579bed34a906d3d783fe60f2fa297ef55b27bb4538438ee49d4dc41/aiohttp-3.13.5-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:6f1cbf0c7926d315c3c26c2da41fd2b5d2fe01ac0e157b78caefc51a782196cf", size = 1863800, upload-time = "2026-03-31T21:59:43.84Z" },
- { url = "https://files.pythonhosted.org/packages/c0/3f/2c1e2f5144cefa889c8afd5cf431994c32f3b29da9961698ff4e3811b79a/aiohttp-3.13.5-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:53fc049ed6390d05423ba33103ded7281fe897cf97878f369a527070bd95795b", size = 1958382, upload-time = "2026-03-31T21:59:46.187Z" },
- { url = "https://files.pythonhosted.org/packages/66/1d/f31ec3f1013723b3babe3609e7f119c2c2fb6ef33da90061a705ef3e1bc8/aiohttp-3.13.5-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:898703aa2667e3c5ca4c54ca36cd73f58b7a38ef87a5606414799ebce4d3fd3a", size = 1803724, upload-time = "2026-03-31T21:59:48.656Z" },
- { url = "https://files.pythonhosted.org/packages/0e/b4/57712dfc6f1542f067daa81eb61da282fab3e6f1966fca25db06c4fc62d5/aiohttp-3.13.5-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0494a01ca9584eea1e5fbd6d748e61ecff218c51b576ee1999c23db7066417d8", size = 1640027, upload-time = "2026-03-31T21:59:51.284Z" },
- { url = "https://files.pythonhosted.org/packages/25/3c/734c878fb43ec083d8e31bf029daae1beafeae582d1b35da234739e82ee7/aiohttp-3.13.5-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:6cf81fe010b8c17b09495cbd15c1d35afbc8fb405c0c9cf4738e5ae3af1d65be", size = 1806644, upload-time = "2026-03-31T21:59:53.753Z" },
- { url = "https://files.pythonhosted.org/packages/20/a5/f671e5cbec1c21d044ff3078223f949748f3a7f86b14e34a365d74a5d21f/aiohttp-3.13.5-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:c564dd5f09ddc9d8f2c2d0a301cd30a79a2cc1b46dd1a73bef8f0038863d016b", size = 1791630, upload-time = "2026-03-31T21:59:56.239Z" },
- { url = "https://files.pythonhosted.org/packages/0b/63/fb8d0ad63a0b8a99be97deac8c04dacf0785721c158bdf23d679a87aa99e/aiohttp-3.13.5-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:2994be9f6e51046c4f864598fd9abeb4fba6e88f0b2152422c9666dcd4aea9c6", size = 1809403, upload-time = "2026-03-31T21:59:59.103Z" },
- { url = "https://files.pythonhosted.org/packages/59/0c/bfed7f30662fcf12206481c2aac57dedee43fe1c49275e85b3a1e1742294/aiohttp-3.13.5-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:157826e2fa245d2ef46c83ea8a5faf77ca19355d278d425c29fda0beb3318037", size = 1634924, upload-time = "2026-03-31T22:00:02.116Z" },
- { url = "https://files.pythonhosted.org/packages/17/d6/fd518d668a09fd5a3319ae5e984d4d80b9a4b3df4e21c52f02251ef5a32e/aiohttp-3.13.5-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:a8aca50daa9493e9e13c0f566201a9006f080e7c50e5e90d0b06f53146a54500", size = 1836119, upload-time = "2026-03-31T22:00:04.756Z" },
- { url = "https://files.pythonhosted.org/packages/78/b7/15fb7a9d52e112a25b621c67b69c167805cb1f2ab8f1708a5c490d1b52fe/aiohttp-3.13.5-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:3b13560160d07e047a93f23aaa30718606493036253d5430887514715b67c9d9", size = 1772072, upload-time = "2026-03-31T22:00:07.494Z" },
- { url = "https://files.pythonhosted.org/packages/7e/df/57ba7f0c4a553fc2bd8b6321df236870ec6fd64a2a473a8a13d4f733214e/aiohttp-3.13.5-cp314-cp314t-win32.whl", hash = "sha256:9a0f4474b6ea6818b41f82172d799e4b3d29e22c2c520ce4357856fced9af2f8", size = 471819, upload-time = "2026-03-31T22:00:10.277Z" },
- { url = "https://files.pythonhosted.org/packages/62/29/2f8418269e46454a26171bfdd6a055d74febf32234e474930f2f60a17145/aiohttp-3.13.5-cp314-cp314t-win_amd64.whl", hash = "sha256:18a2f6c1182c51baa1d28d68fea51513cb2a76612f038853c0ad3c145423d3d9", size = 505441, upload-time = "2026-03-31T22:00:12.791Z" },
+sdist = { url = "https://files.pythonhosted.org/packages/58/d9/22ce5786ac0c1653ae8b6c23bded02c1686d11f0dbb45b31ce128e0df985/aiohttp-3.14.3.tar.gz", hash = "sha256:9491196535a88924a60afd5b5f434b5b203b6cc616250878dbdb223a8f7844bc", size = 7971213, upload-time = "2026-07-23T01:57:27.037Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/18/d4/eb96299230e20acf2efae207cb8d69051f1f68e357e5ea5e479bf6fb097a/aiohttp-3.14.3-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:39aded8c7f3b935b54aab1d8d73c70ec0ee2d3ec3b943e0e86611bc150ba47f5", size = 754690, upload-time = "2026-07-23T01:53:47.332Z" },
+ { url = "https://files.pythonhosted.org/packages/88/11/e7a70a209eb9a067c0d3212b518a0134e3484f5178c7533878b6b514d469/aiohttp-3.14.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:5bcb6ff3fdab1258a192679ff1a05d44f59626430aa05cd1a9d2447423599228", size = 509484, upload-time = "2026-07-23T01:53:51.159Z" },
+ { url = "https://files.pythonhosted.org/packages/30/07/4bbc222cc8dbe31d4c3e8a5baad2286e4d42026ac0c570027b89afce6344/aiohttp-3.14.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:617105e2c3018ee38d0c8ce5ee3c84f621a6d8b9f723202aacaff28449ca91ee", size = 511949, upload-time = "2026-07-23T01:53:55.083Z" },
+ { url = "https://files.pythonhosted.org/packages/54/b9/42e74c46b7b7c794b995bbc1f573fb48950c38b19d8600c62a6804ee2d67/aiohttp-3.14.3-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f631fe87a6f30df5fbe6d79640b25e4cffb38c31c7fb6f10871517b84b0f8c1a", size = 1765282, upload-time = "2026-07-23T01:53:59.662Z" },
+ { url = "https://files.pythonhosted.org/packages/6b/ed/62bc4d74363ad346d518e0720363a949f63e2e23439a79eb5813d4d29bb3/aiohttp-3.14.3-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:a94dbaae5ae27bd849c93570669bff91e0510f33a80805738e3de72a7be0447b", size = 1741511, upload-time = "2026-07-23T01:54:04.063Z" },
+ { url = "https://files.pythonhosted.org/packages/d0/9f/181e8a8bc79e47d13c7fc4540bd7a3b729d9505609c61f392a8dd2fbfe55/aiohttp-3.14.3-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8f2f1c4c032c7cedd7d8da6f54c97b70266c6570c3108d3fdffee7188bb70529", size = 1810680, upload-time = "2026-07-23T01:54:09.882Z" },
+ { url = "https://files.pythonhosted.org/packages/5c/9a/dec94d6ad694552fe3424e3f1928d7a606a5d9d9433a04e7ecdd9d38ae7f/aiohttp-3.14.3-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ea05e1f97ceea523942d9b2a7d7c0359d781d683d6b043f5943a602b14da4787", size = 1905646, upload-time = "2026-07-23T01:54:13.475Z" },
+ { url = "https://files.pythonhosted.org/packages/52/b7/7cd31f29d6055bd711ae6e669367fba6f5ae9de463910a793e30556a8db7/aiohttp-3.14.3-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:543906c127fb1d929b95076db19b83fa2d46751006ff1e23b093aa5ac4d8db42", size = 1792122, upload-time = "2026-07-23T01:54:15.752Z" },
+ { url = "https://files.pythonhosted.org/packages/66/73/10b1ef93afa61f4963c746257b70ced619cf31a4798671de5fdb2608501d/aiohttp-3.14.3-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0a5ff2dfbb9ce645fa5b8ef3e02c6c0b9cc3f6030ff863d0c51fffc50cb5541b", size = 1591127, upload-time = "2026-07-23T01:54:19.489Z" },
+ { url = "https://files.pythonhosted.org/packages/49/ed/3b203fa6de1b338c14acdc06bf6ca9b043b7944f005966958c2ced932cde/aiohttp-3.14.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:041badb8f84396357c4d3ad26de6afd7a32b112f43d3c63045c0c8278cfd2043", size = 1725210, upload-time = "2026-07-23T01:54:24.129Z" },
+ { url = "https://files.pythonhosted.org/packages/28/b7/1c2aab8c706436dcc28598452488ac9cd7c409da815237c28c27d58993e6/aiohttp-3.14.3-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:530125ee1163c4219af35dc3aa1206e541e7b31b6efc1a3f93b70a136f65d427", size = 1764848, upload-time = "2026-07-23T01:54:27.973Z" },
+ { url = "https://files.pythonhosted.org/packages/54/50/94c28f08b131c4bf10984ea2c7a536c9920608bb2d6e7f95642c30cc87b7/aiohttp-3.14.3-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:c8653fd547c93a61aadc612007790f5555cdd18946fa48cf45e26d8ea4ea473d", size = 1777102, upload-time = "2026-07-23T01:54:31.775Z" },
+ { url = "https://files.pythonhosted.org/packages/13/d4/e7d09ba7d345fb2d74440fd2fa033c5e079fac05552927705986f41a364f/aiohttp-3.14.3-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:89176250f686cb9853c0fb7ead90e639e915b84a6f43eedc2a4e7ec21f1037f0", size = 1580205, upload-time = "2026-07-23T01:54:34.518Z" },
+ { url = "https://files.pythonhosted.org/packages/a3/84/072a91d68e1e1eb587985b54baab94221277f877e8ef274fc213a0ceae28/aiohttp-3.14.3-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:3a26434dafe408229ff3403458ca58de24fb51936504decac49ce6755f77e59d", size = 1797219, upload-time = "2026-07-23T01:54:36.995Z" },
+ { url = "https://files.pythonhosted.org/packages/e0/eb/aad34e897e668424d6e995da5dff8a4a09af93363d3392488772957a63aa/aiohttp-3.14.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:d1558173930a5a8d3069cee5c92fc91c87c4dbcb099debbb3622053717145a19", size = 1768629, upload-time = "2026-07-23T01:54:40.103Z" },
+ { url = "https://files.pythonhosted.org/packages/b6/2b/6bb88ddba0fecd9122aa3ebcad25996cf6c083a4a7040dbb3a4f97972af6/aiohttp-3.14.3-cp312-cp312-win32.whl", hash = "sha256:16100ad3ab8d649fdfbee87602d9d2dcdca9df0b9eda8a1b5fdc0d41f96da559", size = 451481, upload-time = "2026-07-23T01:54:42.547Z" },
+ { url = "https://files.pythonhosted.org/packages/76/9b/f2f8f108da17ecef2cc3efc424e8b7ad3782b1a8360f7b8eae8ced84f6ea/aiohttp-3.14.3-cp312-cp312-win_amd64.whl", hash = "sha256:33a2d7c28d33797a2e99923dffa63f83d908a19b6bf26cfe80fa790aa5e1a75a", size = 476845, upload-time = "2026-07-23T01:54:44.853Z" },
+ { url = "https://files.pythonhosted.org/packages/3e/44/28dac80a8941b604f4da10ce21097614ca1bf905ce93dca28d8d7de9c1e7/aiohttp-3.14.3-cp312-cp312-win_arm64.whl", hash = "sha256:362a3fd481769cac1a824514bcd86fda51c65e8fe6e051099e008fddde6db17c", size = 448050, upload-time = "2026-07-23T01:54:47.087Z" },
+ { url = "https://files.pythonhosted.org/packages/57/be/5afd201cc0ab139029aadb75392efe85a293403d9dd3a3226161c21ce00c/aiohttp-3.14.3-cp313-cp313-android_21_arm64_v8a.whl", hash = "sha256:2e9878ae68e4a5f1c0abe4dd497dbc3d51946f5837b56759e2a02e78fa90ef86", size = 506269, upload-time = "2026-07-23T01:54:49.075Z" },
+ { url = "https://files.pythonhosted.org/packages/22/09/dec8189d62b45ade009f6792a2264b942a90cb88aeaf181239933cd72c3c/aiohttp-3.14.3-cp313-cp313-android_21_x86_64.whl", hash = "sha256:f3d2669fe7dec7fc359ecdb5984b29b50d85d5d00f8c1cb61de4f4a24ee42627", size = 515166, upload-time = "2026-07-23T01:54:51.894Z" },
+ { url = "https://files.pythonhosted.org/packages/28/24/2854869d29ed8a8b19d74f9ec6629515f7e04d02dd329d9d179201e58e47/aiohttp-3.14.3-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:cc7cb243a68167172f48c1fd43cee91ec4b1d40cefd190edd43369d1a6bc9c82", size = 486263, upload-time = "2026-07-23T01:54:54.223Z" },
+ { url = "https://files.pythonhosted.org/packages/d4/dd/57187c8be2a35aea65eaee3bd2c3dcbbcf0204f5106c89637e3610380cd1/aiohttp-3.14.3-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:78253b573e6ffab5028924fc98bc281aae05445969982a10864bc360dea2016c", size = 492299, upload-time = "2026-07-23T01:54:56.236Z" },
+ { url = "https://files.pythonhosted.org/packages/b9/11/06ae6ed8f0d414edf4068861e233d8fe23ee699bfd4b3ceb8663db948a62/aiohttp-3.14.3-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:7041d52c3a7fa20c9e8c182b534704abb19502c8bdcbde7ab23bfda6f642394f", size = 502235, upload-time = "2026-07-23T01:54:58.377Z" },
+ { url = "https://files.pythonhosted.org/packages/7e/a3/559639c34a345d2cf7c52dff6838119f2eaf29eb508227b5b83f573af813/aiohttp-3.14.3-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:ac74facc01463f138b0da5580329cfcc82818dea5656e83ddcd11268fc12ff80", size = 750883, upload-time = "2026-07-23T01:55:00.65Z" },
+ { url = "https://files.pythonhosted.org/packages/91/cd/41e131f13afd1e7b0172a9d9eda085ef90eb8439f41f0d279db81ed3ae60/aiohttp-3.14.3-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:d6218d92e450824e9b4881f44e8c09f1853b490f9a64130801024a4793b1b3b0", size = 508473, upload-time = "2026-07-23T01:55:02.945Z" },
+ { url = "https://files.pythonhosted.org/packages/bc/6b/e7f13410d391c6e55b4c007a8de024355389d7d459e3d64c42b2d33617e5/aiohttp-3.14.3-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:11fb37ef075669eee52ab1928fbf6e1741fada40409fa309ebde9607a962aebf", size = 509190, upload-time = "2026-07-23T01:55:05.173Z" },
+ { url = "https://files.pythonhosted.org/packages/97/21/6464573e53d69672cc1eada3e5c5cb2d2efa82701e8305a0f2047a576967/aiohttp-3.14.3-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:55bdcc472aafe2de4a253045cc128007a64f1e0264fb675791e132ea5edaa3bd", size = 1761478, upload-time = "2026-07-23T01:55:07.383Z" },
+ { url = "https://files.pythonhosted.org/packages/1a/81/d217043a4c17fbce360905e3b2bdd20139ebc9a2de836d035d179c4da006/aiohttp-3.14.3-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c39846c3aad97a8530c89d7a3869a8f8e9e3762c6ac0504481e5c80948f7e807", size = 1735092, upload-time = "2026-07-23T01:55:09.803Z" },
+ { url = "https://files.pythonhosted.org/packages/a1/66/e13a02d0eeb1a9a502402a977abb4e4abff9fe4051c26f80558c57a7c975/aiohttp-3.14.3-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5895ef58c4620afe02fa16044f023dc4dafec08158f9d08874a46a7dbc0341b8", size = 1800546, upload-time = "2026-07-23T01:55:12.012Z" },
+ { url = "https://files.pythonhosted.org/packages/26/5e/57d42fca1d18cb5acc1cad945d017fabc5d6ae71d8a08ad66be8dc3ee544/aiohttp-3.14.3-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:fa9467a8113aa69d3d7c55a70ef0b7c636010a40993f3df9d9d0d73b3eb7ef24", size = 1895250, upload-time = "2026-07-23T01:55:14.357Z" },
+ { url = "https://files.pythonhosted.org/packages/ca/1c/7da8d08e74d56f00070822f9638ff3f1c563f8ad87d1efa996c87bfc8644/aiohttp-3.14.3-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d7d2deec16eeedf55f2c7cf75b521ea3856a5177e123844f8fd0f114ce252cb5", size = 1789289, upload-time = "2026-07-23T01:55:16.668Z" },
+ { url = "https://files.pythonhosted.org/packages/cd/0f/cf16bcf56896981c1a0319f5d5db9337994b5165730c48a8fa07e9b34be6/aiohttp-3.14.3-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:dd54d0e8717de95939766febac482ac0474d8ac3b048115f9f2b1d23a16e7db4", size = 1586706, upload-time = "2026-07-23T01:55:18.913Z" },
+ { url = "https://files.pythonhosted.org/packages/fe/6f/76eac12a7f2480e1e304f842efdb07db33256b0d9165b866b6ef0806c202/aiohttp-3.14.3-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:df82f3787c940c94986b34222d59c9e38843fba85139f36e85255a82ad5355a9", size = 1724652, upload-time = "2026-07-23T01:55:21.296Z" },
+ { url = "https://files.pythonhosted.org/packages/39/b6/19c8c592baeeb94b75f966547d40c02ac7590902306ec5863d5c027cf506/aiohttp-3.14.3-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:42a67efc36300d052fb4508a53e8b6901b9284b599ae63945c377569c5fcc1e1", size = 1756239, upload-time = "2026-07-23T01:55:23.705Z" },
+ { url = "https://files.pythonhosted.org/packages/dc/c9/4e9383150296f97f873b680c4de8fb2cd88608fb9f48c79edcb111611abc/aiohttp-3.14.3-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:7a75aa63cbf9b21cfaf60dc2657e19df2c2867d91707d653fee171ffeedd1371", size = 1769161, upload-time = "2026-07-23T01:55:26.082Z" },
+ { url = "https://files.pythonhosted.org/packages/aa/1e/147bdc6cc5de5f3ab011be8bf5d6e786633249f22c20bae06f85e45f5387/aiohttp-3.14.3-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:e92eb8acc45eb6a9f4935071a77edf5b85cc6f8dfad5cd99e97653c26593cdde", size = 1578759, upload-time = "2026-07-23T01:55:28.846Z" },
+ { url = "https://files.pythonhosted.org/packages/fd/31/78388a9d6040ece2e11df62ea229a822cf5e52d238374b220ae9975b2623/aiohttp-3.14.3-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:b014a6ed7cf912e787149fdc529166d3ceabac23f26efeea3158c9aba2354e7e", size = 1792025, upload-time = "2026-07-23T01:55:31.457Z" },
+ { url = "https://files.pythonhosted.org/packages/03/51/a3d29fdf2c25d796746af8ad6fe56a45d6256c38b0a8a2ed752e1160b3a2/aiohttp-3.14.3-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:3d4f72af88ac2474bb5bca640030320e3d38a0163a1d7533500e87be458eef71", size = 1768477, upload-time = "2026-07-23T01:55:33.87Z" },
+ { url = "https://files.pythonhosted.org/packages/29/a6/442e18b5afeade534d877a2dc3c3e392aff8d49787890b0cf84790410267/aiohttp-3.14.3-cp313-cp313-win32.whl", hash = "sha256:5f08ec777f35ee70720233b8b9811d3bb5d728137f30ac91b7457709c3261ac0", size = 451069, upload-time = "2026-07-23T01:55:36.121Z" },
+ { url = "https://files.pythonhosted.org/packages/9d/69/3d876ac02659f271cf7f6769f14a8e3de5b6e888ed8b5a7e998086a4cec8/aiohttp-3.14.3-cp313-cp313-win_amd64.whl", hash = "sha256:dff9461ec275f22135650d5ba4b4931a11f3958df7dfbb8db630000d4dee0883", size = 476518, upload-time = "2026-07-23T01:55:38.303Z" },
+ { url = "https://files.pythonhosted.org/packages/b2/0e/50d6e6471cd31edce8b282bdec59375a3a69124d8a989a0b1313355cae52/aiohttp-3.14.3-cp313-cp313-win_arm64.whl", hash = "sha256:ddcac3c6b382e81f1dd0499199d4136b877beb4cb5ef770bbbfba56c4b8f55d2", size = 447676, upload-time = "2026-07-23T01:55:40.451Z" },
+ { url = "https://files.pythonhosted.org/packages/c8/20/887fdcf832326571b370ffc347b3e70abe101096f3720126aac161b1d872/aiohttp-3.14.3-cp314-cp314-android_24_arm64_v8a.whl", hash = "sha256:49f7325beb0f85ef4aef5f48f490269575f83e6e2acad00a1d80b807eb027062", size = 509067, upload-time = "2026-07-23T01:55:42.618Z" },
+ { url = "https://files.pythonhosted.org/packages/ad/a3/92cec936f78cc4bf0fa5554ebe593b73459d94e3c62303e1902a4cccb6f7/aiohttp-3.14.3-cp314-cp314-android_24_x86_64.whl", hash = "sha256:e3be98a7c30b8c25d573dafba7171d66dfb05ee6a9070fc46535464ff97700a6", size = 514774, upload-time = "2026-07-23T01:55:44.937Z" },
+ { url = "https://files.pythonhosted.org/packages/29/ba/2a0c38df3fc557620b6a5acd98364af050053b6285b4dc7ee74100c63c18/aiohttp-3.14.3-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:614c61d478b83953e261d02bb2df750f17227cd33ef8002945bf5aebbde21919", size = 488134, upload-time = "2026-07-23T01:55:47.135Z" },
+ { url = "https://files.pythonhosted.org/packages/48/d6/d51b7d4bf309af3693940d8ffd2b9ed0b682434ef85959b7c9c137f60cf8/aiohttp-3.14.3-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:1caa7b0d05f3e3a36f87788c59e970a7ee1cefcfcbb924a9f138c4a6551c9cb7", size = 494201, upload-time = "2026-07-23T01:55:49.451Z" },
+ { url = "https://files.pythonhosted.org/packages/3f/5a/8f624384e5f1efabb5229b94157eb966b021e97bdb188c62860c2ae243c2/aiohttp-3.14.3-cp314-cp314-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:dfa68deb2a443bdaa3ea5297b0699c1464f08aef3812b486d1348eee61b07dc0", size = 502766, upload-time = "2026-07-23T01:55:51.656Z" },
+ { url = "https://files.pythonhosted.org/packages/a6/26/4ff0164370deec18fb19254ee4ab10b7a73304ac0c860b13f5f84663759b/aiohttp-3.14.3-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:e72ee89e28d907a18f46959b4eb0bb06701cc7f8cf4366e00029e2ccfaaf5924", size = 756557, upload-time = "2026-07-23T01:55:53.964Z" },
+ { url = "https://files.pythonhosted.org/packages/97/a3/7056b86dc0d9ec709ea9777eae3b0161428f943372f8b98c01c11593b682/aiohttp-3.14.3-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:ad4c8b7488d745d2ca4838ebd8ae5ba9b56341d30b1da43640e4ce87f9f49646", size = 510168, upload-time = "2026-07-23T01:55:56.22Z" },
+ { url = "https://files.pythonhosted.org/packages/85/ed/0357a015892fd68058bf2d39d3fd1958e459b997a7db30aaa6aaa434ae96/aiohttp-3.14.3-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:db332af25642007330fca8be5c4d194caf2bea7a7fc84415aff3497af5dfee6b", size = 512957, upload-time = "2026-07-23T01:55:58.437Z" },
+ { url = "https://files.pythonhosted.org/packages/47/d1/8aba53f15ccb2238405f5e9d30e2a8ca44f93878c26e7165ade00d374b1c/aiohttp-3.14.3-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:25bd2708db6bdf6a6630dd37bdcdfcb47c4434d22ac69c64665b802910140b30", size = 1750149, upload-time = "2026-07-23T01:56:00.856Z" },
+ { url = "https://files.pythonhosted.org/packages/49/bd/40c3fee327529284375c6701cbb0fa4600cc2e8432af1378f897e2ef7d3a/aiohttp-3.14.3-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:cef89a58e628c4efcac3275c2d68083f82426dcdc89c1492a6f654f9f7ea6ab9", size = 1707685, upload-time = "2026-07-23T01:56:03.371Z" },
+ { url = "https://files.pythonhosted.org/packages/2a/a3/ca0cc6724cca8114b05694abd916060758c79894c3aa5b012cdadc1bc28e/aiohttp-3.14.3-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c23ec8ee9d5ab2f5421f9c7fffce208435607af27fd46d4a44e031954352838f", size = 1803911, upload-time = "2026-07-23T01:56:05.817Z" },
+ { url = "https://files.pythonhosted.org/packages/95/b5/85b099c299c3ffd38ad9b3e43694c8a346934e4a30c88c4fd5a841234f77/aiohttp-3.14.3-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:e2667f0bbe7eb6c74eae5e9691441ad186e5845ca3cff63230fc09c4e7514f5d", size = 1876929, upload-time = "2026-07-23T01:56:08.413Z" },
+ { url = "https://files.pythonhosted.org/packages/d5/b7/1da684a04175473fa4cddbf9a2f572e79514c3fd27a74597f43057d4f3da/aiohttp-3.14.3-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:18cb43369747b2ae007bd2655fb8e63a099c2ff1d207962943636dac989b3147", size = 1761112, upload-time = "2026-07-23T01:56:10.918Z" },
+ { url = "https://files.pythonhosted.org/packages/d1/16/bc4b55e3e5cb175fd69c53c90d60d2f47797cb343da5106e23863dc4dba4/aiohttp-3.14.3-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:d77640cc618c1d99fc4f8589c0f24a730adfa54eb1e57ef7bf0c8dfb78da898c", size = 1583500, upload-time = "2026-07-23T01:56:13.613Z" },
+ { url = "https://files.pythonhosted.org/packages/2a/e8/13a9d957a1ee40837f46aa30f0f4c657e673ad86a2e6362a9f9be20d26d9/aiohttp-3.14.3-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:53e5179d8abb5710f8e83ba207c41c8d1261fcffd4616500e15ca2b7a33be10a", size = 1713940, upload-time = "2026-07-23T01:56:15.969Z" },
+ { url = "https://files.pythonhosted.org/packages/38/05/d33c680c1bcf1c7e130f9cbfc1fc02fe8bb0c4af2a94a53dd5fb56131e5c/aiohttp-3.14.3-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:cd817772b2fcf2b8c0905795318485f9ec16eae60b29feb7f4c77085311637f0", size = 1724413, upload-time = "2026-07-23T01:56:18.591Z" },
+ { url = "https://files.pythonhosted.org/packages/85/1d/af798d306f7a74b6a632dbcabcf62a4c91391b7582d2a8c6d7712e2cc54e/aiohttp-3.14.3-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:4e3ac92d90e92773b2362d506068e9a948192bd553e743c5b2429e28527c8661", size = 1770748, upload-time = "2026-07-23T01:56:21.074Z" },
+ { url = "https://files.pythonhosted.org/packages/a8/92/ad720d472556a995049206867765e9410969684f86ee09423ff9969044c1/aiohttp-3.14.3-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:3f42e9b78301f11c8f861746175d8b9c1ccef713fcad9eab396e2f6db8ed4a22", size = 1577564, upload-time = "2026-07-23T01:56:23.475Z" },
+ { url = "https://files.pythonhosted.org/packages/60/ad/0ed7586cbef7a884e23a752fa2bb987a122e6a5dd50dab109258d0a95193/aiohttp-3.14.3-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:9d9edccfe496b476db5f398d97b865e9a6752bcf8aec4eef8390ce20fb64bb41", size = 1782080, upload-time = "2026-07-23T01:56:25.994Z" },
+ { url = "https://files.pythonhosted.org/packages/97/ea/dbaed0d73e8a69aad653b045dab451c67c2454bb731a37b45a86593e9422/aiohttp-3.14.3-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:1c5ec8fb1bcc31a8466f74aaf26c345d5c386fa4bd08a3f0eb9c7a4a3fe8b5bf", size = 1745813, upload-time = "2026-07-23T01:56:28.604Z" },
+ { url = "https://files.pythonhosted.org/packages/81/1b/6893d4bc57e434fc93a6c9217c637d967a0b651d989f6e3265179375754a/aiohttp-3.14.3-cp314-cp314-win32.whl", hash = "sha256:38901a84da3ce22249f6e860bf8f90d141bcab7da090cc398f8bb58c0e44b7da", size = 455872, upload-time = "2026-07-23T01:56:31.031Z" },
+ { url = "https://files.pythonhosted.org/packages/f5/8b/c7baa1ba1eda4db6989baefe5de6d99834921b84ebd7918624febcb9f290/aiohttp-3.14.3-cp314-cp314-win_amd64.whl", hash = "sha256:8b3b60de05f3dcb6f6a00f818bb2ec781cee4de0645f59ccaf99b1d1823b6100", size = 481030, upload-time = "2026-07-23T01:56:33.365Z" },
+ { url = "https://files.pythonhosted.org/packages/22/8c/c29d067df825a2df88ca432db848aa2fe8199598359cc06c12b09320cac9/aiohttp-3.14.3-cp314-cp314-win_arm64.whl", hash = "sha256:1576145bdceeb92382d899751e12743a3a5b8e460a841e3e50543859e54864dc", size = 453669, upload-time = "2026-07-23T01:56:35.731Z" },
+ { url = "https://files.pythonhosted.org/packages/6a/a4/9c033beb355d39b6147980597ec9645e4729243f686ee4dc73945de72030/aiohttp-3.14.3-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:8800c996b01c2772a783e3e46f3e1abd5823029adca0df54231960de9bfefa5b", size = 791403, upload-time = "2026-07-23T01:56:37.972Z" },
+ { url = "https://files.pythonhosted.org/packages/80/ca/87c32a0a7704583cfc49660bd817889bae5b830bf53b5dcb4e92145ac2da/aiohttp-3.14.3-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:ebe8e504f058fe91223351cecd2d9d6946c9d241bb0250d898ffbdf584cc72b0", size = 526413, upload-time = "2026-07-23T01:56:40.523Z" },
+ { url = "https://files.pythonhosted.org/packages/9e/d8/8ec0e471248c500acdce2be3f46db8fb62b5eb60efef072529cc85ee1d26/aiohttp-3.14.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:30402d03a7c0ff52bce290b57e564e9079fd9d0cb545c8aba73f86a103162d2e", size = 532135, upload-time = "2026-07-23T01:56:42.876Z" },
+ { url = "https://files.pythonhosted.org/packages/fe/45/f8919fd936e8b79fcd9bda7b6d8e62613462a713f4f17987fd7c34399142/aiohttp-3.14.3-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9fc7b5bfec6573f3ae844f457fdde5adeb713f8b8e4a81ad64fc207b49383716", size = 1922742, upload-time = "2026-07-23T01:56:45.528Z" },
+ { url = "https://files.pythonhosted.org/packages/f6/ec/9ca76b28a27525b0cc53e20842e0228b022f301ce1f436b7d814b4aaf2df/aiohttp-3.14.3-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:8a5fd34f7f7410d1730d5c2ba873cacb2eed3fede366feb268a70ba22581ed8f", size = 1787371, upload-time = "2026-07-23T01:56:48.045Z" },
+ { url = "https://files.pythonhosted.org/packages/b1/04/6acdbf17315f7b55f1937e3387acb89a3cddeb4995689553d064af8e92ab/aiohttp-3.14.3-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:270d3dace9ca2f10f0da5d8ebe519b7a310fc6112ed916e32df5866df0888553", size = 1912623, upload-time = "2026-07-23T01:56:50.605Z" },
+ { url = "https://files.pythonhosted.org/packages/86/e6/438b0c79ca6f45eb9fd9817dd4c01a91919a38c0de5ee9e05e2b4dc0ece7/aiohttp-3.14.3-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:3ae5b3a59436d089b5395d910121a390feed4d00578eb95a0fd1a329fe963100", size = 2005515, upload-time = "2026-07-23T01:56:53.153Z" },
+ { url = "https://files.pythonhosted.org/packages/bb/6b/62cbd6577758699525f5c712d1ddef57d9875fbab0ae8d5f5a202fd598f8/aiohttp-3.14.3-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2498f0fe69ead802f9675beca44a7c21c62fdaa4ec5145ea1c3ad6edbee29f85", size = 1879906, upload-time = "2026-07-23T01:56:55.818Z" },
+ { url = "https://files.pythonhosted.org/packages/00/95/18bcbf830a21dc3aae24d8f6b6feaf3db1d2090242d00a7868db2ffb0b67/aiohttp-3.14.3-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:a0dc483c00da8b673abbb367eb6f8d8f4bcec30eb58529ea13cb42e7fd2dfa33", size = 1675849, upload-time = "2026-07-23T01:56:58.861Z" },
+ { url = "https://files.pythonhosted.org/packages/a9/19/47f4968659c5e23606c3790c80fc624e691c153d036148449ee84d31b287/aiohttp-3.14.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:c7d3a97c678d34fc5b59da671ee9cd630096ddc643e7b5a30d54a2a6f3574d3f", size = 1843496, upload-time = "2026-07-23T01:57:01.591Z" },
+ { url = "https://files.pythonhosted.org/packages/64/af/38c33c4dd82fddcb4e56c4653b6f1072a8edbc6b7fa15809f14932c41e2d/aiohttp-3.14.3-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:f8fb78a83c9e5f741ca3a68cfb455c1f5bb83b4e7249a3848b3cd78d0a8563b0", size = 1827746, upload-time = "2026-07-23T01:57:05.131Z" },
+ { url = "https://files.pythonhosted.org/packages/a1/9d/0537cda4885ac8f5b7053d164dd06312f4c483a4edcb8ee5b8aaf2a989bf/aiohttp-3.14.3-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:74ab5b6a9fb13e873e5a90946588baecaf488745e1db1a4a5c433f971f035098", size = 1853810, upload-time = "2026-07-23T01:57:08.043Z" },
+ { url = "https://files.pythonhosted.org/packages/19/fe/26f9c5e6458385aa86497836b0dea6fb2f027827d63f37c7856cce9286ee/aiohttp-3.14.3-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:bd52f811e65f6fb634b1047159657c98f52b407f8efec907bcfc09da9a4c0a25", size = 1668895, upload-time = "2026-07-23T01:57:10.837Z" },
+ { url = "https://files.pythonhosted.org/packages/ec/4c/618b1db9b9ba079b8875d2cdf78e7c4a3bf72903bd5850fee7dd9544600a/aiohttp-3.14.3-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:f0f177d1b195b9e06376cfd7d308d8a1b920909a609d03ac82a8c73bbb16d3b9", size = 1883833, upload-time = "2026-07-23T01:57:13.672Z" },
+ { url = "https://files.pythonhosted.org/packages/94/c6/bd959bd1e4771f9fd944e9e436224c48c77b018b73b519b5aad346335bcc/aiohttp-3.14.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:498c6c623134f8e09a3c4e60bcd607a0b4590dd7dbf08dd40851b27cbb520ccb", size = 1844251, upload-time = "2026-07-23T01:57:16.593Z" },
+ { url = "https://files.pythonhosted.org/packages/5e/19/08d41839658bdd44a0ed2480f3891705ecb487ce28c0dde62c9040c997e0/aiohttp-3.14.3-cp314-cp314t-win32.whl", hash = "sha256:b304db572b4368edd8dda8a2274f73156fe15558fca4a917cb8a09fc47af5963", size = 474180, upload-time = "2026-07-23T01:57:19.306Z" },
+ { url = "https://files.pythonhosted.org/packages/99/5d/3cd6ef0a2b2851f7ab913b5b079334781bd50ff56a323e4454063377a080/aiohttp-3.14.3-cp314-cp314t-win_amd64.whl", hash = "sha256:b20032766aedf6261c7a566585a40867d092ac03a0d81592d5370ef9b054f99b", size = 500528, upload-time = "2026-07-23T01:57:21.762Z" },
+ { url = "https://files.pythonhosted.org/packages/a4/37/cfd1ed540a4d318da025590d96b728e63713c09e9377950fc655dadeb856/aiohttp-3.14.3-cp314-cp314t-win_arm64.whl", hash = "sha256:2e1161602f45a54de2ce0905243a95f58cb42dcd378402f3697f5e0b21e9d2e7", size = 469280, upload-time = "2026-07-23T01:57:24.241Z" },
]
[[package]]
@@ -236,6 +287,19 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/af/77/43b27c14865dd4204ef353b875b4251e270b2518296e90b9bda479776c58/apswutils-0.1.2-py3-none-any.whl", hash = "sha256:9cd73744f9ae83c2e6f4337d4fcb092f5ea2f1814037e9ff7d953e2bc9c8362a", size = 48171, upload-time = "2025-12-18T06:24:31.312Z" },
]
+[[package]]
+name = "astunparse"
+version = "1.6.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "six" },
+ { name = "wheel" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f3/af/4182184d3c338792894f34a62672919db7ca008c89abee9b564dd34d8029/astunparse-1.6.3.tar.gz", hash = "sha256:5ad93a8456f0d084c3456d059fd9a92cce667963232cbf763eac3bc5b7940872", size = 18290, upload-time = "2019-12-22T18:12:13.129Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/2b/03/13dde6512ad7b4557eb792fbcf0c653af6076b81e5941d36ec61f7ce6028/astunparse-1.6.3-py2.py3-none-any.whl", hash = "sha256:c2652417f2c8b5bb325c885ae329bdf3f86424075c4fd1a128674bc6fba4b8e8", size = 12732, upload-time = "2019-12-22T18:12:11.297Z" },
+]
+
[[package]]
name = "attrs"
version = "26.1.0"
@@ -343,7 +407,8 @@ dependencies = [
{ name = "tokenizers" },
{ name = "torch" },
{ name = "transformers" },
- { name = "triton" },
+ { name = "triton", version = "3.6.0", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform == 'win32'" },
+ { name = "triton", version = "3.7.1", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform != 'win32'" },
{ name = "typing-extensions" },
{ name = "zstandard" },
]
@@ -373,6 +438,15 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/5f/af/0a6e1d2a845988039f6c197fa7269b5e9abbe17354fb41cc9d75bb260fcb/av-17.0.1-cp314-cp314t-win_arm64.whl", hash = "sha256:a87a42c36e29f75e7dff7281944f2a6876a2c8875e225ccbf6c1ae62748b4caa", size = 22072676, upload-time = "2026-04-18T17:12:31.836Z" },
]
+[[package]]
+name = "backports-strenum"
+version = "1.3.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/35/c7/2ed54c32fed313591ffb21edbd48db71e68827d43a61938e5a0bc2b6ec91/backports_strenum-1.3.1.tar.gz", hash = "sha256:77c52407342898497714f0596e86188bb7084f89063226f4ba66863482f42414", size = 7257, upload-time = "2023-12-09T14:36:40.937Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/d6/50/56cf20e2ee5127b603b81d5a69580a1a325083e2b921aa8f067da83927c0/backports_strenum-1.3.1-py3-none-any.whl", hash = "sha256:cdcfe36dc897e2615dc793b7d3097f54d359918fc448754a517e6f23044ccf83", size = 8304, upload-time = "2023-12-09T14:36:39.905Z" },
+]
+
[[package]]
name = "beautifulsoup4"
version = "4.14.3"
@@ -734,7 +808,7 @@ name = "cuda-bindings"
version = "13.2.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "cuda-pathfinder" },
+ { name = "cuda-pathfinder", marker = "sys_platform != 'win32'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/52/c8/b2589d68acf7e3d63e2be330b84bc25712e97ed799affbca7edd7eae25d6/cuda_bindings-13.2.0-cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e865447abfb83d6a98ad5130ed3c70b1fc295ae3eeee39fd07b4ddb0671b6788", size = 5722404, upload-time = "2026-03-11T00:12:44.041Z" },
@@ -757,45 +831,51 @@ wheels = [
[[package]]
name = "cuda-toolkit"
-version = "13.0.2"
+version = "13.0.3.0"
source = { registry = "https://pypi.org/simple" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/57/b2/453099f5f3b698d7d0eab38916aac44c7f76229f451709e2eb9db6615dcd/cuda_toolkit-13.0.2-py2.py3-none-any.whl", hash = "sha256:b198824cf2f54003f50d64ada3a0f184b42ca0846c1c94192fa269ecd97a66eb", size = 2364, upload-time = "2025-12-19T23:24:07.328Z" },
+ { url = "https://files.pythonhosted.org/packages/d1/c7/a79086a62c98befcdb8349656c6f114e2db3b8b2422f6e25c97a7f2a9a3c/cuda_toolkit-13.0.3.0-py2.py3-none-any.whl", hash = "sha256:d693caaa261214ddd7dbb60d68e71cbed884e68c2be7509778f3051da0b91c3f", size = 2512, upload-time = "2026-04-14T00:50:08.173Z" },
]
[package.optional-dependencies]
cublas = [
- { name = "nvidia-cublas", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-cublas", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+ { name = "nvidia-cuda-nvrtc", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
cudart = [
- { name = "nvidia-cuda-runtime", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-cuda-runtime", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
cufft = [
- { name = "nvidia-cufft", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-cufft", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+ { name = "nvidia-nvjitlink", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
cufile = [
- { name = "nvidia-cufile", marker = "sys_platform == 'linux'" },
+ { name = "nvidia-cufile", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
cupti = [
- { name = "nvidia-cuda-cupti", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-cuda-cupti", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
curand = [
- { name = "nvidia-curand", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-curand", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
cusolver = [
- { name = "nvidia-cusolver", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-cublas", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+ { name = "nvidia-cusolver", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+ { name = "nvidia-cusparse", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+ { name = "nvidia-nvjitlink", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
cusparse = [
- { name = "nvidia-cusparse", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-cusparse", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+ { name = "nvidia-nvjitlink", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
nvjitlink = [
- { name = "nvidia-nvjitlink", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-nvjitlink", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
nvrtc = [
- { name = "nvidia-cuda-nvrtc", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-cuda-nvrtc", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
nvtx = [
- { name = "nvidia-nvtx", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+ { name = "nvidia-nvtx", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
]
[[package]]
@@ -855,7 +935,7 @@ wheels = [
[[package]]
name = "diffusers"
-version = "0.37.1"
+version = "0.39.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "filelock" },
@@ -868,9 +948,9 @@ dependencies = [
{ name = "requests" },
{ name = "safetensors" },
]
-sdist = { url = "https://files.pythonhosted.org/packages/46/5c/f4c2eb8d481fe8784a7e2331fbaab820079c06676185fa6d2177b386d590/diffusers-0.37.1.tar.gz", hash = "sha256:2346c21f77f835f273b7aacbaada1c34a596a3a2cc6ddc99d149efcd0ec298fa", size = 4135139, upload-time = "2026-03-25T08:04:04.515Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/1a/81/6095237b86a3116c4789f28c4435d5296c00c0fc74ffde99008fd6b3a36c/diffusers-0.39.0.tar.gz", hash = "sha256:14bb1d98c85a0e463d734c99aaa73b480a7bc9bad22af30fbf730ef8f09c1d67", size = 4651240, upload-time = "2026-07-03T08:48:47.904Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/9c/dd/51c38785ce5e1c287b5ad17ba550edaaaffce0deb0da4857019c6700fbaf/diffusers-0.37.1-py3-none-any.whl", hash = "sha256:0537c0b28cb53cf39d6195489bcf8f833986df556c10f5e28ab7427b86fc8b90", size = 5001536, upload-time = "2026-03-25T08:04:02.385Z" },
+ { url = "https://files.pythonhosted.org/packages/3f/3f/7469c46e9d22307ea686bab687d70e6bf328722952f9d10339f5e913e608/diffusers-0.39.0-py3-none-any.whl", hash = "sha256:912aca51b5787365110806e984d5555735bf8a461073bb8459029d0bca7870ef", size = 5631176, upload-time = "2026-07-03T08:48:45.337Z" },
]
[[package]]
@@ -921,18 +1001,9 @@ dependencies = [
]
sdist = { url = "https://files.pythonhosted.org/packages/62/59/e47bbd0542d0e6f4ce9983d5eb458a01d4b42c81e5c410cb9e159b1061ae/encodec-0.1.1.tar.gz", hash = "sha256:36dde98ccfe6c51a15576476cadfcb3b35a63507b8b8555abd69889a6fba6772", size = 3736037, upload-time = "2022-10-25T16:13:21.471Z" }
-[[package]]
-name = "execnet"
-version = "2.1.2"
-source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/bf/89/780e11f9588d9e7128a3f87788354c7946a9cbb1401ad38a48c4db9a4f07/execnet-2.1.2.tar.gz", hash = "sha256:63d83bfdd9a23e35b9c6a3261412324f964c2ec8dcd8d3c6916ee9373e0befcd", size = 166622, upload-time = "2025-11-12T09:56:37.75Z" }
-wheels = [
- { url = "https://files.pythonhosted.org/packages/ab/84/02fc1827e8cdded4aa65baef11296a9bbe595c474f0d6d758af082d849fd/execnet-2.1.2-py3-none-any.whl", hash = "sha256:67fba928dd5a544b783f6056f449e5e3931a5c378b128bc18501f7ea79e296ec", size = 40708, upload-time = "2025-11-12T09:56:36.333Z" },
-]
-
[[package]]
name = "executorch"
-version = "1.2.0"
+version = "1.4.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "coremltools", marker = "(python_full_version < '3.14' and sys_platform == 'darwin') or (python_full_version < '3.14' and sys_platform == 'linux')" },
@@ -947,10 +1018,6 @@ dependencies = [
{ name = "packaging", marker = "python_full_version < '3.14'" },
{ name = "pandas", marker = "python_full_version < '3.14'" },
{ name = "parameterized", marker = "python_full_version < '3.14'" },
- { name = "pytest", version = "8.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.14'" },
- { name = "pytest-json-report", marker = "python_full_version < '3.14'" },
- { name = "pytest-rerunfailures", marker = "python_full_version < '3.14'" },
- { name = "pytest-xdist", marker = "python_full_version < '3.14'" },
{ name = "pytorch-tokenizers", marker = "python_full_version < '3.14'" },
{ name = "pyyaml", marker = "python_full_version < '3.14'" },
{ name = "ruamel-yaml", marker = "python_full_version < '3.14'" },
@@ -962,14 +1029,18 @@ dependencies = [
{ name = "typing-extensions", marker = "python_full_version < '3.14'" },
]
wheels = [
- { url = "https://files.pythonhosted.org/packages/38/2a/7b624076e67d71a82b68f07fad90fc9b1e6f3f887460f915cd665d0a27b3/executorch-1.2.0-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:ea6848902acabf21c7a6fe078947f50c386bacd9081161761bff27931b3e2af1", size = 12027466, upload-time = "2026-04-01T21:10:32.961Z" },
- { url = "https://files.pythonhosted.org/packages/a0/59/55a707e4a97667b4e3c03c60d4883891c482e033edd32f3d30966abdb358/executorch-1.2.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:ebcf508bafed6a7eb4f186459ccc921aabd7b1f6c13a2045ed16654075b58a36", size = 12828287, upload-time = "2026-04-01T21:10:35.672Z" },
- { url = "https://files.pythonhosted.org/packages/11/27/d88311cad6ce5181562fb6cc2123ff197d774e8ec236a2d5a5f0b6bdb361/executorch-1.2.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:44090bb8361de77ae551efae045ec577abfb9a4b7f42365627768d2e5735f542", size = 20526090, upload-time = "2026-04-01T21:10:39.094Z" },
- { url = "https://files.pythonhosted.org/packages/6c/eb/92ace5cfefbba29217137edabf5ec68b2b60b9e00ae3cc7ecebf79973d5b/executorch-1.2.0-cp312-cp312-win_amd64.whl", hash = "sha256:1e0a885d8ac567c9ff9cf97902745e5090d91b12cd57921edcd3b8cd5a1e2393", size = 9705134, upload-time = "2026-04-01T21:10:42.509Z" },
- { url = "https://files.pythonhosted.org/packages/9e/a1/fc395be78ae85cd3c19f1436322dc4e8622317cf8ed2ece1155f0ad39e24/executorch-1.2.0-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:9e14736f2357aaec6930eda98dc870b08e78ece604a3cdb11fe66d493e6c68c3", size = 12027961, upload-time = "2026-04-01T21:10:45.57Z" },
- { url = "https://files.pythonhosted.org/packages/bd/43/8b94829c2df0ed57d4286afc8ff2fe54d7108c378c02ecd5190f50c6e74e/executorch-1.2.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:148fc9ac6e477fd3421f2795fb9918a0e1c0136d54f50c229f77a5bed6d3138c", size = 12829767, upload-time = "2026-04-01T21:10:48.53Z" },
- { url = "https://files.pythonhosted.org/packages/a4/c5/d3b77ec2cc6ff4550ffd63f0154e000670730fad5134b4df80419afb89e6/executorch-1.2.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:b93a56a7e080ecc8b0648662d39b0d23d74554e475e97960a8c1ff487488b4c0", size = 20526098, upload-time = "2026-04-01T21:10:52.116Z" },
- { url = "https://files.pythonhosted.org/packages/07/7b/18fbc1f0f21401e013b91fca3f67d4ae2c426a1ede84a88846ddb4322185/executorch-1.2.0-cp313-cp313-win_amd64.whl", hash = "sha256:c66a5289f9319a64c9a962432aa9cacf4ffa0ac8c99f5f2a694408ebf5102a02", size = 9705167, upload-time = "2026-04-01T21:10:55.374Z" },
+ { url = "https://files.pythonhosted.org/packages/3a/40/067a21f1ca501fda2f673b410a766309b0641042e8d2b93f614b78f61044/executorch-1.4.0-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:1391236f2b72815004e6b43d1073cb62668a7588f640662dafce92f4b5f6996d", size = 16668345, upload-time = "2026-08-07T19:53:28.415Z" },
+ { url = "https://files.pythonhosted.org/packages/f5/06/c37c63ed08a0d7717d90e078b36768c6a300a40bd7fa3fa61efa25247b45/executorch-1.4.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:29871cde0ec60431f1056c531e7903616320695f4036f503827f3a81a4e542fd", size = 16074566, upload-time = "2026-08-07T19:53:31.095Z" },
+ { url = "https://files.pythonhosted.org/packages/1c/55/afb80f13ffd785b1235718c74d2fda17748b53bc111b4c3168d9c2055ca6/executorch-1.4.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:fb13ebf8fab5b5c7ef586fa7b347ecd38be8fd42684545193a495ecb21287b55", size = 17567549, upload-time = "2026-08-07T19:53:33.751Z" },
+ { url = "https://files.pythonhosted.org/packages/06/cd/5235f6a5116005ac68ae3c7a12158e8973a0052a391f26d055cc0cfd9320/executorch-1.4.0-cp312-cp312-win_amd64.whl", hash = "sha256:4c5bb4a2f3e30e7b9819b75bd8f41e083922d341d9147730fda973f8dbf1ef7d", size = 12803395, upload-time = "2026-08-07T19:53:36.471Z" },
+ { url = "https://files.pythonhosted.org/packages/86/79/b23fb1ed45f9d8ac0193b0209677a1398a73f5638aee16dd8a67a19ac02c/executorch-1.4.0-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:7677c6808c10831b0ee3d534ad4c9acb038ea4f50f8c6feb5e5ef4c6c7e25c7f", size = 16668640, upload-time = "2026-08-07T19:53:38.997Z" },
+ { url = "https://files.pythonhosted.org/packages/13/ab/1fb5bd2d3577ecad6a4b0b25b35f23e952ce8b6ae6bddc0df768ff90ba38/executorch-1.4.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:8a6647beda101543eff7f1ed6568f907dc6edcf85451eb45c20c771aff254eee", size = 16074966, upload-time = "2026-08-07T19:53:41.452Z" },
+ { url = "https://files.pythonhosted.org/packages/13/0f/fef87d44aafb6afd8db087e8b84aae784d0977c7fff4ba10f7298193dc17/executorch-1.4.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:79f8418ed0f86c5591940d4c3044934444d63add39251b3e843f2b695d7e6426", size = 17568357, upload-time = "2026-08-07T19:53:43.846Z" },
+ { url = "https://files.pythonhosted.org/packages/bc/4e/f4461f32e59f783f264ab243da216b589197fff2defbf8419dafbae99f84/executorch-1.4.0-cp313-cp313-win_amd64.whl", hash = "sha256:77549dd42496fd3e2af48d4b0dee33c559f6d6ee480bf17ca32355be091ceada", size = 12803501, upload-time = "2026-08-07T19:53:46.404Z" },
+ { url = "https://files.pythonhosted.org/packages/87/6d/c5f203bf78342a93d6374cf28fb66ba082f27748c0f760e034642f3d01c5/executorch-1.4.0-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:b568c66abb381ae95e9a2aab7d9459687fa54c20c374b49f6d24682816afb025", size = 16670480, upload-time = "2026-08-07T19:53:48.904Z" },
+ { url = "https://files.pythonhosted.org/packages/74/68/b8149335b8e6ab57f257b333d1ec5f5d16aecdc7ea205e07ea5cde5d7285/executorch-1.4.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:1e5576141b755aad344184326266561180040a0c86dbb30ac3c77e21f720c245", size = 16078636, upload-time = "2026-08-07T19:53:51.375Z" },
+ { url = "https://files.pythonhosted.org/packages/70/36/1974f95a5ae0b853262a41cb2acccce2cdaf2f7e0289fb1c081371811faf/executorch-1.4.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:cd24213687eac2ba5f255d3b0a02420879bae3e86666c372d018c38d72e37921", size = 17568993, upload-time = "2026-08-07T19:53:53.789Z" },
+ { url = "https://files.pythonhosted.org/packages/3a/8e/6f46aa84c676962926e1e5de260314d0ad177298d7450958e520e12eb663/executorch-1.4.0-cp314-cp314-win_amd64.whl", hash = "sha256:bf76414dc0062e2d4f42ea8a226b26977550e80930c35ee718f9fbf50fc37612", size = 12805319, upload-time = "2026-08-07T19:53:56.834Z" },
]
[[package]]
@@ -1193,9 +1264,71 @@ http = [
{ name = "aiohttp" },
]
+[[package]]
+name = "gast"
+version = "0.7.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/91/f6/e73969782a2ecec280f8a176f2476149dd9dba69d5f8779ec6108a7721e6/gast-0.7.0.tar.gz", hash = "sha256:0bb14cd1b806722e91ddbab6fb86bba148c22b40e7ff11e248974e04c8adfdae", size = 33630, upload-time = "2025-11-29T15:30:05.266Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/1d/33/f1c6a276de27b7d7339a34749cc33fa87f077f921969c47185d34a887ae2/gast-0.7.0-py3-none-any.whl", hash = "sha256:99cbf1365633a74099f69c59bd650476b96baa5ef196fec88032b00b31ba36f7", size = 22966, upload-time = "2025-11-29T15:30:03.983Z" },
+]
+
+[[package]]
+name = "google-pasta"
+version = "0.2.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "six" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/35/4a/0bd53b36ff0323d10d5f24ebd67af2de10a1117f5cf4d7add90df92756f1/google-pasta-0.2.0.tar.gz", hash = "sha256:c9f2c8dfc8f96d0d5808299920721be30c9eec37f2389f28904f454565c8a16e", size = 40430, upload-time = "2020-03-13T18:57:50.34Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/a3/de/c648ef6835192e6e2cc03f40b19eeda4382c49b5bafb43d88b931c4c74ac/google_pasta-0.2.0-py3-none-any.whl", hash = "sha256:b32482794a366b5366a32c92a9a9201b107821889935a02b3e51f6b432ea84ed", size = 57471, upload-time = "2020-03-13T18:57:48.872Z" },
+]
+
+[[package]]
+name = "grpcio"
+version = "1.81.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b0/b5/1ff353970a87eda4c98251e34d2dfd214abd4982dc89119c9252a2a482d2/grpcio-1.81.1.tar.gz", hash = "sha256:6fa10a767143a5e82e8eaab53918af0cd8909a57a27f8cb2288b80a613ac671b", size = 13026582, upload-time = "2026-06-11T12:46:51.673Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/85/07/9a979c81738863a738dc23d65177056e71fbb2db817740ed870b33434e7a/grpcio-1.81.1-cp312-cp312-linux_armv7l.whl", hash = "sha256:8b39472beafc0bdcafc4c8c73ad082ebfdb449d566897a61e7acb4fa88089115", size = 6053264, upload-time = "2026-06-11T12:45:21.017Z" },
+ { url = "https://files.pythonhosted.org/packages/75/95/539706ca0d3bd40dbad583dc56fd883da941f37556b629132da5762781b9/grpcio-1.81.1-cp312-cp312-macosx_11_0_universal2.whl", hash = "sha256:12b7524c88d4026d3dcb7b0ebe16b6714f3b4af402ddd0f0639ab064a00c87c3", size = 12052560, upload-time = "2026-06-11T12:45:23.652Z" },
+ { url = "https://files.pythonhosted.org/packages/e0/44/f257b7e0bd69c93b06c6cb8ac8d1b901ccb42bedabd83c1a4c77a71f8810/grpcio-1.81.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:1e123f9b37edb8375fd74130d1f69c944bbf0a7b06761ae7211154b8759e94d2", size = 6595983, upload-time = "2026-06-11T12:45:26.963Z" },
+ { url = "https://files.pythonhosted.org/packages/b9/f3/19782aa04c960968bef8c5539329d8e3bbc3364e2e46d19eb5e5cc5e43b7/grpcio-1.81.1-cp312-cp312-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:2c2e2ae6867c2966b8daccc836d54a13218e0007e9a490aeb81dd05be64d22d7", size = 7303455, upload-time = "2026-06-11T12:45:29.707Z" },
+ { url = "https://files.pythonhosted.org/packages/eb/8c/dea020b6d91508cd84463917a63149ec196ee7db505d032ae43fcb3303b9/grpcio-1.81.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:766bc7c9a9c340342f4c864ccbda8e78111e4751f13b895812b9c148fb79e9d0", size = 6809167, upload-time = "2026-06-11T12:45:32.52Z" },
+ { url = "https://files.pythonhosted.org/packages/1c/c7/3030dd940408083bd32cd95d634777a71605ade4887154d93e8a89244946/grpcio-1.81.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:b259a04a737cb3496be0901328eb8b7552ed8df4865d8c8f1cf1bffcfc0776a3", size = 7412536, upload-time = "2026-06-11T12:45:35.403Z" },
+ { url = "https://files.pythonhosted.org/packages/e0/dd/1172a9e42b168edcafefad6115346ef619a3fc02158bb170e66ced24bcdd/grpcio-1.81.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:85b10a45b8993d195c4f3ff57025b8d1e11834909ee475c403bfa60cb4caefaf", size = 8408276, upload-time = "2026-06-11T12:45:37.78Z" },
+ { url = "https://files.pythonhosted.org/packages/25/7a/71437c7f3596e5246155c515852795a85a1a8d228190212432b13b97a95d/grpcio-1.81.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:8ea1936c26b99999b27479853039a7f34713f56c49375ad52b38535ec93a796c", size = 7849660, upload-time = "2026-06-11T12:45:40.627Z" },
+ { url = "https://files.pythonhosted.org/packages/65/40/7debc0da45d2efebafb82da75644be347497fe4ee250514b8cd3b86ae8bf/grpcio-1.81.1-cp312-cp312-win32.whl", hash = "sha256:a185a04039df6cae8648bc8ab6d6fde7bf94f7188ecf7828e76ac52eef1e41d6", size = 4185819, upload-time = "2026-06-11T12:45:43.027Z" },
+ { url = "https://files.pythonhosted.org/packages/2e/b9/8fe3ba5ed462067774ebc1f9c7f26aa7ebcc280ddd476be107153de1339e/grpcio-1.81.1-cp312-cp312-win_amd64.whl", hash = "sha256:3ad74f8bb1a18963914c5452d289422830b39459e8776ebbcd207be1fbfb1d94", size = 4930461, upload-time = "2026-06-11T12:45:45.775Z" },
+ { url = "https://files.pythonhosted.org/packages/7a/42/dcc2e4b600538ef18327c0839d56b7d3c3812337c5d710df5877dbb39b1e/grpcio-1.81.1-cp313-cp313-linux_armv7l.whl", hash = "sha256:b10e1ff4756ed27d5a29d7fc79cfce7ef1ff56ad20025b89bac7cf79e09abbbe", size = 6054466, upload-time = "2026-06-11T12:45:48.43Z" },
+ { url = "https://files.pythonhosted.org/packages/7b/4a/a36e03210183a8a7d4c80c3936acee679f4bd77d5861f369db47b2cc5f05/grpcio-1.81.1-cp313-cp313-macosx_11_0_universal2.whl", hash = "sha256:819edbdcb42ab8598b494bcf0222684bbb7a3c772bd1b1f0be7e029a6063c28e", size = 12048795, upload-time = "2026-06-11T12:45:54.011Z" },
+ { url = "https://files.pythonhosted.org/packages/b0/d5/d68e30b29098f63beab6fe501100fe82674ff142b32c672532da86a99b3a/grpcio-1.81.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:c5bf2dc311127d91230cc79b92188c082634a06cf66c5234db49a43b910183b0", size = 6599094, upload-time = "2026-06-11T12:45:57.799Z" },
+ { url = "https://files.pythonhosted.org/packages/3d/b3/e837954d279754f638a11cca5dcf6b24a005efb398984cefaf7735945a54/grpcio-1.81.1-cp313-cp313-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:e8ca6a1fcdb2943c9cbc1804a1baf3acb6071d72a471591678ded84218006e14", size = 7307182, upload-time = "2026-06-11T12:46:00.568Z" },
+ { url = "https://files.pythonhosted.org/packages/0d/1e/b47957057e729adc6cdf519a47f8be2562b7140e280f1418443eb4022192/grpcio-1.81.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:e64dd101d380a115cc5a0c7856788adb535f1a4e21fc543775602f8be95180ae", size = 6810962, upload-time = "2026-06-11T12:46:03.312Z" },
+ { url = "https://files.pythonhosted.org/packages/40/26/569868e364e05b19ec8f969da53d230bcd89c962cd198f7c29943155c4d3/grpcio-1.81.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:98a07f9bf591e3a8919797bee1c53f026ba4acd587e5a4404c8e57c9ec36b2a5", size = 7415698, upload-time = "2026-06-11T12:46:06.005Z" },
+ { url = "https://files.pythonhosted.org/packages/36/0c/5440a0582cb5653fc42a6e262eeb22700943313f8076f9dc927491b20a59/grpcio-1.81.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:c261d74b1a945cf895a9d6eccd1685a8e837531beaab782da4d630a8d12deffb", size = 8407779, upload-time = "2026-06-11T12:46:08.84Z" },
+ { url = "https://files.pythonhosted.org/packages/ff/aa/66fe9f39871d766987d869a03ee0842a026f499c7b1e62decb9e78a8088e/grpcio-1.81.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:58ad1131c300d3c9b933802b3cc4dc69d380822935ba50b28703156ea826fbf7", size = 7844521, upload-time = "2026-06-11T12:46:12.171Z" },
+ { url = "https://files.pythonhosted.org/packages/f0/9e/69bb7194861bcd28fb3193261d4f9c3831b4446993f002cf59068943e7ab/grpcio-1.81.1-cp313-cp313-win32.whl", hash = "sha256:78e29211f26da2fdd0e9c6d2b79f489476140cf7029b6a64808ade7ca4156a42", size = 4182786, upload-time = "2026-06-11T12:46:15.192Z" },
+ { url = "https://files.pythonhosted.org/packages/0d/20/3da8bb0d637feccdc3e1e419bb511ce93651ce7d54164f95de22cc0b8b34/grpcio-1.81.1-cp313-cp313-win_amd64.whl", hash = "sha256:edb59506291b647a30884b1d51a599d605f40b20af4a7dc3d33786a47a31de60", size = 4928648, upload-time = "2026-06-11T12:46:17.823Z" },
+ { url = "https://files.pythonhosted.org/packages/b6/58/19414622b1bf6981bc9c05a365bd548e71876c89000083b3af489251e9c0/grpcio-1.81.1-cp314-cp314-linux_armv7l.whl", hash = "sha256:506f48f2f9c29b143fca3dad7b0d518c188b6c9648c75a2ae6e2d9f2c13a060b", size = 6055336, upload-time = "2026-06-11T12:46:20.557Z" },
+ { url = "https://files.pythonhosted.org/packages/32/f1/2ec88adb92b0eba970dd0e0e7dd086341daa3c75eba4f735f9e44bf684b0/grpcio-1.81.1-cp314-cp314-macosx_11_0_universal2.whl", hash = "sha256:d865db4a6318e1c1bea83292e0ed231090538fc4ca45425b0f0480eb338bbc6e", size = 12056279, upload-time = "2026-06-11T12:46:24.255Z" },
+ { url = "https://files.pythonhosted.org/packages/41/36/e8c5f8c6ec71de73733695ebc809e98b178b534ec6d8eaa31a7ebab4ad4c/grpcio-1.81.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:e2aa72e3ce1770317ef534f63d397b55e130725f5149bd36077c3b539019db27", size = 6608225, upload-time = "2026-06-11T12:46:27.601Z" },
+ { url = "https://files.pythonhosted.org/packages/30/22/96fc577a845ab093326d9ab1adb874bd4936c8cf98ac8ed2f3db13a0a2fb/grpcio-1.81.1-cp314-cp314-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:0490c30c261eded63f3f354979f9dc4502a9fb944cccb60cd9dc85f5a7349854", size = 7306576, upload-time = "2026-06-11T12:46:30.514Z" },
+ { url = "https://files.pythonhosted.org/packages/76/7b/61dab5d5969f28d97fb1009cead1df0a5cd987d3315e1b37f18a4449f8bc/grpcio-1.81.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:410482da976329fe5f4067270401b12cf2bd552ff8020f054ecfaddb5475f9d6", size = 6812165, upload-time = "2026-06-11T12:46:33.699Z" },
+ { url = "https://files.pythonhosted.org/packages/82/78/6e501929d4f5f96462fd82fd9f0f06e5f9612207582b862868d68757b27d/grpcio-1.81.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:e3657301562ac3cb8018d30d0d3ebfa39932239f7b5703422057ef14b69949f5", size = 7422962, upload-time = "2026-06-11T12:46:36.511Z" },
+ { url = "https://files.pythonhosted.org/packages/2a/7e/f2157589e66daa78ebb3165942d05a08bdea93b9d11c2bc1e172aef89685/grpcio-1.81.1-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:24c8e57504c8f45b237e40b99262d181071e5099a07053695b75d97bb53053a0", size = 8408176, upload-time = "2026-06-11T12:46:39.803Z" },
+ { url = "https://files.pythonhosted.org/packages/da/df/c6717fef716e00d235ffb96123baf6dce76d6004f6233fa767c502861460/grpcio-1.81.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:b427c19380991a4eaab2f6144b64b99b412043314c6bf4ab544f97bb31ee4190", size = 7846681, upload-time = "2026-06-11T12:46:43.013Z" },
+ { url = "https://files.pythonhosted.org/packages/36/84/3502e9f210a6a5c4438c8aca3f88edd2e04f6a27f3d41b26cf0a0024b096/grpcio-1.81.1-cp314-cp314-win32.whl", hash = "sha256:61233fe8951e5c85dff81c2458b6528624760166946b5b47ea150a589168411f", size = 4264615, upload-time = "2026-06-11T12:46:45.741Z" },
+ { url = "https://files.pythonhosted.org/packages/ff/b0/4af731ff7492c68a96e4c71bfd0f4590acde92b31c6fe4894e6465c10ff6/grpcio-1.81.1-cp314-cp314-win_amd64.whl", hash = "sha256:3768a5ff1b2125e6f552e561b6b2dca0e64982d8949689b4df145cf8b98d7821", size = 5070275, upload-time = "2026-06-11T12:46:48.486Z" },
+]
+
[[package]]
name = "gst-python-ml"
-version = "1.1.0"
+version = "1.1.1"
source = { editable = "." }
dependencies = [
{ name = "accelerate" },
@@ -1228,6 +1361,17 @@ dependencies = [
]
[package.optional-dependencies]
+all = [
+ { name = "ai-edge-litert" },
+ { name = "faster-whisper" },
+ { name = "llama-cpp-python" },
+ { name = "onnxruntime" },
+ { name = "openvino" },
+ { name = "pysilero" },
+ { name = "pysilero-vad" },
+ { name = "tensorflow" },
+ { name = "tinygrad" },
+]
executorch = [
{ name = "executorch", marker = "python_full_version < '3.14'" },
]
@@ -1240,6 +1384,9 @@ jax-gpu = [
jax-tpu = [
{ name = "jax", extra = ["tpu"] },
]
+litert = [
+ { name = "ai-edge-litert" },
+]
llamacpp = [
{ name = "llama-cpp-python" },
]
@@ -1253,6 +1400,12 @@ onnx = [
onnx-gpu = [
{ name = "onnxruntime-gpu" },
]
+openvino = [
+ { name = "openvino" },
+]
+tensorflow = [
+ { name = "tensorflow" },
+]
tinygrad = [
{ name = "tinygrad" },
]
@@ -1264,17 +1417,16 @@ vad = [
[package.dev-dependencies]
dev = [
- { name = "pytest", version = "8.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.14'" },
- { name = "pytest", version = "9.0.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.14'" },
+ { name = "pytest" },
]
test = [
- { name = "pytest", version = "8.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.14'" },
- { name = "pytest", version = "9.0.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.14'" },
+ { name = "pytest" },
]
[package.metadata]
requires-dist = [
{ name = "accelerate", specifier = ">=1.13.0" },
+ { name = "ai-edge-litert", marker = "extra == 'litert'" },
{ name = "autoawq", specifier = ">=0.2.9" },
{ name = "bitsandbytes", specifier = ">=0.49.2" },
{ name = "confluent-kafka" },
@@ -1282,6 +1434,7 @@ requires-dist = [
{ name = "easydict" },
{ name = "executorch", marker = "python_full_version < '3.14' and extra == 'executorch'" },
{ name = "faster-whisper", marker = "extra == 'vad'" },
+ { name = "gst-python-ml", extras = ["onnx", "tinygrad", "llamacpp", "openvino", "tensorflow", "litert", "vad"], marker = "extra == 'all'" },
{ name = "huggingface-hub" },
{ name = "jax", extras = ["cpu"], marker = "extra == 'jax-cpu'" },
{ name = "jax", extras = ["cuda12"], marker = "extra == 'jax-gpu'" },
@@ -1297,6 +1450,7 @@ requires-dist = [
{ name = "opencv-contrib-python", marker = "python_full_version >= '3.14'", specifier = ">=4.13.0" },
{ name = "opencv-python", marker = "python_full_version < '3.14'", specifier = ">=4.9.0" },
{ name = "opencv-python", marker = "python_full_version >= '3.14'", specifier = ">=4.13.0" },
+ { name = "openvino", marker = "extra == 'openvino'", specifier = ">=2024.0" },
{ name = "protobuf" },
{ name = "pycairo" },
{ name = "pyflann-py3" },
@@ -1308,6 +1462,7 @@ requires-dist = [
{ name = "sentencepiece" },
{ name = "soundfile" },
{ name = "speechbrain" },
+ { name = "tensorflow", marker = "extra == 'tensorflow'", specifier = ">=2.16.0" },
{ name = "tinygrad", marker = "extra == 'tinygrad'" },
{ name = "torch", marker = "python_full_version < '3.14'", specifier = ">=2.7.0" },
{ name = "torch", marker = "python_full_version >= '3.14'", specifier = ">=2.11.0" },
@@ -1321,11 +1476,11 @@ requires-dist = [
{ name = "webdataset" },
{ name = "whisperspeech" },
]
-provides-extras = ["onnx", "onnx-gpu", "tinygrad", "mlx", "executorch", "llamacpp", "jax-cpu", "jax-gpu", "jax-tpu", "vad"]
+provides-extras = ["onnx", "onnx-gpu", "tinygrad", "mlx", "executorch", "llamacpp", "jax-cpu", "jax-gpu", "jax-tpu", "openvino", "tensorflow", "litert", "vad", "all"]
[package.metadata.requires-dev]
-dev = [{ name = "pytest", specifier = ">=7.0" }]
-test = [{ name = "pytest", specifier = ">=7.0" }]
+dev = [{ name = "pytest", specifier = ">=9.0.3" }]
+test = [{ name = "pytest", specifier = ">=9.0.3" }]
[[package]]
name = "h11"
@@ -1336,6 +1491,27 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/04/4b/29cac41a4d98d144bf5f6d33995617b185d14b22401f75ca86f384e87ff1/h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86", size = 37515, upload-time = "2025-04-24T03:35:24.344Z" },
]
+[[package]]
+name = "h5py"
+version = "3.14.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/5d/57/dfb3c5c3f1bf5f5ef2e59a22dec4ff1f3d7408b55bfcefcfb0ea69ef21c6/h5py-3.14.0.tar.gz", hash = "sha256:2372116b2e0d5d3e5e705b7f663f7c8d96fa79a4052d250484ef91d24d6a08f4", size = 424323, upload-time = "2025-06-06T14:06:15.01Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/3e/77/8f651053c1843391e38a189ccf50df7e261ef8cd8bfd8baba0cbe694f7c3/h5py-3.14.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:e0045115d83272090b0717c555a31398c2c089b87d212ceba800d3dc5d952e23", size = 3312740, upload-time = "2025-06-06T14:05:01.193Z" },
+ { url = "https://files.pythonhosted.org/packages/ff/10/20436a6cf419b31124e59fefc78d74cb061ccb22213226a583928a65d715/h5py-3.14.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:6da62509b7e1d71a7d110478aa25d245dd32c8d9a1daee9d2a42dba8717b047a", size = 2829207, upload-time = "2025-06-06T14:05:05.061Z" },
+ { url = "https://files.pythonhosted.org/packages/3f/19/c8bfe8543bfdd7ccfafd46d8cfd96fce53d6c33e9c7921f375530ee1d39a/h5py-3.14.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:554ef0ced3571366d4d383427c00c966c360e178b5fb5ee5bb31a435c424db0c", size = 4708455, upload-time = "2025-06-06T14:05:11.528Z" },
+ { url = "https://files.pythonhosted.org/packages/86/f9/f00de11c82c88bfc1ef22633557bfba9e271e0cb3189ad704183fc4a2644/h5py-3.14.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0cbd41f4e3761f150aa5b662df991868ca533872c95467216f2bec5fcad84882", size = 4929422, upload-time = "2025-06-06T14:05:18.399Z" },
+ { url = "https://files.pythonhosted.org/packages/7a/6d/6426d5d456f593c94b96fa942a9b3988ce4d65ebaf57d7273e452a7222e8/h5py-3.14.0-cp312-cp312-win_amd64.whl", hash = "sha256:bf4897d67e613ecf5bdfbdab39a1158a64df105827da70ea1d90243d796d367f", size = 2862845, upload-time = "2025-06-06T14:05:23.699Z" },
+ { url = "https://files.pythonhosted.org/packages/6c/c2/7efe82d09ca10afd77cd7c286e42342d520c049a8c43650194928bcc635c/h5py-3.14.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:aa4b7bbce683379b7bf80aaba68e17e23396100336a8d500206520052be2f812", size = 3289245, upload-time = "2025-06-06T14:05:28.24Z" },
+ { url = "https://files.pythonhosted.org/packages/4f/31/f570fab1239b0d9441024b92b6ad03bb414ffa69101a985e4c83d37608bd/h5py-3.14.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:ef9603a501a04fcd0ba28dd8f0995303d26a77a980a1f9474b3417543d4c6174", size = 2807335, upload-time = "2025-06-06T14:05:31.997Z" },
+ { url = "https://files.pythonhosted.org/packages/0d/ce/3a21d87896bc7e3e9255e0ad5583ae31ae9e6b4b00e0bcb2a67e2b6acdbc/h5py-3.14.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e8cbaf6910fa3983c46172666b0b8da7b7bd90d764399ca983236f2400436eeb", size = 4700675, upload-time = "2025-06-06T14:05:37.38Z" },
+ { url = "https://files.pythonhosted.org/packages/e7/ec/86f59025306dcc6deee5fda54d980d077075b8d9889aac80f158bd585f1b/h5py-3.14.0-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d90e6445ab7c146d7f7981b11895d70bc1dd91278a4f9f9028bc0c95e4a53f13", size = 4921632, upload-time = "2025-06-06T14:05:43.464Z" },
+ { url = "https://files.pythonhosted.org/packages/3f/6d/0084ed0b78d4fd3e7530c32491f2884140d9b06365dac8a08de726421d4a/h5py-3.14.0-cp313-cp313-win_amd64.whl", hash = "sha256:ae18e3de237a7a830adb76aaa68ad438d85fe6e19e0d99944a3ce46b772c69b3", size = 2852929, upload-time = "2025-06-06T14:05:47.659Z" },
+]
+
[[package]]
name = "hf-xet"
version = "1.4.3"
@@ -1495,11 +1671,11 @@ wheels = [
[[package]]
name = "idna"
-version = "3.13"
+version = "3.18"
source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/ce/cc/762dfb036166873f0059f3b7de4565e1b5bc3d6f28a414c13da27e442f99/idna-3.13.tar.gz", hash = "sha256:585ea8fe5d69b9181ec1afba340451fba6ba764af97026f92a91d4eef164a242", size = 194210, upload-time = "2026-04-22T16:42:42.314Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/cd/63/9496c57188a2ee585e0f1db071d75089a11e98aa86eb99d9d7618fc1edce/idna-3.18.tar.gz", hash = "sha256:ffb385a7e039654cef1ab9ef32c6fafe283c0c0467bba1d9029738ce4a14a848", size = 196711, upload-time = "2026-06-02T14:34:07.794Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/5d/13/ad7d7ca3808a898b4612b6fe93cde56b53f3034dcde235acb1f0e1df24c6/idna-3.13-py3-none-any.whl", hash = "sha256:892ea0cde124a99ce773decba204c5552b69c3c67ffd5f232eb7696135bc8bb3", size = 68629, upload-time = "2026-04-22T16:42:40.909Z" },
+ { url = "https://files.pythonhosted.org/packages/1e/5e/d4e9f1a599fb8e573b7b87160658329fbf28d19eac2718f51fc3def3aa5a/idna-3.18-py3-none-any.whl", hash = "sha256:7f952cbe720b688055e3f87de14f5c3e5fdaa8bc3928985c4077ca689de849a2", size = 65455, upload-time = "2026-06-02T14:34:06.319Z" },
]
[[package]]
@@ -1655,6 +1831,25 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/7b/91/984aca2ec129e2757d1e4e3c81c3fcda9d0f85b74670a094cc443d9ee949/joblib-1.5.3-py3-none-any.whl", hash = "sha256:5fc3c5039fc5ca8c0276333a188bbd59d6b7ab37fe6632daa76bc7f9ec18e713", size = 309071, upload-time = "2025-12-15T08:41:44.973Z" },
]
+[[package]]
+name = "keras"
+version = "3.15.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "absl-py" },
+ { name = "h5py" },
+ { name = "ml-dtypes" },
+ { name = "namex" },
+ { name = "numpy" },
+ { name = "optree" },
+ { name = "packaging" },
+ { name = "rich" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/ec/a6/4cebb4196f4e6284b35e68dc17dda9a4111a2e2bf21db8211de12881ef89/keras-3.15.1.tar.gz", hash = "sha256:92ae7c1dd7f61041953dc2d42253181ee099086f0b58ae36fcf98147a6f08a29", size = 1919818, upload-time = "2026-07-29T18:20:13.909Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/8d/b3/c9b848bbdba18e765a8051917c0cc82585b64278ce87895f2e521a27438f/keras-3.15.1-py3-none-any.whl", hash = "sha256:836460e480930acbd19bb7a17e62f9ecad40e8a9af9a651fc8d0586a96d27e20", size = 2398595, upload-time = "2026-07-29T18:20:12.345Z" },
+]
+
[[package]]
name = "kgb"
version = "7.3"
@@ -1797,6 +1992,23 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/8a/a1/8d812e53a5da1687abb10445275d41a8b13adb781bbf7196ddbcf8d88505/lazy_loader-0.5-py3-none-any.whl", hash = "sha256:ab0ea149e9c554d4ffeeb21105ac60bed7f3b4fd69b1d2360a4add51b170b005", size = 8044, upload-time = "2026-03-06T15:45:07.668Z" },
]
+[[package]]
+name = "libclang"
+version = "18.1.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6e/5c/ca35e19a4f142adffa27e3d652196b7362fa612243e2b916845d801454fc/libclang-18.1.1.tar.gz", hash = "sha256:a1214966d08d73d971287fc3ead8dfaf82eb07fb197680d8b3859dbbbbf78250", size = 39612, upload-time = "2024-03-17T16:04:37.434Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/4b/49/f5e3e7e1419872b69f6f5e82ba56e33955a74bd537d8a1f5f1eff2f3668a/libclang-18.1.1-1-py2.py3-none-macosx_11_0_arm64.whl", hash = "sha256:0b2e143f0fac830156feb56f9231ff8338c20aecfe72b4ffe96f19e5a1dbb69a", size = 25836045, upload-time = "2024-06-30T17:40:31.646Z" },
+ { url = "https://files.pythonhosted.org/packages/e2/e5/fc61bbded91a8830ccce94c5294ecd6e88e496cc85f6704bf350c0634b70/libclang-18.1.1-py2.py3-none-macosx_10_9_x86_64.whl", hash = "sha256:6f14c3f194704e5d09769108f03185fce7acaf1d1ae4bbb2f30a72c2400cb7c5", size = 26502641, upload-time = "2024-03-18T15:52:26.722Z" },
+ { url = "https://files.pythonhosted.org/packages/db/ed/1df62b44db2583375f6a8a5e2ca5432bbdc3edb477942b9b7c848c720055/libclang-18.1.1-py2.py3-none-macosx_11_0_arm64.whl", hash = "sha256:83ce5045d101b669ac38e6da8e58765f12da2d3aafb3b9b98d88b286a60964d8", size = 26420207, upload-time = "2024-03-17T15:00:26.63Z" },
+ { url = "https://files.pythonhosted.org/packages/1d/fc/716c1e62e512ef1c160e7984a73a5fc7df45166f2ff3f254e71c58076f7c/libclang-18.1.1-py2.py3-none-manylinux2010_x86_64.whl", hash = "sha256:c533091d8a3bbf7460a00cb6c1a71da93bffe148f172c7d03b1c31fbf8aa2a0b", size = 24515943, upload-time = "2024-03-17T16:03:45.942Z" },
+ { url = "https://files.pythonhosted.org/packages/3c/3d/f0ac1150280d8d20d059608cf2d5ff61b7c3b7f7bcf9c0f425ab92df769a/libclang-18.1.1-py2.py3-none-manylinux2014_aarch64.whl", hash = "sha256:54dda940a4a0491a9d1532bf071ea3ef26e6dbaf03b5000ed94dd7174e8f9592", size = 23784972, upload-time = "2024-03-17T16:12:47.677Z" },
+ { url = "https://files.pythonhosted.org/packages/fe/2f/d920822c2b1ce9326a4c78c0c2b4aa3fde610c7ee9f631b600acb5376c26/libclang-18.1.1-py2.py3-none-manylinux2014_armv7l.whl", hash = "sha256:cf4a99b05376513717ab5d82a0db832c56ccea4fd61a69dbb7bccf2dfb207dbe", size = 20259606, upload-time = "2024-03-17T16:17:42.437Z" },
+ { url = "https://files.pythonhosted.org/packages/2d/c2/de1db8c6d413597076a4259cea409b83459b2db997c003578affdd32bf66/libclang-18.1.1-py2.py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:69f8eb8f65c279e765ffd28aaa7e9e364c776c17618af8bff22a8df58677ff4f", size = 24921494, upload-time = "2024-03-17T16:14:20.132Z" },
+ { url = "https://files.pythonhosted.org/packages/0b/2d/3f480b1e1d31eb3d6de5e3ef641954e5c67430d5ac93b7fa7e07589576c7/libclang-18.1.1-py2.py3-none-win_amd64.whl", hash = "sha256:4dd2d3b82fab35e2bf9ca717d7b63ac990a3519c7e312f19fa8e86dcc712f7fb", size = 26415083, upload-time = "2024-03-17T16:42:21.703Z" },
+ { url = "https://files.pythonhosted.org/packages/71/cf/e01dc4cc79779cd82d77888a88ae2fa424d93b445ad4f6c02bfc18335b70/libclang-18.1.1-py2.py3-none-win_arm64.whl", hash = "sha256:3f0e1f49f04d3cd198985fea0511576b0aee16f9ff0e0f0cad7f9c57ec3c20e8", size = 22361112, upload-time = "2024-03-17T16:42:59.565Z" },
+]
+
[[package]]
name = "librosa"
version = "0.11.0"
@@ -2127,46 +2339,54 @@ wheels = [
[[package]]
name = "msgpack"
-version = "1.1.2"
-source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/4d/f2/bfb55a6236ed8725a96b0aa3acbd0ec17588e6a2c3b62a93eb513ed8783f/msgpack-1.1.2.tar.gz", hash = "sha256:3b60763c1373dd60f398488069bcdc703cd08a711477b5d480eecc9f9626f47e", size = 173581, upload-time = "2025-10-08T09:15:56.596Z" }
-wheels = [
- { url = "https://files.pythonhosted.org/packages/ad/bd/8b0d01c756203fbab65d265859749860682ccd2a59594609aeec3a144efa/msgpack-1.1.2-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:70a0dff9d1f8da25179ffcf880e10cf1aad55fdb63cd59c9a49a1b82290062aa", size = 81939, upload-time = "2025-10-08T09:15:01.472Z" },
- { url = "https://files.pythonhosted.org/packages/34/68/ba4f155f793a74c1483d4bdef136e1023f7bcba557f0db4ef3db3c665cf1/msgpack-1.1.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:446abdd8b94b55c800ac34b102dffd2f6aa0ce643c55dfc017ad89347db3dbdb", size = 85064, upload-time = "2025-10-08T09:15:03.764Z" },
- { url = "https://files.pythonhosted.org/packages/f2/60/a064b0345fc36c4c3d2c743c82d9100c40388d77f0b48b2f04d6041dbec1/msgpack-1.1.2-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c63eea553c69ab05b6747901b97d620bb2a690633c77f23feb0c6a947a8a7b8f", size = 417131, upload-time = "2025-10-08T09:15:05.136Z" },
- { url = "https://files.pythonhosted.org/packages/65/92/a5100f7185a800a5d29f8d14041f61475b9de465ffcc0f3b9fba606e4505/msgpack-1.1.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:372839311ccf6bdaf39b00b61288e0557916c3729529b301c52c2d88842add42", size = 427556, upload-time = "2025-10-08T09:15:06.837Z" },
- { url = "https://files.pythonhosted.org/packages/f5/87/ffe21d1bf7d9991354ad93949286f643b2bb6ddbeab66373922b44c3b8cc/msgpack-1.1.2-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:2929af52106ca73fcb28576218476ffbb531a036c2adbcf54a3664de124303e9", size = 404920, upload-time = "2025-10-08T09:15:08.179Z" },
- { url = "https://files.pythonhosted.org/packages/ff/41/8543ed2b8604f7c0d89ce066f42007faac1eaa7d79a81555f206a5cdb889/msgpack-1.1.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:be52a8fc79e45b0364210eef5234a7cf8d330836d0a64dfbb878efa903d84620", size = 415013, upload-time = "2025-10-08T09:15:09.83Z" },
- { url = "https://files.pythonhosted.org/packages/41/0d/2ddfaa8b7e1cee6c490d46cb0a39742b19e2481600a7a0e96537e9c22f43/msgpack-1.1.2-cp312-cp312-win32.whl", hash = "sha256:1fff3d825d7859ac888b0fbda39a42d59193543920eda9d9bea44d958a878029", size = 65096, upload-time = "2025-10-08T09:15:11.11Z" },
- { url = "https://files.pythonhosted.org/packages/8c/ec/d431eb7941fb55a31dd6ca3404d41fbb52d99172df2e7707754488390910/msgpack-1.1.2-cp312-cp312-win_amd64.whl", hash = "sha256:1de460f0403172cff81169a30b9a92b260cb809c4cb7e2fc79ae8d0510c78b6b", size = 72708, upload-time = "2025-10-08T09:15:12.554Z" },
- { url = "https://files.pythonhosted.org/packages/c5/31/5b1a1f70eb0e87d1678e9624908f86317787b536060641d6798e3cf70ace/msgpack-1.1.2-cp312-cp312-win_arm64.whl", hash = "sha256:be5980f3ee0e6bd44f3a9e9dea01054f175b50c3e6cdb692bc9424c0bbb8bf69", size = 64119, upload-time = "2025-10-08T09:15:13.589Z" },
- { url = "https://files.pythonhosted.org/packages/6b/31/b46518ecc604d7edf3a4f94cb3bf021fc62aa301f0cb849936968164ef23/msgpack-1.1.2-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:4efd7b5979ccb539c221a4c4e16aac1a533efc97f3b759bb5a5ac9f6d10383bf", size = 81212, upload-time = "2025-10-08T09:15:14.552Z" },
- { url = "https://files.pythonhosted.org/packages/92/dc/c385f38f2c2433333345a82926c6bfa5ecfff3ef787201614317b58dd8be/msgpack-1.1.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:42eefe2c3e2af97ed470eec850facbe1b5ad1d6eacdbadc42ec98e7dcf68b4b7", size = 84315, upload-time = "2025-10-08T09:15:15.543Z" },
- { url = "https://files.pythonhosted.org/packages/d3/68/93180dce57f684a61a88a45ed13047558ded2be46f03acb8dec6d7c513af/msgpack-1.1.2-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1fdf7d83102bf09e7ce3357de96c59b627395352a4024f6e2458501f158bf999", size = 412721, upload-time = "2025-10-08T09:15:16.567Z" },
- { url = "https://files.pythonhosted.org/packages/5d/ba/459f18c16f2b3fc1a1ca871f72f07d70c07bf768ad0a507a698b8052ac58/msgpack-1.1.2-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fac4be746328f90caa3cd4bc67e6fe36ca2bf61d5c6eb6d895b6527e3f05071e", size = 424657, upload-time = "2025-10-08T09:15:17.825Z" },
- { url = "https://files.pythonhosted.org/packages/38/f8/4398c46863b093252fe67368b44edc6c13b17f4e6b0e4929dbf0bdb13f23/msgpack-1.1.2-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:fffee09044073e69f2bad787071aeec727183e7580443dfeb8556cbf1978d162", size = 402668, upload-time = "2025-10-08T09:15:19.003Z" },
- { url = "https://files.pythonhosted.org/packages/28/ce/698c1eff75626e4124b4d78e21cca0b4cc90043afb80a507626ea354ab52/msgpack-1.1.2-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:5928604de9b032bc17f5099496417f113c45bc6bc21b5c6920caf34b3c428794", size = 419040, upload-time = "2025-10-08T09:15:20.183Z" },
- { url = "https://files.pythonhosted.org/packages/67/32/f3cd1667028424fa7001d82e10ee35386eea1408b93d399b09fb0aa7875f/msgpack-1.1.2-cp313-cp313-win32.whl", hash = "sha256:a7787d353595c7c7e145e2331abf8b7ff1e6673a6b974ded96e6d4ec09f00c8c", size = 65037, upload-time = "2025-10-08T09:15:21.416Z" },
- { url = "https://files.pythonhosted.org/packages/74/07/1ed8277f8653c40ebc65985180b007879f6a836c525b3885dcc6448ae6cb/msgpack-1.1.2-cp313-cp313-win_amd64.whl", hash = "sha256:a465f0dceb8e13a487e54c07d04ae3ba131c7c5b95e2612596eafde1dccf64a9", size = 72631, upload-time = "2025-10-08T09:15:22.431Z" },
- { url = "https://files.pythonhosted.org/packages/e5/db/0314e4e2db56ebcf450f277904ffd84a7988b9e5da8d0d61ab2d057df2b6/msgpack-1.1.2-cp313-cp313-win_arm64.whl", hash = "sha256:e69b39f8c0aa5ec24b57737ebee40be647035158f14ed4b40e6f150077e21a84", size = 64118, upload-time = "2025-10-08T09:15:23.402Z" },
- { url = "https://files.pythonhosted.org/packages/22/71/201105712d0a2ff07b7873ed3c220292fb2ea5120603c00c4b634bcdafb3/msgpack-1.1.2-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:e23ce8d5f7aa6ea6d2a2b326b4ba46c985dbb204523759984430db7114f8aa00", size = 81127, upload-time = "2025-10-08T09:15:24.408Z" },
- { url = "https://files.pythonhosted.org/packages/1b/9f/38ff9e57a2eade7bf9dfee5eae17f39fc0e998658050279cbb14d97d36d9/msgpack-1.1.2-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:6c15b7d74c939ebe620dd8e559384be806204d73b4f9356320632d783d1f7939", size = 84981, upload-time = "2025-10-08T09:15:25.812Z" },
- { url = "https://files.pythonhosted.org/packages/8e/a9/3536e385167b88c2cc8f4424c49e28d49a6fc35206d4a8060f136e71f94c/msgpack-1.1.2-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:99e2cb7b9031568a2a5c73aa077180f93dd2e95b4f8d3b8e14a73ae94a9e667e", size = 411885, upload-time = "2025-10-08T09:15:27.22Z" },
- { url = "https://files.pythonhosted.org/packages/2f/40/dc34d1a8d5f1e51fc64640b62b191684da52ca469da9cd74e84936ffa4a6/msgpack-1.1.2-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:180759d89a057eab503cf62eeec0aa61c4ea1200dee709f3a8e9397dbb3b6931", size = 419658, upload-time = "2025-10-08T09:15:28.4Z" },
- { url = "https://files.pythonhosted.org/packages/3b/ef/2b92e286366500a09a67e03496ee8b8ba00562797a52f3c117aa2b29514b/msgpack-1.1.2-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:04fb995247a6e83830b62f0b07bf36540c213f6eac8e851166d8d86d83cbd014", size = 403290, upload-time = "2025-10-08T09:15:29.764Z" },
- { url = "https://files.pythonhosted.org/packages/78/90/e0ea7990abea5764e4655b8177aa7c63cdfa89945b6e7641055800f6c16b/msgpack-1.1.2-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:8e22ab046fa7ede9e36eeb4cfad44d46450f37bb05d5ec482b02868f451c95e2", size = 415234, upload-time = "2025-10-08T09:15:31.022Z" },
- { url = "https://files.pythonhosted.org/packages/72/4e/9390aed5db983a2310818cd7d3ec0aecad45e1f7007e0cda79c79507bb0d/msgpack-1.1.2-cp314-cp314-win32.whl", hash = "sha256:80a0ff7d4abf5fecb995fcf235d4064b9a9a8a40a3ab80999e6ac1e30b702717", size = 66391, upload-time = "2025-10-08T09:15:32.265Z" },
- { url = "https://files.pythonhosted.org/packages/6e/f1/abd09c2ae91228c5f3998dbd7f41353def9eac64253de3c8105efa2082f7/msgpack-1.1.2-cp314-cp314-win_amd64.whl", hash = "sha256:9ade919fac6a3e7260b7f64cea89df6bec59104987cbea34d34a2fa15d74310b", size = 73787, upload-time = "2025-10-08T09:15:33.219Z" },
- { url = "https://files.pythonhosted.org/packages/6a/b0/9d9f667ab48b16ad4115c1935d94023b82b3198064cb84a123e97f7466c1/msgpack-1.1.2-cp314-cp314-win_arm64.whl", hash = "sha256:59415c6076b1e30e563eb732e23b994a61c159cec44deaf584e5cc1dd662f2af", size = 66453, upload-time = "2025-10-08T09:15:34.225Z" },
- { url = "https://files.pythonhosted.org/packages/16/67/93f80545eb1792b61a217fa7f06d5e5cb9e0055bed867f43e2b8e012e137/msgpack-1.1.2-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:897c478140877e5307760b0ea66e0932738879e7aa68144d9b78ea4c8302a84a", size = 85264, upload-time = "2025-10-08T09:15:35.61Z" },
- { url = "https://files.pythonhosted.org/packages/87/1c/33c8a24959cf193966ef11a6f6a2995a65eb066bd681fd085afd519a57ce/msgpack-1.1.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:a668204fa43e6d02f89dbe79a30b0d67238d9ec4c5bd8a940fc3a004a47b721b", size = 89076, upload-time = "2025-10-08T09:15:36.619Z" },
- { url = "https://files.pythonhosted.org/packages/fc/6b/62e85ff7193663fbea5c0254ef32f0c77134b4059f8da89b958beb7696f3/msgpack-1.1.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5559d03930d3aa0f3aacb4c42c776af1a2ace2611871c84a75afe436695e6245", size = 435242, upload-time = "2025-10-08T09:15:37.647Z" },
- { url = "https://files.pythonhosted.org/packages/c1/47/5c74ecb4cc277cf09f64e913947871682ffa82b3b93c8dad68083112f412/msgpack-1.1.2-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:70c5a7a9fea7f036b716191c29047374c10721c389c21e9ffafad04df8c52c90", size = 432509, upload-time = "2025-10-08T09:15:38.794Z" },
- { url = "https://files.pythonhosted.org/packages/24/a4/e98ccdb56dc4e98c929a3f150de1799831c0a800583cde9fa022fa90602d/msgpack-1.1.2-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:f2cb069d8b981abc72b41aea1c580ce92d57c673ec61af4c500153a626cb9e20", size = 415957, upload-time = "2025-10-08T09:15:40.238Z" },
- { url = "https://files.pythonhosted.org/packages/da/28/6951f7fb67bc0a4e184a6b38ab71a92d9ba58080b27a77d3e2fb0be5998f/msgpack-1.1.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:d62ce1f483f355f61adb5433ebfd8868c5f078d1a52d042b0a998682b4fa8c27", size = 422910, upload-time = "2025-10-08T09:15:41.505Z" },
- { url = "https://files.pythonhosted.org/packages/f0/03/42106dcded51f0a0b5284d3ce30a671e7bd3f7318d122b2ead66ad289fed/msgpack-1.1.2-cp314-cp314t-win32.whl", hash = "sha256:1d1418482b1ee984625d88aa9585db570180c286d942da463533b238b98b812b", size = 75197, upload-time = "2025-10-08T09:15:42.954Z" },
- { url = "https://files.pythonhosted.org/packages/15/86/d0071e94987f8db59d4eeb386ddc64d0bb9b10820a8d82bcd3e53eeb2da6/msgpack-1.1.2-cp314-cp314t-win_amd64.whl", hash = "sha256:5a46bf7e831d09470ad92dff02b8b1ac92175ca36b087f904a0519857c6be3ff", size = 85772, upload-time = "2025-10-08T09:15:43.954Z" },
- { url = "https://files.pythonhosted.org/packages/81/f2/08ace4142eb281c12701fc3b93a10795e4d4dc7f753911d836675050f886/msgpack-1.1.2-cp314-cp314t-win_arm64.whl", hash = "sha256:d99ef64f349d5ec3293688e91486c5fdb925ed03807f64d98d205d2713c60b46", size = 70868, upload-time = "2025-10-08T09:15:44.959Z" },
+version = "1.2.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/31/f9/c0a1c127f9049db9155afc316952ea571720dd01833ff5e4d7e8e6352dbb/msgpack-1.2.1.tar.gz", hash = "sha256:04c721c2c7448767e9e3f2520a475663d8ee0f09c31890f6d2bd70fd636a9647", size = 183960, upload-time = "2026-06-18T16:13:52.594Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/bc/dd/9e8cbd8f5582ca4b590336f2b91ee5662f6a6ca562b565abaf696a0f81ff/msgpack-1.2.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:2ef59c659f289eddf8aa6623823f19fa2f40a4029266889eac7a2505dd210c35", size = 83531, upload-time = "2026-06-18T16:12:58.249Z" },
+ { url = "https://files.pythonhosted.org/packages/50/2e/ebdb85a8da151397a2790363676b7ed7c125924fe618e4c6d8befb0cc62c/msgpack-1.2.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:d3567748a5107cb40cdf66a275430c2f87c07777698f4bfd25c35f44d533258c", size = 82657, upload-time = "2026-06-18T16:12:59.396Z" },
+ { url = "https://files.pythonhosted.org/packages/26/aa/753ad8b007b464e1d8aa0c8e650b9c5f4f725e658fc5ac8a7635c55b7f6e/msgpack-1.2.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:60926b75d00c8e816ef98f3034f484a8bc64242d66839cef4cf7e503142316a0", size = 410634, upload-time = "2026-06-18T16:13:00.383Z" },
+ { url = "https://files.pythonhosted.org/packages/6a/fd/6adabd4f6d5e686f97dd02ce7fce3fe4cf672cbac36b8f67ff4040e8ad8b/msgpack-1.2.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:020e881a764b20d8d7ca1a54fc01b8175519d108e3c3f194fddc200bda95951a", size = 419989, upload-time = "2026-06-18T16:13:01.776Z" },
+ { url = "https://files.pythonhosted.org/packages/5a/cc/85039b7b0eb168aaad7383a23c97e291a11f08351cb45a606ce865e4e3f1/msgpack-1.2.1-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4202c74688ca06591f78cb18988228bd4cca2cc75d57b60008372892d2f1e6e6", size = 377544, upload-time = "2026-06-18T16:13:03.637Z" },
+ { url = "https://files.pythonhosted.org/packages/ed/bf/35963899493b32030c85fc513b723ae66144ac70c11ebc52e889e16e3d99/msgpack-1.2.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:8b267ce94efb76fbd1b3373511420074ee3187f0f7811bf394531de13294735a", size = 400842, upload-time = "2026-06-18T16:13:05.012Z" },
+ { url = "https://files.pythonhosted.org/packages/a6/df/8e2ac970c8f99264cd9997d1c73df5466bc19da3301d7dc5500862a9b089/msgpack-1.2.1-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:e4f1d0f8f98ade9634e01fb704a408f9336c0a8f1117b369f5db83dc7551d8b1", size = 374108, upload-time = "2026-06-18T16:13:06.232Z" },
+ { url = "https://files.pythonhosted.org/packages/17/dd/fa8bd265110dfa51c20cb529f9e6d240a16fafe7e645004c6af2d01353ba/msgpack-1.2.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:f02cf17a6ca1abe29b5f980644f7551f94d71f2011509b26d8625ce038f0df64", size = 414939, upload-time = "2026-06-18T16:13:07.478Z" },
+ { url = "https://files.pythonhosted.org/packages/2e/b9/8377a5ad8953fc0437c70cc98d9ae29f27fe5ac5109fbec0812085865735/msgpack-1.2.1-cp312-cp312-win32.whl", hash = "sha256:0c0d9802354507bcba62af19c17918e3eb437cc25e6f50657d511b5856a77aac", size = 64504, upload-time = "2026-06-18T16:13:08.822Z" },
+ { url = "https://files.pythonhosted.org/packages/57/7f/ce1e377df7e62461fefd9eb23bfb93a4a523f40a517b377b8f844d836828/msgpack-1.2.1-cp312-cp312-win_amd64.whl", hash = "sha256:5c24aa15d5963051e1a5c62b12c50cd705992502b5ec1f3bece6046f33c9fc24", size = 71421, upload-time = "2026-06-18T16:13:09.828Z" },
+ { url = "https://files.pythonhosted.org/packages/8f/32/ebfe84c9929f08f188d56c7a2fd913406a9ddad76a634697c1c43b8112e6/msgpack-1.2.1-cp312-cp312-win_arm64.whl", hash = "sha256:4227224aaec8f7fbcbfbd4272319347b2bb4030366502600f8c45588c5187b07", size = 64775, upload-time = "2026-06-18T16:13:11.056Z" },
+ { url = "https://files.pythonhosted.org/packages/b0/ac/dcddcab6f6c20ecb387ca5e980371cdb3f87ff69aeca388be97eebc4c074/msgpack-1.2.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:0a70e3cf2804a300d921bb0940426e35f4e489a23adfb77a808892241db0a064", size = 83151, upload-time = "2026-06-18T16:13:12.173Z" },
+ { url = "https://files.pythonhosted.org/packages/64/71/fbcfa83a1d6a9c6091942d1cfd070962244664b87427a9a49a6897b1b219/msgpack-1.2.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:491cc39455ca765fad51fb451bf2915eb2cf41192ab5801ce8d67c1d614fe056", size = 82351, upload-time = "2026-06-18T16:13:13.194Z" },
+ { url = "https://files.pythonhosted.org/packages/e3/10/ddf7b06db879e8792d13934ddda09ff20bd2a583fd84c9b59aae9b0e650b/msgpack-1.2.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f310233ef7fb9c14e201c93639fe5f5260b005f56f0b29048e999c30935596cc", size = 407518, upload-time = "2026-06-18T16:13:14.233Z" },
+ { url = "https://files.pythonhosted.org/packages/79/d3/36a46a8ed992b781acbc05928bd5bee3c810cb0c3563bf81a7b0c04a1a76/msgpack-1.2.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:787c9bebb5833e8f6fc8abca3c0597683d8d87f56a8842b6b89c75a5f3176e2d", size = 416405, upload-time = "2026-06-18T16:13:15.435Z" },
+ { url = "https://files.pythonhosted.org/packages/f9/84/e8e9598b557c0ba6ddae901a73780a4c75ac667dddf59414b1e56a42fb34/msgpack-1.2.1-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:dc871b997a9370d855b7394465f2f350e847a5b806dd38dcc9c989e7d87da155", size = 376257, upload-time = "2026-06-18T16:13:17.022Z" },
+ { url = "https://files.pythonhosted.org/packages/40/16/738fe6d875ad7e2a9429c165322a4ec088f4f273cdfae63d96a89c467961/msgpack-1.2.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:85f57e960d877f2977f6430896191b04a21f8901b3b4baf2e4604329f4db5402", size = 397469, upload-time = "2026-06-18T16:13:18.287Z" },
+ { url = "https://files.pythonhosted.org/packages/ca/be/6d5952df75a7f24f35833af764c3a6860780364cb3a0030beb8099e1b2b4/msgpack-1.2.1-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:1233ee2dd0cefba127583de50ea654677277047d238303521db35def3d7b2e7c", size = 372802, upload-time = "2026-06-18T16:13:19.685Z" },
+ { url = "https://files.pythonhosted.org/packages/e1/39/e2ef7dbf0473bcb8dc7c50bf782a892d67414877b63e47fc88eb189ef5e6/msgpack-1.2.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e3dc2feb0876209d9c38aa56cb1de169bd6c4348f1aa48271f241226590993e6", size = 411273, upload-time = "2026-06-18T16:13:21.028Z" },
+ { url = "https://files.pythonhosted.org/packages/ef/c5/133f4512a56e983a93445c836c9d94d88f3bc2e0980ff4b9e577bd8416ce/msgpack-1.2.1-cp313-cp313-win32.whl", hash = "sha256:6d09badf350af2be9d189184e04e64cf54ad93569ab3d96fca58bd3e84aad707", size = 64471, upload-time = "2026-06-18T16:13:22.293Z" },
+ { url = "https://files.pythonhosted.org/packages/e2/98/577e10b055096a7dd40732358cabaf7180a20c79ed1dcdbb618e4b9deac7/msgpack-1.2.1-cp313-cp313-win_amd64.whl", hash = "sha256:33f14fba63278b714efe6ad07e50ea5f03d91537aa6a1c5f1ceca4cf44013ca9", size = 71274, upload-time = "2026-06-18T16:13:23.455Z" },
+ { url = "https://files.pythonhosted.org/packages/ba/ee/0c0048e7cfbef23c6a94791b8959ab28155232e7956de8a305b5ff588f05/msgpack-1.2.1-cp313-cp313-win_arm64.whl", hash = "sha256:afc5febcd4c99effbc02b528e49d6fd0760b2b7d48c05239e345a5fa6e743d9a", size = 64795, upload-time = "2026-06-18T16:13:24.687Z" },
+ { url = "https://files.pythonhosted.org/packages/77/58/cce442852c6b9e1639c7c8ac8fd9143121cb32dab0f308df4d1426a8eb9c/msgpack-1.2.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:05f340e47e7e47d2da8db9b53e1bb1d294369e9ef45a747441309f6650b8351d", size = 83610, upload-time = "2026-06-18T16:13:25.724Z" },
+ { url = "https://files.pythonhosted.org/packages/60/5c/15b4c7a0182f75ffa90751958ba36a9c01cafee367d49a3edc10ed140b01/msgpack-1.2.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:810b916696c86ef0deb3b74588480224df4c1b071136c34183e4a2a4284d7ac7", size = 83138, upload-time = "2026-06-18T16:13:26.781Z" },
+ { url = "https://files.pythonhosted.org/packages/b8/a6/99e58722feaffc5f2fbcc0c8c0d1451ab9f84097f7af87291b46af2390f4/msgpack-1.2.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ca0dacff965c47afdc3749a8469d7302a8f801d6a28758d55120d75e66ce6889", size = 406090, upload-time = "2026-06-18T16:13:28.072Z" },
+ { url = "https://files.pythonhosted.org/packages/19/03/8c63e8cf52958534ef688625965ab04c269a6cadd8caef16758b380a821a/msgpack-1.2.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0e2bf9280bceb5efca998435904b5d3e9fdbcc11d90dc9df30aec7973252b720", size = 412106, upload-time = "2026-06-18T16:13:29.427Z" },
+ { url = "https://files.pythonhosted.org/packages/63/d2/155d9e71b40e41fd934bc0c48b9b2770f22263e1ac20aad8e29fdca7be3f/msgpack-1.2.1-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:aa6c4be5d1c02a42b066ca6ddb71adf36432868fdcdb6ee87e634e86e0674190", size = 374851, upload-time = "2026-06-18T16:13:30.631Z" },
+ { url = "https://files.pythonhosted.org/packages/98/48/deaf2326262a8d5ea3295ce9649912ecd3f551ba7ec8e33c665d2ba583f3/msgpack-1.2.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:ec0e675d59150a6269ddc9139087c722292664a37d071a849c05c473350f1f2d", size = 396168, upload-time = "2026-06-18T16:13:31.977Z" },
+ { url = "https://files.pythonhosted.org/packages/10/2a/b4410f906c2ec0008f1608d3ab5143afc3ad3f4e6da0fed3ea2231d0bef4/msgpack-1.2.1-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:dd3bfe82d53edfe4b7fc9a7ec9761e23a7a5b1dac22264505af428253c29ed24", size = 371959, upload-time = "2026-06-18T16:13:33.282Z" },
+ { url = "https://files.pythonhosted.org/packages/59/86/1edc67270099a528fa2093ea60fe191233cd238e4bd30cfacf7db79fc959/msgpack-1.2.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:5ad5467fc3f68b5468e06c5f788d712e9f8ffc8b0cd1bcb160c105c1ee92dae7", size = 408457, upload-time = "2026-06-18T16:13:34.567Z" },
+ { url = "https://files.pythonhosted.org/packages/82/90/8b630fef07d8c5ab457b71ff2c217910c83d333c7a68472c186e87cc504a/msgpack-1.2.1-cp314-cp314-win32.whl", hash = "sha256:98b58bdb89c46190e4609bb36abe17c6d4105ad13f9c5f8f6f64d320f8ced3fb", size = 65942, upload-time = "2026-06-18T16:13:36.056Z" },
+ { url = "https://files.pythonhosted.org/packages/16/f1/467b81e98b24dd3885d7b1857728797b4ffc76a7a7483af4fb321a07de3c/msgpack-1.2.1-cp314-cp314-win_amd64.whl", hash = "sha256:74847557e28ce71bd3c438a447ca90e4b507e997ddbdef8a12a7b283b86c156b", size = 72627, upload-time = "2026-06-18T16:13:37.079Z" },
+ { url = "https://files.pythonhosted.org/packages/a7/1d/5d8c4c89985feb6acefb82a09e501c60392261856d2408d20bfe4f0360b1/msgpack-1.2.1-cp314-cp314-win_arm64.whl", hash = "sha256:b50b727bd652bdc37d950336c848ef20ec54a4cafc38dce19b1cd86ad625d0f7", size = 66908, upload-time = "2026-06-18T16:13:38.23Z" },
+ { url = "https://files.pythonhosted.org/packages/1b/02/ad2afb678b4de94496cd432b581759b756a92c1192d8c767edd6b132efdc/msgpack-1.2.1-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:8d00f177ca88a77c1cf848d204a38f249751650b601cb6532acc68805d8a8273", size = 86000, upload-time = "2026-06-18T16:13:39.44Z" },
+ { url = "https://files.pythonhosted.org/packages/54/74/0b797484013128837f3b1cbb6cea019277c4de4e377dc512b4d9a0f92940/msgpack-1.2.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:5bb9c386f0a329c035ddbab4b72d1028bf9627add8dda41070288563d57ed1b1", size = 86544, upload-time = "2026-06-18T16:13:40.447Z" },
+ { url = "https://files.pythonhosted.org/packages/a9/b4/b774d7eb95561739907fec675582f83203cf41c597a418c2589b4bfb8e9d/msgpack-1.2.1-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:20466cca18c49c7292a8984bc15d65857b171e7264bdcb5f96baf8be238791fc", size = 427661, upload-time = "2026-06-18T16:13:41.574Z" },
+ { url = "https://files.pythonhosted.org/packages/b2/f9/3243191dc9937e00756c8bc1b0272fed8f23758e43df2a3b46f533e5090f/msgpack-1.2.1-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:196300e7e5d6e74d50f1607ab9c06c4a1484c383cd22defd727902591f7e8dde", size = 426375, upload-time = "2026-06-18T16:13:42.936Z" },
+ { url = "https://files.pythonhosted.org/packages/23/c7/1693111db9944ba4ad4b67a1e788400d78a0b6af7a6523dc7e4e58f8274b/msgpack-1.2.1-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:575957e79cd51903a4e8495a242442949641e08f1efd5197b43bebd3ea7682b4", size = 380495, upload-time = "2026-06-18T16:13:44.306Z" },
+ { url = "https://files.pythonhosted.org/packages/3e/2b/92f86956a0c13e8662f7e2ad630c4eb4db07497b967589bd5245e018b2c1/msgpack-1.2.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:8c2ed1e48cc0f460bf3c7780e7137ff21a4e18433451916f2442c1b21036cd7d", size = 410897, upload-time = "2026-06-18T16:13:45.629Z" },
+ { url = "https://files.pythonhosted.org/packages/da/ea/1479f72d200313a76fc2f823a79d1e07ed052ab7b8a0280640aa7b95de42/msgpack-1.2.1-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:5f6277e5f783c36786a145e0247fc189a03f35f84b251646e53592d2bc12b355", size = 378519, upload-time = "2026-06-18T16:13:46.998Z" },
+ { url = "https://files.pythonhosted.org/packages/f5/4d/fa006060ffa1011d32bfae826fe766fe73e02982183601633b7121058ab3/msgpack-1.2.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f9389552ecf4784886345ead0647e4edc96bee37cbab05b75540f542f766c48c", size = 419815, upload-time = "2026-06-18T16:13:48.205Z" },
+ { url = "https://files.pythonhosted.org/packages/2f/e1/aab6c946570496b78e67804721f3d5e2d62a93081b9b37df77764ef56347/msgpack-1.2.1-cp314-cp314t-win32.whl", hash = "sha256:c1c79a604a2969a868a78b6ebd27a887e00c624f14f66b3038e0590cb23332d1", size = 70914, upload-time = "2026-06-18T16:13:49.385Z" },
+ { url = "https://files.pythonhosted.org/packages/13/0a/e608956488a2af014cfe6e3d665e090b8ee42aa14b07f8f95b8880d66b09/msgpack-1.2.1-cp314-cp314t-win_amd64.whl", hash = "sha256:f12038a35fabd52e56a3547bab42401af49a45caa6dd00b34c44de235bc93ee2", size = 77999, upload-time = "2026-06-18T16:13:50.467Z" },
+ { url = "https://files.pythonhosted.org/packages/d2/8a/27e2e57055176e366a46b85d02d68e7a5bcfbdd8474c9706375d965f24d3/msgpack-1.2.1-cp314-cp314t-win_arm64.whl", hash = "sha256:0adcf06ffde0777c0e1a9b771a2b1c4226ba1bbf748c8efcc02fcdeca3299107", size = 71160, upload-time = "2026-06-18T16:13:51.498Z" },
]
[[package]]
@@ -2285,6 +2505,15 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/7e/82/69e539c4c2027f1e1697e09aaa2449243085a0edf81ae2c6341e84d769b6/multiprocess-0.70.19-py39-none-any.whl", hash = "sha256:0d4b4397ed669d371c81dcd1ef33fd384a44d6c3de1bd0ca7ac06d837720d3c5", size = 133477, upload-time = "2026-01-19T06:47:38.619Z" },
]
+[[package]]
+name = "namex"
+version = "0.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/0c/c0/ee95b28f029c73f8d49d8f52edaed02a1d4a9acb8b69355737fdb1faa191/namex-0.1.0.tar.gz", hash = "sha256:117f03ccd302cc48e3f5c58a296838f6b89c83455ab8683a1e85f2a430aa4306", size = 6649, upload-time = "2025-05-26T23:17:38.918Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/b2/bc/465daf1de06409cdd4532082806770ee0d8d7df434da79c76564d0f69741/namex-0.1.0-py3-none-any.whl", hash = "sha256:e2012a474502f1e2251267062aae3114611f07df4224b6e06334c57b0f2ce87c", size = 5905, upload-time = "2025-05-26T23:17:37.695Z" },
+]
+
[[package]]
name = "networkx"
version = "3.6.1"
@@ -2385,11 +2614,14 @@ wheels = [
[[package]]
name = "nvidia-cublas"
-version = "13.1.0.3"
+version = "13.1.1.3"
source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "nvidia-cuda-nvrtc", marker = "sys_platform != 'win32'" },
+]
wheels = [
- { url = "https://files.pythonhosted.org/packages/e1/a5/fce49e2ae977e0ccc084e5adafceb4f0ac0c8333cb6863501618a7277f67/nvidia_cublas-13.1.0.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:c86fc7f7ae36d7528288c5d88098edcb7b02c633d262e7ddbb86b0ad91be5df2", size = 542851226, upload-time = "2025-10-09T08:59:04.818Z" },
- { url = "https://files.pythonhosted.org/packages/e7/44/423ac00af4dd95a5aeb27207e2c0d9b7118702149bf4704c3ddb55bb7429/nvidia_cublas-13.1.0.3-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:ee8722c1f0145ab246bccb9e452153b5e0515fd094c3678df50b2a0888b8b171", size = 423133236, upload-time = "2025-10-09T08:59:32.536Z" },
+ { url = "https://files.pythonhosted.org/packages/a7/a1/0bd24ee8c8d03adac032fd2909426a00c88f8c57961b1277ded97f91119f/nvidia_cublas-13.1.1.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:b7a210458267ac818974c53038fbec2e969d5c99f305ab15c72522fa9f001dd5", size = 542848918, upload-time = "2026-04-08T18:46:22.985Z" },
+ { url = "https://files.pythonhosted.org/packages/3b/cd/154ca20c38269e05eff77c1464e6c1da89f50a6390b565e9d82e06bc11e1/nvidia_cublas-13.1.1.3-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:37936a16db8fe4ac1f065c2139360608a543a09275cb1a1af612e08cfa065436", size = 423138758, upload-time = "2026-04-08T18:46:58.655Z" },
]
[[package]]
@@ -2490,14 +2722,14 @@ wheels = [
[[package]]
name = "nvidia-cudnn-cu13"
-version = "9.19.0.56"
+version = "9.20.0.48"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "nvidia-cublas" },
+ { name = "nvidia-cublas", marker = "sys_platform != 'win32'" },
]
wheels = [
- { url = "https://files.pythonhosted.org/packages/f1/84/26025437c1e6b61a707442184fa0c03d083b661adf3a3eecfd6d21677740/nvidia_cudnn_cu13-9.19.0.56-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:6ed29ffaee1176c612daf442e4dd6cfeb6a0caa43ddcbeb59da94953030b1be4", size = 433781201, upload-time = "2026-02-03T20:40:53.805Z" },
- { url = "https://files.pythonhosted.org/packages/a3/22/0b4b932655d17a6da1b92fa92ab12844b053bb2ac2475e179ba6f043da1e/nvidia_cudnn_cu13-9.19.0.56-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:d20e1734305e9d68889a96e3f35094d733ff1f83932ebe462753973e53a572bf", size = 366066321, upload-time = "2026-02-03T20:44:52.837Z" },
+ { url = "https://files.pythonhosted.org/packages/56/c5/83384d846b2fd17c44bd499b36c75a45ed4f095fbbb2252294e89cea5c5c/nvidia_cudnn_cu13-9.20.0.48-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:e31454ae00094b0c55319d9d15b6fa2fc50a9e1c0f5c8c80fb75258234e731e1", size = 444574296, upload-time = "2026-03-09T19:28:27.751Z" },
+ { url = "https://files.pythonhosted.org/packages/6e/5e/edb9c0ae051602c3ccaffe424256463636d639e27d7f302dde9975ef9e7a/nvidia_cudnn_cu13-9.20.0.48-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:0c45dd8eeb50b603f07995b1b300c62ffe6a1980482b82b3bcf94a4ca9d49304", size = 366173588, upload-time = "2026-03-09T19:29:34.474Z" },
]
[[package]]
@@ -2505,7 +2737,7 @@ name = "nvidia-cufft"
version = "12.0.0.61"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "nvidia-nvjitlink" },
+ { name = "nvidia-nvjitlink", marker = "sys_platform != 'win32'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/8b/ae/f417a75c0259e85c1d2f83ca4e960289a5f814ed0cea74d18c353d3e989d/nvidia_cufft-12.0.0.61-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:2708c852ef8cd89d1d2068bdbece0aa188813a0c934db3779b9b1faa8442e5f5", size = 214053554, upload-time = "2025-09-04T08:31:38.196Z" },
@@ -2547,9 +2779,9 @@ name = "nvidia-cusolver"
version = "12.0.4.66"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "nvidia-cublas" },
- { name = "nvidia-cusparse" },
- { name = "nvidia-nvjitlink" },
+ { name = "nvidia-cublas", marker = "sys_platform != 'win32'" },
+ { name = "nvidia-cusparse", marker = "sys_platform != 'win32'" },
+ { name = "nvidia-nvjitlink", marker = "sys_platform != 'win32'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/c8/c3/b30c9e935fc01e3da443ec0116ed1b2a009bb867f5324d3f2d7e533e776b/nvidia_cusolver-12.0.4.66-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:02c2457eaa9e39de20f880f4bd8820e6a1cfb9f9a34f820eb12a155aa5bc92d2", size = 223467760, upload-time = "2025-09-04T08:33:04.222Z" },
@@ -2575,7 +2807,7 @@ name = "nvidia-cusparse"
version = "12.6.3.3"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "nvidia-nvjitlink" },
+ { name = "nvidia-nvjitlink", marker = "sys_platform != 'win32'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/f8/94/5c26f33738ae35276672f12615a64bd008ed5be6d1ebcb23579285d960a9/nvidia_cusparse-12.6.3.3-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:80bcc4662f23f1054ee334a15c72b8940402975e0eab63178fc7e670aa59472c", size = 162155568, upload-time = "2025-09-04T08:33:42.864Z" },
@@ -2596,11 +2828,11 @@ wheels = [
[[package]]
name = "nvidia-cusparselt-cu13"
-version = "0.8.0"
+version = "0.8.1"
source = { registry = "https://pypi.org/simple" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/46/10/8dcd1175260706a2fc92a16a52e306b71d4c1ea0b0cc4a9484183399818a/nvidia_cusparselt_cu13-0.8.0-py3-none-manylinux2014_aarch64.whl", hash = "sha256:400c6ed1cf6780fc6efedd64ec9f1345871767e6a1a0a552a1ea0578117ea77c", size = 220791277, upload-time = "2025-08-13T19:22:40.982Z" },
- { url = "https://files.pythonhosted.org/packages/fd/53/43b0d71f4e702fa9733f8b4571fdca50a8813f1e450b656c239beff12315/nvidia_cusparselt_cu13-0.8.0-py3-none-manylinux2014_x86_64.whl", hash = "sha256:25e30a8a7323935d4ad0340b95a0b69926eee755767e8e0b1cf8dd85b197d3fd", size = 169884119, upload-time = "2025-08-13T19:23:41.967Z" },
+ { url = "https://files.pythonhosted.org/packages/46/e1/cdc1797eadf82d3a9a575a19b33fdc871a97edbec42c00b5b5e914f4aff4/nvidia_cusparselt_cu13-0.8.1-py3-none-manylinux2014_aarch64.whl", hash = "sha256:4dca476c50bf4780d46cd0bfbd82e2bc10a08e4fef7950917ce8d7578d22a23f", size = 221051344, upload-time = "2025-09-05T18:49:51.289Z" },
+ { url = "https://files.pythonhosted.org/packages/34/7d/2661f2fb3ac4302f3a246f5fc030213ac60c1fe0bce84f9783dbd831dbb7/nvidia_cusparselt_cu13-0.8.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:786ce87568c303fadb5afcc7102d454cd3040d75f6f8626f5db460d1871f4dd0", size = 170148586, upload-time = "2025-09-05T18:50:50.248Z" },
]
[[package]]
@@ -2614,11 +2846,11 @@ wheels = [
[[package]]
name = "nvidia-nccl-cu13"
-version = "2.28.9"
+version = "2.29.7"
source = { registry = "https://pypi.org/simple" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/39/55/1920646a2e43ffd4fc958536b276197ed740e9e0c54105b4bb3521591fc7/nvidia_nccl_cu13-2.28.9-py3-none-manylinux_2_18_aarch64.whl", hash = "sha256:01c873ba1626b54caa12272ed228dc5b2781545e0ae8ba3f432a8ef1c6d78643", size = 196561677, upload-time = "2025-11-18T05:49:03.45Z" },
- { url = "https://files.pythonhosted.org/packages/b0/b4/878fefaad5b2bcc6fcf8d474a25e3e3774bc5133e4b58adff4d0bca238bc/nvidia_nccl_cu13-2.28.9-py3-none-manylinux_2_18_x86_64.whl", hash = "sha256:e4553a30f34195f3fa1da02a6da3d6337d28f2003943aa0a3d247bbc25fefc42", size = 196493177, upload-time = "2025-11-18T05:49:17.677Z" },
+ { url = "https://files.pythonhosted.org/packages/72/0d/daf50d44177ee0cbc7ff0a0c91eb5ff676c82be42f9a970bc7597f440c3a/nvidia_nccl_cu13-2.29.7-py3-none-manylinux_2_18_aarch64.whl", hash = "sha256:674a12383e3c38a1bcccae7d4f3633b37852230b6047883cb2f4c2d1b36d9bf5", size = 206014712, upload-time = "2026-03-03T05:34:20.843Z" },
+ { url = "https://files.pythonhosted.org/packages/67/f4/58e4e91b6919367c7aafb8e36fce9aad1a3047e536bf7e2fd560927d3a4c/nvidia_nccl_cu13-2.29.7-py3-none-manylinux_2_18_x86_64.whl", hash = "sha256:edd81538446786ec3b73972543e53bb43bcaf0bfc8ef76cb679fcc390ffe136d", size = 205976000, upload-time = "2026-03-03T05:36:24.472Z" },
]
[[package]]
@@ -2780,6 +3012,42 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/e9/a5/1be1516390333ff9be3a9cb648c9f33df79d5096e5884b5df71a588af463/opencv_python-4.13.0.92-cp37-abi3-win_amd64.whl", hash = "sha256:423d934c9fafb91aad38edf26efb46da91ffbc05f3f59c4b0c72e699720706f5", size = 40212062, upload-time = "2026-02-05T07:02:12.724Z" },
]
+[[package]]
+name = "openvino"
+version = "2026.2.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "numpy" },
+ { name = "openvino-telemetry" },
+]
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/c9/25/5c3e836afd92c9c07a411ea9af9e9ac3fff40aae612e93c5c0a997da83a1/openvino-2026.2.1-21919-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:7afed0219cd29fa73f54b7351ffab8e7c7fefb64290dd5264d7c447617ee09ff", size = 31169975, upload-time = "2026-06-17T09:20:58.625Z" },
+ { url = "https://files.pythonhosted.org/packages/62/c2/46b88970c6d8ff40703b58a96bec40c354ca1e6977511e01af0fa0f12b8a/openvino-2026.2.1-21919-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:efae5ceb94943aeab8a3be04241cbb89110070ce9bacfbfbe947d8d7ef691a82", size = 58367861, upload-time = "2026-06-17T09:21:02.968Z" },
+ { url = "https://files.pythonhosted.org/packages/a0/1f/e43283e3a2879b0ebdf1b4591461504255bfc639471ea3698a35ebcbcab2/openvino-2026.2.1-21919-cp312-cp312-manylinux_2_35_aarch64.whl", hash = "sha256:6154b26f63804cb671e20b8872abccafbe88b15fdd41fbd35ffa08bb4ac5e29b", size = 28337803, upload-time = "2026-06-17T09:21:06.642Z" },
+ { url = "https://files.pythonhosted.org/packages/93/81/f4605ab2bdbb3daff7cadf7d5265e87c4fe95d01af00ae8e1638b9a4abe9/openvino-2026.2.1-21919-cp312-cp312-win_amd64.whl", hash = "sha256:6928a707c447dbd43dab62c8bbc24a7953fd1b5e1fbea0f89a4135ca1ee4083b", size = 76171228, upload-time = "2026-06-17T09:21:11.156Z" },
+ { url = "https://files.pythonhosted.org/packages/64/b5/5788f1ce7906f33f003de3bb55c2477d9911c19f72d7ed3f7b719538d4a3/openvino-2026.2.1-21919-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:3ddd393566ab5cdcf106028e532f0f3789d7796fac848284c4d2ef31e3b81ef8", size = 31170279, upload-time = "2026-06-17T09:21:15.214Z" },
+ { url = "https://files.pythonhosted.org/packages/ba/41/7045a639e3879b0c858533f64c93e9c998f292b810dfaff58c23540c738a/openvino-2026.2.1-21919-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:00483723163c8d4d6f436ad0e4d05247494de4d280e4b065f67c537b440f4b56", size = 58367437, upload-time = "2026-06-17T09:21:20.215Z" },
+ { url = "https://files.pythonhosted.org/packages/ba/91/3f402a78755c1c848128126f145fc32a6d091c2257c543522d9ca979456a/openvino-2026.2.1-21919-cp313-cp313-manylinux_2_35_aarch64.whl", hash = "sha256:f9619c51027b6b0514e82ade09794b109a45553e8afcb073eedb1c8c1ffd43c4", size = 28344673, upload-time = "2026-06-17T09:21:23.873Z" },
+ { url = "https://files.pythonhosted.org/packages/ee/e5/13b9bfc20c02706d015e96e5b3e932227d658b3826b4aa5007cdc2455ed4/openvino-2026.2.1-21919-cp313-cp313-win_amd64.whl", hash = "sha256:197435104d18657a7cc4d05850615b913618ebb0d3baaec052fda25a2d831c74", size = 76171396, upload-time = "2026-06-17T09:21:29.03Z" },
+ { url = "https://files.pythonhosted.org/packages/1d/c0/1bd1e01be4eb7434711638b551521b3859f6df13a7b5cdaf591870ee9065/openvino-2026.2.1-21919-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:3f1dfd9d8ffbe21c2d40089c8137903432535b89ee8eb3d177b75604f4273ad4", size = 31151204, upload-time = "2026-06-17T09:21:32.973Z" },
+ { url = "https://files.pythonhosted.org/packages/43/65/b7f2a382e1da48c6cce109e7960714aa0a0608f2fb6b00388594b5d8edd2/openvino-2026.2.1-21919-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:0ec716ddd84a63534613171d27c8657c8417734a1cccedeb40533e8673a12c6d", size = 58369839, upload-time = "2026-06-17T09:21:37.636Z" },
+ { url = "https://files.pythonhosted.org/packages/45/4e/c0730932f7e0c34f627cedb569ac1f21222c4c5e0ed904a7f86e4508849d/openvino-2026.2.1-21919-cp314-cp314-manylinux_2_35_aarch64.whl", hash = "sha256:6fb513f36dde11dcccf5143e44b37bb64368689e3d746c283e9907e181e32226", size = 28351270, upload-time = "2026-06-17T09:21:41.833Z" },
+ { url = "https://files.pythonhosted.org/packages/5c/2c/c4f651c675c0271efdf7733c1a1adc9bfbf6cfcb5fce384565582f65e646/openvino-2026.2.1-21919-cp314-cp314-win_amd64.whl", hash = "sha256:c53788fcf66b5059eb5feeaba8f96782f3b672ea760df7cebbb96a28467eb472", size = 76173523, upload-time = "2026-06-17T09:21:47.05Z" },
+ { url = "https://files.pythonhosted.org/packages/18/e6/8c6942a06b86bdea8e55c630770946f5ef5d17a4131cec2ff677a1d836fb/openvino-2026.2.1-21919-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:8a4d99ffc01216fe7fbb41b246db2d92ae00a7e980de9e1c8ead5eac81ca3202", size = 31379331, upload-time = "2026-06-17T09:21:50.722Z" },
+ { url = "https://files.pythonhosted.org/packages/38/71/890efb337ad5d0ab1e70bfb0e3382bbb9efaed1f67dd27b77a275acbec13/openvino-2026.2.1-21919-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:863c55be074494a0e879f780eb2a7b25cf102105fd1835af1046376fbcb35a93", size = 58413666, upload-time = "2026-06-17T09:21:55.225Z" },
+ { url = "https://files.pythonhosted.org/packages/aa/06/56015860afcb19f20a684a2b1fa6f8662b53db7dde0b5219ce73e250d5f4/openvino-2026.2.1-21919-cp314-cp314t-manylinux_2_35_aarch64.whl", hash = "sha256:daf95cfd841070e17588f66826a59cccb562beceb98ce78637e3296dfecbce3c", size = 25494110, upload-time = "2026-06-17T09:21:59.279Z" },
+ { url = "https://files.pythonhosted.org/packages/a5/60/a8fd84f8289003dec261840f32a06476c793830d4562c2c2d76724ffdcc6/openvino-2026.2.1-21919-cp314-cp314t-win_amd64.whl", hash = "sha256:7ebc506896b082acd744b479d400e090ed566c7d1ff64a0071f2d54460032f9a", size = 76338896, upload-time = "2026-06-17T09:22:05.122Z" },
+]
+
+[[package]]
+name = "openvino-telemetry"
+version = "2025.2.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/71/8a/89d82f1a9d913fb266c2e6dc2f6030935db24b7152963a8db6c4f039787f/openvino_telemetry-2025.2.0.tar.gz", hash = "sha256:8bf8127218e51e99547bf38b8fb85a8b31c9bf96e6f3a82eb0b3b6a34155977c", size = 18894, upload-time = "2025-07-07T10:29:51.159Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/3b/ac/5ab0ca0aa269ad3c73f7bfc3801b10e5f56f75a31bf68c1ae8bd51cf70a4/openvino_telemetry-2025.2.0-py3-none-any.whl", hash = "sha256:bcb667e83a44f202ecf4cfa49281715c6d7e21499daec04ff853b7f964833599", size = 25227, upload-time = "2025-07-07T10:29:50.189Z" },
+]
+
[[package]]
name = "opt-einsum"
version = "3.4.0"
@@ -2789,6 +3057,78 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/23/cd/066e86230ae37ed0be70aae89aabf03ca8d9f39c8aea0dec8029455b5540/opt_einsum-3.4.0-py3-none-any.whl", hash = "sha256:69bb92469f86a1565195ece4ac0323943e83477171b91d24c35afe028a90d7cd", size = 71932, upload-time = "2024-09-26T14:33:23.039Z" },
]
+[[package]]
+name = "optree"
+version = "0.19.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/44/63/92328a17ab7836562fe0129e605f685a88db35ce98427c34ff48ee4ec157/optree-0.19.1.tar.gz", hash = "sha256:4497d1c9197b8c6842e511368163d318ce536521ebdcff8bebb7551dcdfac532", size = 177531, upload-time = "2026-05-06T02:32:39.704Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/ba/a7/cb5567029a608a296b0ca224025d03bba0365b41df19085b9b580191f6f2/optree-0.19.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:96e5c7c3b9144f08ae40c3d9848cfbcfa36b6bead0f8215ad071d5922ee6c4a5", size = 424023, upload-time = "2026-05-06T02:30:57.732Z" },
+ { url = "https://files.pythonhosted.org/packages/b9/a1/3651fb32fa8617108204aa4056d283af742020e0987d106f41402005d800/optree-0.19.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:5d9d198343e1e6ced18bef0cbff84091c1877964fc4a121df33f18840e073a01", size = 394782, upload-time = "2026-05-06T02:30:59.239Z" },
+ { url = "https://files.pythonhosted.org/packages/c2/1e/676470909aa64d7aba7c5edf83b171dc83b7af901d9ebb8e6d7512fe913a/optree-0.19.1-cp312-cp312-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7a1202371d9fe3aa75f3e886b1f871aac4991a655aadb65e54f58a3ae9388ab2", size = 413157, upload-time = "2026-05-06T02:31:00.339Z" },
+ { url = "https://files.pythonhosted.org/packages/f4/41/1a4c58f2af5742b9d9e21ea9e45c6c3c49463b5e2a0537e84ead1e9597ca/optree-0.19.1-cp312-cp312-manylinux_2_26_i686.manylinux_2_28_i686.whl", hash = "sha256:d41ccc4c20bfeae01d1d221c057a6d026e84e32229664952eddcdbe4b9b71417", size = 476923, upload-time = "2026-05-06T02:31:01.492Z" },
+ { url = "https://files.pythonhosted.org/packages/10/c1/f62167bd9d6f6c948b191a0943923404678d47100f777f4a8fb37816e6f8/optree-0.19.1-cp312-cp312-manylinux_2_26_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:7d934f240b109c6891dd06b2e30400b123b8a4b6ed31dcd0db2ae2378d30a6e8", size = 475385, upload-time = "2026-05-06T02:31:02.836Z" },
+ { url = "https://files.pythonhosted.org/packages/30/5e/5323c5fa3024fdd900bdd8f14621139ed844c2247bf1a26e7cf5c1116188/optree-0.19.1-cp312-cp312-manylinux_2_26_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ddeefb7ca799c09647e332ebc1a5f6c09888a5a0e51f2dff4ca55e65b42a8c14", size = 474406, upload-time = "2026-05-06T02:31:04.023Z" },
+ { url = "https://files.pythonhosted.org/packages/e2/6a/54e4c47e61a51504a5224c933722e0c8a69925aacec4c08175e9675aeb81/optree-0.19.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f0ce49f64f804f7f35f2f9c2a21e3ba94c090199fccdcfd40e3ded4426c5c175", size = 457596, upload-time = "2026-05-06T02:31:05.695Z" },
+ { url = "https://files.pythonhosted.org/packages/a7/12/bba07c0b769586c6bd54e81f1f734cad103dbe30abbadee940fe7d3e330e/optree-0.19.1-cp312-cp312-manylinux_2_39_riscv64.whl", hash = "sha256:e0f02600832ab8d0f6c934dcb5c339e17a36938d477641a45798e02625ebe107", size = 417900, upload-time = "2026-05-06T02:31:07.251Z" },
+ { url = "https://files.pythonhosted.org/packages/9f/8f/6ae994bb47f9394b33912a14593f9247737dd6c3303811550e5a3e918107/optree-0.19.1-cp312-cp312-win32.whl", hash = "sha256:f10d58c1a17e1b32f9d9b5e1b9d1ad964d99c1113d9df0b9f62f2fe7dde19909", size = 317302, upload-time = "2026-05-06T02:31:08.627Z" },
+ { url = "https://files.pythonhosted.org/packages/31/97/d7e3ec79dcdde81f785a0446acf75fea77723f5ca4b98556350d7877986f/optree-0.19.1-cp312-cp312-win_amd64.whl", hash = "sha256:06f5c8a4cf356a1a276ce5cec1be44719ed260690f79c036d04b4d427e801258", size = 341362, upload-time = "2026-05-06T02:31:09.689Z" },
+ { url = "https://files.pythonhosted.org/packages/33/97/813afb84a81fd8ae65444730907c05f0775fd6c79d3359c9e84bd3370445/optree-0.19.1-cp312-cp312-win_arm64.whl", hash = "sha256:a33bd23fc5c67ecb9ff491b75fde10cd9b53f47f8a876de842090e8c7a2437e1", size = 351838, upload-time = "2026-05-06T02:31:11.086Z" },
+ { url = "https://files.pythonhosted.org/packages/c2/7b/0f2f3c9d55dda5127624daf68ff802ab624b739dd4b32aef505dac0c8e02/optree-0.19.1-cp313-cp313-android_24_arm64_v8a.whl", hash = "sha256:f144cfd65fb17c6aa2c51818614eb009e6052d3d6ace91f7e570b1318cdcac4c", size = 929090, upload-time = "2026-05-06T02:31:12.267Z" },
+ { url = "https://files.pythonhosted.org/packages/15/e2/670d260dfd0532d64272dd6f7edd540a09d7040c0342b6cc6cf773568ea4/optree-0.19.1-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:39a006735d2a0a68751a3bc33d670184fddcd86db63b0293e1e819739e8105e4", size = 391528, upload-time = "2026-05-06T02:31:14.212Z" },
+ { url = "https://files.pythonhosted.org/packages/f4/96/46c15e80b0c97e2ba6aba11339008a37cabc5ccf55c31c6c60aecdb79638/optree-0.19.1-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:d2cb43c36638f469f5d8f4cf638e914de90c62242d8bed29f1b4487e0346ab94", size = 398231, upload-time = "2026-05-06T02:31:15.519Z" },
+ { url = "https://files.pythonhosted.org/packages/7e/39/9d7d22cdaeb9a40ace2485f91c5b7c5f3a7f688575e2621e436561211cc1/optree-0.19.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:e70faa00ab69331f49f8337d45021bed09ae2265d1db72eea9d7817af2b73c64", size = 429852, upload-time = "2026-05-06T02:31:16.992Z" },
+ { url = "https://files.pythonhosted.org/packages/79/4c/1da9e8375e7b7fd9671dc5987682b042f6412c4d6fd9da03296403818d9f/optree-0.19.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:1c5d21176b670407f4555aae40711668832599c4fb0627000c5ce3ed0d6e2dae", size = 398688, upload-time = "2026-05-06T02:31:18.113Z" },
+ { url = "https://files.pythonhosted.org/packages/d3/50/cd2d178099618093f5a9fd1c9de80af2b428879922eae1e9f27f1002c8be/optree-0.19.1-cp313-cp313-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f658fa46305b2bdccdc5bb2cb07818aeaef88a1085499deda5be48a0a58d2971", size = 417560, upload-time = "2026-05-06T02:31:19.391Z" },
+ { url = "https://files.pythonhosted.org/packages/d7/b0/f22ff5632083b5032caa80208dd202f8e963ed4aac11afa0a0f6a307fd68/optree-0.19.1-cp313-cp313-manylinux_2_26_i686.manylinux_2_28_i686.whl", hash = "sha256:e757079d44a00319447f43df5c51e55bf9b62d9f05eea0e2db5ff7c7ca5ec71d", size = 482937, upload-time = "2026-05-06T02:31:20.799Z" },
+ { url = "https://files.pythonhosted.org/packages/7d/d4/7499d28be8b11eb40668262d27802119fe7e6ec4cd8816b76a1acd7b08f5/optree-0.19.1-cp313-cp313-manylinux_2_26_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9690c132822d9dee479cf7dff8cc52a67c8af42a4f7529d21f0f4f1d99e4c84e", size = 477864, upload-time = "2026-05-06T02:31:22.077Z" },
+ { url = "https://files.pythonhosted.org/packages/b1/6e/6c6fa6f1159ac68f4ee7666610127fb4c14d47a2fa7a0a48de3aecc24d4b/optree-0.19.1-cp313-cp313-manylinux_2_26_s390x.manylinux_2_28_s390x.whl", hash = "sha256:544b70958dbd7e732bc6874e0180c609c9052115937d0ec28123bb49c1a574aa", size = 478319, upload-time = "2026-05-06T02:31:23.266Z" },
+ { url = "https://files.pythonhosted.org/packages/68/b5/8a2427bbe4ee59e2ce26a14125728e3b48c7030c80984ba07d0e5d804d37/optree-0.19.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9dde5b756946c1f1458aeab248a7a9b0c01bb06b5787de9f06d52ad38b745557", size = 462379, upload-time = "2026-05-06T02:31:24.543Z" },
+ { url = "https://files.pythonhosted.org/packages/ee/0c/a073eeaea4d4f68e02d5883ed8268746a296e6749e3c46e0124ca45f306c/optree-0.19.1-cp313-cp313-manylinux_2_39_riscv64.whl", hash = "sha256:f1d7838e8b1b62258abd73a5911afad1153ed76822070558c3ba7e0bb5b44192", size = 423061, upload-time = "2026-05-06T02:31:25.652Z" },
+ { url = "https://files.pythonhosted.org/packages/5f/34/637b151d071ca94aea0087322f470ce84c5828ef6b9c0de7dc7b4420a1cf/optree-0.19.1-cp313-cp313-win32.whl", hash = "sha256:9870d33ec50cca0c46c2b431cea24c6247457da15fd4ad66ccb8ab78145c1490", size = 317439, upload-time = "2026-05-06T02:31:27.304Z" },
+ { url = "https://files.pythonhosted.org/packages/50/52/49b8a8d9e94c57c6fa5008953f84a1c36a4119a3b90dcb7df745f1f05a00/optree-0.19.1-cp313-cp313-win_amd64.whl", hash = "sha256:aa0845b725bcd0029e179cf9b4bc2cc016c7358e56fc7c0d2c43bf4d514c96cf", size = 343906, upload-time = "2026-05-06T02:31:28.774Z" },
+ { url = "https://files.pythonhosted.org/packages/c6/a9/1ae0a9685f5301f454f01d2490065b98df6956f90b1b2fd1cea9daa6d820/optree-0.19.1-cp313-cp313-win_arm64.whl", hash = "sha256:6f0b1efc177bed6495f78d39d5aa495ccb31cc20bcf64bb1b806ca4c919f4049", size = 353146, upload-time = "2026-05-06T02:31:29.976Z" },
+ { url = "https://files.pythonhosted.org/packages/9c/77/4c8108cbce2c8ae2aa4b6adc7874082882e32cf131cb64b3a4411f50dec4/optree-0.19.1-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:b964bcdb5cfe367cdf56447e80ba5a49123098d8c4e8e68b41c20890eec6e58e", size = 469723, upload-time = "2026-05-06T02:31:31.425Z" },
+ { url = "https://files.pythonhosted.org/packages/64/33/ce9b54646ed4ab5773a9dc59767dadfe3de8bb2e97a3ed19205b995a7a31/optree-0.19.1-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:08ccec0ee5a565eb5aa4fe30383016a358627ea23d968ec8ab28b1f2ce4ce3d8", size = 437071, upload-time = "2026-05-06T02:31:33.027Z" },
+ { url = "https://files.pythonhosted.org/packages/79/55/04260128a726e3550b49467a65bff859452897144b68bae54b2f2e5c27f1/optree-0.19.1-cp313-cp313t-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:672588408906051d3e9a99aca6c0af93c6e0b638137a701418088eaa0bb6c719", size = 433503, upload-time = "2026-05-06T02:31:34.423Z" },
+ { url = "https://files.pythonhosted.org/packages/d6/99/6a4cc29389667efa089a0c476b7c36b7d0a66e10dd2d8c2d19c776977566/optree-0.19.1-cp313-cp313t-manylinux_2_26_i686.manylinux_2_28_i686.whl", hash = "sha256:d16cef4d0555d49ce221d80249f1285a2d3faf932e451c3ce6cb8ccb6a846767", size = 496305, upload-time = "2026-05-06T02:31:35.835Z" },
+ { url = "https://files.pythonhosted.org/packages/7f/46/506aa1a64abce69e2f4cec9cdac3da0cae207cf04c5e70e7f143bf8b29d8/optree-0.19.1-cp313-cp313t-manylinux_2_26_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:dc2db0b449baff53aa7e583306101de0ade5e5ae9e6fce78400eb2319bbd23dc", size = 492759, upload-time = "2026-05-06T02:31:37.265Z" },
+ { url = "https://files.pythonhosted.org/packages/f5/28/2210de9a68722007fe007da3cae1a5971b92fc8113b5eecef66a04637959/optree-0.19.1-cp313-cp313t-manylinux_2_26_s390x.manylinux_2_28_s390x.whl", hash = "sha256:76b3e9e5d37e6b05ec82fff91758c8c0e27e159b35faea4b33d5eb975d720257", size = 495447, upload-time = "2026-05-06T02:31:38.505Z" },
+ { url = "https://files.pythonhosted.org/packages/d9/61/40c3463e52914d552c66c760ae15e673137c4cc1d1d9f8da0d745656193a/optree-0.19.1-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:03faa8e23fdaf3a18f9a1568c2c0eb0641a6aa05baf3a20639bd11fb34664700", size = 475564, upload-time = "2026-05-06T02:31:39.732Z" },
+ { url = "https://files.pythonhosted.org/packages/0a/66/1603680fa924e68e5697c1229510c0645db0a9c633a12d1a9bfdbfc9cb74/optree-0.19.1-cp313-cp313t-manylinux_2_39_riscv64.whl", hash = "sha256:a9b9c7e9148ec470124dc4c1d1cd1485dbeb35973357b5911b181a79090426d2", size = 442414, upload-time = "2026-05-06T02:31:40.908Z" },
+ { url = "https://files.pythonhosted.org/packages/a5/58/34820bab11f28ba6b03fe9e151880ad591b43f26648f058c94451fbdfc3a/optree-0.19.1-cp313-cp313t-win32.whl", hash = "sha256:ab8ad9803376d553a2958471b6bb6842b7e15888e19cc6aeb76da96c6afd948d", size = 348644, upload-time = "2026-05-06T02:31:42.038Z" },
+ { url = "https://files.pythonhosted.org/packages/d9/2b/0be3f8b9765f366e3e12d0590e9c6514de110d0c5b3b9002f49e56bf15b1/optree-0.19.1-cp313-cp313t-win_amd64.whl", hash = "sha256:afd4abeb2783b2367093287bc6268ac9af244b20c8d9b01696ccfe817483b66c", size = 382445, upload-time = "2026-05-06T02:31:43.166Z" },
+ { url = "https://files.pythonhosted.org/packages/fc/fa/8c0882cdd42e28a23c1998297c8ad1202194510cbba8b050251429c641c0/optree-0.19.1-cp313-cp313t-win_arm64.whl", hash = "sha256:b9120510d3f951e268e417a3f64f335bc1c539e1e80bff2129ddc6fb60ac7b56", size = 388040, upload-time = "2026-05-06T02:31:44.661Z" },
+ { url = "https://files.pythonhosted.org/packages/e3/da/4e16e26375c56c9e40760697af4e2b72f196c2099e96cc783b63dcc862a8/optree-0.19.1-cp314-cp314-android_24_arm64_v8a.whl", hash = "sha256:e1951ddc870f67430310fd17393971c30510ee9fd290525b44c12afe25f3c307", size = 927808, upload-time = "2026-05-06T02:31:45.954Z" },
+ { url = "https://files.pythonhosted.org/packages/6b/87/ff1c6bb6b79a5d0b70b83f7ae8b78811a406a749b3ae4478a2122a7afb66/optree-0.19.1-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:ae9d42718ebf985cdad3182364b5cf829193b8fd2c6d993fbb4111d38e2bdf96", size = 390981, upload-time = "2026-05-06T02:31:47.38Z" },
+ { url = "https://files.pythonhosted.org/packages/82/25/fc648710102960f87d18cd8fc8a24afe14a5ec7827c64dfb1340230c0794/optree-0.19.1-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:930268ebfdebca43a8808f6293910d6ade2fe7c84fa784692017d7120d285226", size = 397756, upload-time = "2026-05-06T02:31:48.76Z" },
+ { url = "https://files.pythonhosted.org/packages/24/f6/a7bf5d75a6481038bbb61846d87d43124d63741385796ef7b37d326f46bd/optree-0.19.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:b2757c5d922aab76cfc9b870c373fb35209c2094e3c912733b326c043e85a0c6", size = 427424, upload-time = "2026-05-06T02:31:49.838Z" },
+ { url = "https://files.pythonhosted.org/packages/49/cc/14dd93887295859457e507fc46a847b68ae8f20c42b2fde4d8a749c94bbc/optree-0.19.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:b17a7b70ff8bd406c2142914c5ab0a57f8bcfb9f52181f7012e32406bbdbfdda", size = 398242, upload-time = "2026-05-06T02:31:51.262Z" },
+ { url = "https://files.pythonhosted.org/packages/17/b5/ac51aa118dd918761519fbc031865b1d6f850453e9a7ac0c3da21109c4f0/optree-0.19.1-cp314-cp314-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:987bba55366917d9829f45b5ee86499ecc87a30e9103072db9ab8d67f9958179", size = 419568, upload-time = "2026-05-06T02:31:52.349Z" },
+ { url = "https://files.pythonhosted.org/packages/ec/41/25144e61f76278b9e0a5d4189c7083fe853164c5f7328a1f5aac43d964c2/optree-0.19.1-cp314-cp314-manylinux_2_26_i686.manylinux_2_28_i686.whl", hash = "sha256:d3bba2af7a5fce0c25e99024688e68dfe9be41e3d6e92720febbefdc879fba38", size = 482797, upload-time = "2026-05-06T02:31:53.471Z" },
+ { url = "https://files.pythonhosted.org/packages/22/47/2c76c7ce937323988770c41126e0e380bcb73a816f68a767f23b5c33aced/optree-0.19.1-cp314-cp314-manylinux_2_26_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:dae6c247cc8751bd2f167951468769f5c98f8cfdae31c0db0f2eb4145a6ec560", size = 479794, upload-time = "2026-05-06T02:31:54.843Z" },
+ { url = "https://files.pythonhosted.org/packages/7c/ca/bd9553f94bec0bc7860f10ae177c14ca265ab19ddb463122be22fa335ee8/optree-0.19.1-cp314-cp314-manylinux_2_26_s390x.manylinux_2_28_s390x.whl", hash = "sha256:17a986fd91ccdc18bb7b587ca1f508c1761580a93517e6db33a13b22e46acb9b", size = 481084, upload-time = "2026-05-06T02:31:56.261Z" },
+ { url = "https://files.pythonhosted.org/packages/9c/1a/4834b1f2fb1847412353d7342eb7a1d001a4f3bd9d24155e057135a4aa44/optree-0.19.1-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3d0e1493429ae1d1a5e34855774ee604c974a8f76656bd0e562cdbf9466c9b1f", size = 462955, upload-time = "2026-05-06T02:31:57.829Z" },
+ { url = "https://files.pythonhosted.org/packages/f4/88/598fb91c06fee3d8b08568779b011225dc2b66140927bd0b2b2d9b40a566/optree-0.19.1-cp314-cp314-manylinux_2_39_riscv64.whl", hash = "sha256:f61a01ed9991193ed6f3db8e956ede05218190a32ca2ddfb71cfc40c8daba1d5", size = 423754, upload-time = "2026-05-06T02:31:59.291Z" },
+ { url = "https://files.pythonhosted.org/packages/20/8a/83c64ecadc686e08310fc9c20bc0bbe6453e89b69257e08887818dac7886/optree-0.19.1-cp314-cp314-win32.whl", hash = "sha256:b0c920579bddc3b18a0e051850f017618e24efcc19ba83dcd415cf74db5fd904", size = 325214, upload-time = "2026-05-06T02:32:00.802Z" },
+ { url = "https://files.pythonhosted.org/packages/96/c3/4f2f318b98465376bbb7a06a33da553c688b3ed39dafbb8307f824eef74a/optree-0.19.1-cp314-cp314-win_amd64.whl", hash = "sha256:50d77b91a8cd01adf422472b7edf39fc445b0268816176a868a385d28f8367c2", size = 351654, upload-time = "2026-05-06T02:32:01.944Z" },
+ { url = "https://files.pythonhosted.org/packages/1c/ab/55d7508e87055c730fe7207cfd0c45183a07ddf1f91d9e73d017a7f8c1f4/optree-0.19.1-cp314-cp314-win_arm64.whl", hash = "sha256:c682ab6711b7a623503711fa661a2bba7886e1c21dc06c3b7febba101b458051", size = 361610, upload-time = "2026-05-06T02:32:03.003Z" },
+ { url = "https://files.pythonhosted.org/packages/ae/2d/4f7facd482d56079b7adb8ce3fede19f41629bc0463e8ee25907f1dba36c/optree-0.19.1-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:068edb89fadd94f6f57fdb51f4ad2c764b5a0bfd00903c55ffe433c2863a8037", size = 469130, upload-time = "2026-05-06T02:32:04.395Z" },
+ { url = "https://files.pythonhosted.org/packages/92/60/f7539012aa8a7488c1e34f66b76eadc384c3152dd9800973f1b5fe045dfd/optree-0.19.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:a609c90e4f64e4f3e2b5b3cc022210314834737e0e61a745485e33b33eae773b", size = 437286, upload-time = "2026-05-06T02:32:05.527Z" },
+ { url = "https://files.pythonhosted.org/packages/9b/3f/a5f8fb3ec3840f885de52d7a793ba57ace17990e3a9b3797218425ffe842/optree-0.19.1-cp314-cp314t-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:dfae64c4c371640a4b3e2a9e3e6aa3a3e8cdf2da5247a88fef5b632614b948a6", size = 431954, upload-time = "2026-05-06T02:32:06.83Z" },
+ { url = "https://files.pythonhosted.org/packages/68/dc/6d0ef14bc82bd54046c1a066d25fa6854123a6b29fd691f1f95dec3ab45f/optree-0.19.1-cp314-cp314t-manylinux_2_26_i686.manylinux_2_28_i686.whl", hash = "sha256:470742544ff2d4b63843023f38dcfb83e82c3a9877c783dee0e69cbb974de6d1", size = 494631, upload-time = "2026-05-06T02:32:08.038Z" },
+ { url = "https://files.pythonhosted.org/packages/b8/9a/9e183c610c414cba581f9afda7610589d89cae229d627b14f8480425d975/optree-0.19.1-cp314-cp314t-manylinux_2_26_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:1a74e0656ccef45b1fec07b9d964ce97f3def8bab73711f56175076c4259884f", size = 491786, upload-time = "2026-05-06T02:32:09.363Z" },
+ { url = "https://files.pythonhosted.org/packages/4d/73/266b9de8eb5b16bfe7010c90c55840517d5d61ee6e0ca64901440296d97a/optree-0.19.1-cp314-cp314t-manylinux_2_26_s390x.manylinux_2_28_s390x.whl", hash = "sha256:f55841132ba8a34dbbd85e0c2cf990602384eea0e4638df986cd3266482f4a17", size = 490876, upload-time = "2026-05-06T02:32:11.388Z" },
+ { url = "https://files.pythonhosted.org/packages/b3/8d/42a8ca6277ef93d47ab0986e30a25134206afe0c6e6c3425c8736b2677ba/optree-0.19.1-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a5f8383952f18d5a4ec6b248d8ae6fe27012434ad9750aa33a821ad4846da5af", size = 475079, upload-time = "2026-05-06T02:32:12.768Z" },
+ { url = "https://files.pythonhosted.org/packages/63/91/e363f4adda292f891ca0cf5748010fea955737bdf494cc11d4c3bcda6935/optree-0.19.1-cp314-cp314t-manylinux_2_39_riscv64.whl", hash = "sha256:de8acbed5965beae6f6b0456fcb8d1afaea1fe300810739e88645e22138849bc", size = 440119, upload-time = "2026-05-06T02:32:14.096Z" },
+ { url = "https://files.pythonhosted.org/packages/3f/eb/489d22ef3cadb2f5f3bbd6e6099d17b5a521ff533e086f78f005c3358017/optree-0.19.1-cp314-cp314t-win32.whl", hash = "sha256:312048e69dc88de26915674f961bf38980a765a6b48ead2f1672858a39402c41", size = 357465, upload-time = "2026-05-06T02:32:15.424Z" },
+ { url = "https://files.pythonhosted.org/packages/e0/34/7f48b7034ff75d2eb3e94e2196709ddbf762798fb621f9508899fa66b44e/optree-0.19.1-cp314-cp314t-win_amd64.whl", hash = "sha256:60e9345405d7b06cafdf1b1dd2e2261ceddddce10f35729240f90e2bab845a0b", size = 397783, upload-time = "2026-05-06T02:32:16.853Z" },
+ { url = "https://files.pythonhosted.org/packages/07/42/6d6f93416c66820cb8753e65b5ff43c47480af9c4911bd2b8406ff0f7f27/optree-0.19.1-cp314-cp314t-win_arm64.whl", hash = "sha256:4e103e212d1e8fe0399ed076eff80a905fb14929729bbd994d3660110a27a252", size = 396064, upload-time = "2026-05-06T02:32:18.077Z" },
+]
+
[[package]]
name = "packaging"
version = "26.2"
@@ -2861,71 +3201,73 @@ wheels = [
[[package]]
name = "pillow"
-version = "12.2.0"
-source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/8c/21/c2bcdd5906101a30244eaffc1b6e6ce71a31bd0742a01eb89e660ebfac2d/pillow-12.2.0.tar.gz", hash = "sha256:a830b1a40919539d07806aa58e1b114df53ddd43213d9c8b75847eee6c0182b5", size = 46987819, upload-time = "2026-04-01T14:46:17.687Z" }
-wheels = [
- { url = "https://files.pythonhosted.org/packages/58/be/7482c8a5ebebbc6470b3eb791812fff7d5e0216c2be3827b30b8bb6603ed/pillow-12.2.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:2d192a155bbcec180f8564f693e6fd9bccff5a7af9b32e2e4bf8c9c69dbad6b5", size = 5308279, upload-time = "2026-04-01T14:43:13.246Z" },
- { url = "https://files.pythonhosted.org/packages/d8/95/0a351b9289c2b5cbde0bacd4a83ebc44023e835490a727b2a3bd60ddc0f4/pillow-12.2.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:f3f40b3c5a968281fd507d519e444c35f0ff171237f4fdde090dd60699458421", size = 4695490, upload-time = "2026-04-01T14:43:15.584Z" },
- { url = "https://files.pythonhosted.org/packages/de/af/4e8e6869cbed569d43c416fad3dc4ecb944cb5d9492defaed89ddd6fe871/pillow-12.2.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:03e7e372d5240cc23e9f07deca4d775c0817bffc641b01e9c3af208dbd300987", size = 6284462, upload-time = "2026-04-01T14:43:18.268Z" },
- { url = "https://files.pythonhosted.org/packages/e9/9e/c05e19657fd57841e476be1ab46c4d501bffbadbafdc31a6d665f8b737b6/pillow-12.2.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:b86024e52a1b269467a802258c25521e6d742349d760728092e1bc2d135b4d76", size = 8094744, upload-time = "2026-04-01T14:43:20.716Z" },
- { url = "https://files.pythonhosted.org/packages/2b/54/1789c455ed10176066b6e7e6da1b01e50e36f94ba584dc68d9eebfe9156d/pillow-12.2.0-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7371b48c4fa448d20d2714c9a1f775a81155050d383333e0a6c15b1123dda005", size = 6398371, upload-time = "2026-04-01T14:43:23.443Z" },
- { url = "https://files.pythonhosted.org/packages/43/e3/fdc657359e919462369869f1c9f0e973f353f9a9ee295a39b1fea8ee1a77/pillow-12.2.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:62f5409336adb0663b7caa0da5c7d9e7bdbaae9ce761d34669420c2a801b2780", size = 7087215, upload-time = "2026-04-01T14:43:26.758Z" },
- { url = "https://files.pythonhosted.org/packages/8b/f8/2f6825e441d5b1959d2ca5adec984210f1ec086435b0ed5f52c19b3b8a6e/pillow-12.2.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:01afa7cf67f74f09523699b4e88c73fb55c13346d212a59a2db1f86b0a63e8c5", size = 6509783, upload-time = "2026-04-01T14:43:29.56Z" },
- { url = "https://files.pythonhosted.org/packages/67/f9/029a27095ad20f854f9dba026b3ea6428548316e057e6fc3545409e86651/pillow-12.2.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:fc3d34d4a8fbec3e88a79b92e5465e0f9b842b628675850d860b8bd300b159f5", size = 7212112, upload-time = "2026-04-01T14:43:32.091Z" },
- { url = "https://files.pythonhosted.org/packages/be/42/025cfe05d1be22dbfdb4f264fe9de1ccda83f66e4fc3aac94748e784af04/pillow-12.2.0-cp312-cp312-win32.whl", hash = "sha256:58f62cc0f00fd29e64b29f4fd923ffdb3859c9f9e6105bfc37ba1d08994e8940", size = 6378489, upload-time = "2026-04-01T14:43:34.601Z" },
- { url = "https://files.pythonhosted.org/packages/5d/7b/25a221d2c761c6a8ae21bfa3874988ff2583e19cf8a27bf2fee358df7942/pillow-12.2.0-cp312-cp312-win_amd64.whl", hash = "sha256:7f84204dee22a783350679a0333981df803dac21a0190d706a50475e361c93f5", size = 7084129, upload-time = "2026-04-01T14:43:37.213Z" },
- { url = "https://files.pythonhosted.org/packages/10/e1/542a474affab20fd4a0f1836cb234e8493519da6b76899e30bcc5d990b8b/pillow-12.2.0-cp312-cp312-win_arm64.whl", hash = "sha256:af73337013e0b3b46f175e79492d96845b16126ddf79c438d7ea7ff27783a414", size = 2463612, upload-time = "2026-04-01T14:43:39.421Z" },
- { url = "https://files.pythonhosted.org/packages/4a/01/53d10cf0dbad820a8db274d259a37ba50b88b24768ddccec07355382d5ad/pillow-12.2.0-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:8297651f5b5679c19968abefd6bb84d95fe30ef712eb1b2d9b2d31ca61267f4c", size = 4100837, upload-time = "2026-04-01T14:43:41.506Z" },
- { url = "https://files.pythonhosted.org/packages/0f/98/f3a6657ecb698c937f6c76ee564882945f29b79bad496abcba0e84659ec5/pillow-12.2.0-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:50d8520da2a6ce0af445fa6d648c4273c3eeefbc32d7ce049f22e8b5c3daecc2", size = 4176528, upload-time = "2026-04-01T14:43:43.773Z" },
- { url = "https://files.pythonhosted.org/packages/69/bc/8986948f05e3ea490b8442ea1c1d4d990b24a7e43d8a51b2c7d8b1dced36/pillow-12.2.0-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:766cef22385fa1091258ad7e6216792b156dc16d8d3fa607e7545b2b72061f1c", size = 3640401, upload-time = "2026-04-01T14:43:45.87Z" },
- { url = "https://files.pythonhosted.org/packages/34/46/6c717baadcd62bc8ed51d238d521ab651eaa74838291bda1f86fe1f864c9/pillow-12.2.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:5d2fd0fa6b5d9d1de415060363433f28da8b1526c1c129020435e186794b3795", size = 5308094, upload-time = "2026-04-01T14:43:48.438Z" },
- { url = "https://files.pythonhosted.org/packages/71/43/905a14a8b17fdb1ccb58d282454490662d2cb89a6bfec26af6d3520da5ec/pillow-12.2.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:56b25336f502b6ed02e889f4ece894a72612fe885889a6e8c4c80239ff6e5f5f", size = 4695402, upload-time = "2026-04-01T14:43:51.292Z" },
- { url = "https://files.pythonhosted.org/packages/73/dd/42107efcb777b16fa0393317eac58f5b5cf30e8392e266e76e51cff28c3d/pillow-12.2.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:f1c943e96e85df3d3478f7b691f229887e143f81fedab9b20205349ab04d73ed", size = 6280005, upload-time = "2026-04-01T14:43:54.242Z" },
- { url = "https://files.pythonhosted.org/packages/a8/68/b93e09e5e8549019e61acf49f65b1a8530765a7f812c77a7461bca7e4494/pillow-12.2.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:03f6fab9219220f041c74aeaa2939ff0062bd5c364ba9ce037197f4c6d498cd9", size = 8090669, upload-time = "2026-04-01T14:43:57.335Z" },
- { url = "https://files.pythonhosted.org/packages/4b/6e/3ccb54ce8ec4ddd1accd2d89004308b7b0b21c4ac3d20fa70af4760a4330/pillow-12.2.0-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5cdfebd752ec52bf5bb4e35d9c64b40826bc5b40a13df7c3cda20a2c03a0f5ed", size = 6395194, upload-time = "2026-04-01T14:43:59.864Z" },
- { url = "https://files.pythonhosted.org/packages/67/ee/21d4e8536afd1a328f01b359b4d3997b291ffd35a237c877b331c1c3b71c/pillow-12.2.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:eedf4b74eda2b5a4b2b2fb4c006d6295df3bf29e459e198c90ea48e130dc75c3", size = 7082423, upload-time = "2026-04-01T14:44:02.74Z" },
- { url = "https://files.pythonhosted.org/packages/78/5f/e9f86ab0146464e8c133fe85df987ed9e77e08b29d8d35f9f9f4d6f917ba/pillow-12.2.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:00a2865911330191c0b818c59103b58a5e697cae67042366970a6b6f1b20b7f9", size = 6505667, upload-time = "2026-04-01T14:44:05.381Z" },
- { url = "https://files.pythonhosted.org/packages/ed/1e/409007f56a2fdce61584fd3acbc2bbc259857d555196cedcadc68c015c82/pillow-12.2.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:1e1757442ed87f4912397c6d35a0db6a7b52592156014706f17658ff58bbf795", size = 7208580, upload-time = "2026-04-01T14:44:08.39Z" },
- { url = "https://files.pythonhosted.org/packages/23/c4/7349421080b12fb35414607b8871e9534546c128a11965fd4a7002ccfbee/pillow-12.2.0-cp313-cp313-win32.whl", hash = "sha256:144748b3af2d1b358d41286056d0003f47cb339b8c43a9ea42f5fea4d8c66b6e", size = 6375896, upload-time = "2026-04-01T14:44:11.197Z" },
- { url = "https://files.pythonhosted.org/packages/3f/82/8a3739a5e470b3c6cbb1d21d315800d8e16bff503d1f16b03a4ec3212786/pillow-12.2.0-cp313-cp313-win_amd64.whl", hash = "sha256:390ede346628ccc626e5730107cde16c42d3836b89662a115a921f28440e6a3b", size = 7081266, upload-time = "2026-04-01T14:44:13.947Z" },
- { url = "https://files.pythonhosted.org/packages/c3/25/f968f618a062574294592f668218f8af564830ccebdd1fa6200f598e65c5/pillow-12.2.0-cp313-cp313-win_arm64.whl", hash = "sha256:8023abc91fba39036dbce14a7d6535632f99c0b857807cbbbf21ecc9f4717f06", size = 2463508, upload-time = "2026-04-01T14:44:16.312Z" },
- { url = "https://files.pythonhosted.org/packages/4d/a4/b342930964e3cb4dce5038ae34b0eab4653334995336cd486c5a8c25a00c/pillow-12.2.0-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:042db20a421b9bafecc4b84a8b6e444686bd9d836c7fd24542db3e7df7baad9b", size = 5309927, upload-time = "2026-04-01T14:44:18.89Z" },
- { url = "https://files.pythonhosted.org/packages/9f/de/23198e0a65a9cf06123f5435a5d95cea62a635697f8f03d134d3f3a96151/pillow-12.2.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:dd025009355c926a84a612fecf58bb315a3f6814b17ead51a8e48d3823d9087f", size = 4698624, upload-time = "2026-04-01T14:44:21.115Z" },
- { url = "https://files.pythonhosted.org/packages/01/a6/1265e977f17d93ea37aa28aa81bad4fa597933879fac2520d24e021c8da3/pillow-12.2.0-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:88ddbc66737e277852913bd1e07c150cc7bb124539f94c4e2df5344494e0a612", size = 6321252, upload-time = "2026-04-01T14:44:23.663Z" },
- { url = "https://files.pythonhosted.org/packages/3c/83/5982eb4a285967baa70340320be9f88e57665a387e3a53a7f0db8231a0cd/pillow-12.2.0-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:d362d1878f00c142b7e1a16e6e5e780f02be8195123f164edf7eddd911eefe7c", size = 8126550, upload-time = "2026-04-01T14:44:26.772Z" },
- { url = "https://files.pythonhosted.org/packages/4e/48/6ffc514adce69f6050d0753b1a18fd920fce8cac87620d5a31231b04bfc5/pillow-12.2.0-cp313-cp313t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2c727a6d53cb0018aadd8018c2b938376af27914a68a492f59dfcaca650d5eea", size = 6433114, upload-time = "2026-04-01T14:44:29.615Z" },
- { url = "https://files.pythonhosted.org/packages/36/a3/f9a77144231fb8d40ee27107b4463e205fa4677e2ca2548e14da5cf18dce/pillow-12.2.0-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:efd8c21c98c5cc60653bcb311bef2ce0401642b7ce9d09e03a7da87c878289d4", size = 7115667, upload-time = "2026-04-01T14:44:32.773Z" },
- { url = "https://files.pythonhosted.org/packages/c1/fc/ac4ee3041e7d5a565e1c4fd72a113f03b6394cc72ab7089d27608f8aaccb/pillow-12.2.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:9f08483a632889536b8139663db60f6724bfcb443c96f1b18855860d7d5c0fd4", size = 6538966, upload-time = "2026-04-01T14:44:35.252Z" },
- { url = "https://files.pythonhosted.org/packages/c0/a8/27fb307055087f3668f6d0a8ccb636e7431d56ed0750e07a60547b1e083e/pillow-12.2.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:dac8d77255a37e81a2efcbd1fc05f1c15ee82200e6c240d7e127e25e365c39ea", size = 7238241, upload-time = "2026-04-01T14:44:37.875Z" },
- { url = "https://files.pythonhosted.org/packages/ad/4b/926ab182c07fccae9fcb120043464e1ff1564775ec8864f21a0ebce6ac25/pillow-12.2.0-cp313-cp313t-win32.whl", hash = "sha256:ee3120ae9dff32f121610bb08e4313be87e03efeadfc6c0d18f89127e24d0c24", size = 6379592, upload-time = "2026-04-01T14:44:40.336Z" },
- { url = "https://files.pythonhosted.org/packages/c2/c4/f9e476451a098181b30050cc4c9a3556b64c02cf6497ea421ac047e89e4b/pillow-12.2.0-cp313-cp313t-win_amd64.whl", hash = "sha256:325ca0528c6788d2a6c3d40e3568639398137346c3d6e66bb61db96b96511c98", size = 7085542, upload-time = "2026-04-01T14:44:43.251Z" },
- { url = "https://files.pythonhosted.org/packages/00/a4/285f12aeacbe2d6dc36c407dfbbe9e96d4a80b0fb710a337f6d2ad978c75/pillow-12.2.0-cp313-cp313t-win_arm64.whl", hash = "sha256:2e5a76d03a6c6dcef67edabda7a52494afa4035021a79c8558e14af25313d453", size = 2465765, upload-time = "2026-04-01T14:44:45.996Z" },
- { url = "https://files.pythonhosted.org/packages/bf/98/4595daa2365416a86cb0d495248a393dfc84e96d62ad080c8546256cb9c0/pillow-12.2.0-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:3adc9215e8be0448ed6e814966ecf3d9952f0ea40eb14e89a102b87f450660d8", size = 4100848, upload-time = "2026-04-01T14:44:48.48Z" },
- { url = "https://files.pythonhosted.org/packages/0b/79/40184d464cf89f6663e18dfcf7ca21aae2491fff1a16127681bf1fa9b8cf/pillow-12.2.0-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:6a9adfc6d24b10f89588096364cc726174118c62130c817c2837c60cf08a392b", size = 4176515, upload-time = "2026-04-01T14:44:51.353Z" },
- { url = "https://files.pythonhosted.org/packages/b0/63/703f86fd4c422a9cf722833670f4f71418fb116b2853ff7da722ea43f184/pillow-12.2.0-cp314-cp314-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:6a6e67ea2e6feda684ed370f9a1c52e7a243631c025ba42149a2cc5934dec295", size = 3640159, upload-time = "2026-04-01T14:44:53.588Z" },
- { url = "https://files.pythonhosted.org/packages/71/e0/fb22f797187d0be2270f83500aab851536101b254bfa1eae10795709d283/pillow-12.2.0-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:2bb4a8d594eacdfc59d9e5ad972aa8afdd48d584ffd5f13a937a664c3e7db0ed", size = 5312185, upload-time = "2026-04-01T14:44:56.039Z" },
- { url = "https://files.pythonhosted.org/packages/ba/8c/1a9e46228571de18f8e28f16fabdfc20212a5d019f3e3303452b3f0a580d/pillow-12.2.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:80b2da48193b2f33ed0c32c38140f9d3186583ce7d516526d462645fd98660ae", size = 4695386, upload-time = "2026-04-01T14:44:58.663Z" },
- { url = "https://files.pythonhosted.org/packages/70/62/98f6b7f0c88b9addd0e87c217ded307b36be024d4ff8869a812b241d1345/pillow-12.2.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:22db17c68434de69d8ecfc2fe821569195c0c373b25cccb9cbdacf2c6e53c601", size = 6280384, upload-time = "2026-04-01T14:45:01.5Z" },
- { url = "https://files.pythonhosted.org/packages/5e/03/688747d2e91cfbe0e64f316cd2e8005698f76ada3130d0194664174fa5de/pillow-12.2.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:7b14cc0106cd9aecda615dd6903840a058b4700fcb817687d0ee4fc8b6e389be", size = 8091599, upload-time = "2026-04-01T14:45:04.5Z" },
- { url = "https://files.pythonhosted.org/packages/f6/35/577e22b936fcdd66537329b33af0b4ccfefaeabd8aec04b266528cddb33c/pillow-12.2.0-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8cbeb542b2ebc6fcdacabf8aca8c1a97c9b3ad3927d46b8723f9d4f033288a0f", size = 6396021, upload-time = "2026-04-01T14:45:07.117Z" },
- { url = "https://files.pythonhosted.org/packages/11/8d/d2532ad2a603ca2b93ad9f5135732124e57811d0168155852f37fbce2458/pillow-12.2.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4bfd07bc812fbd20395212969e41931001fd59eb55a60658b0e5710872e95286", size = 7083360, upload-time = "2026-04-01T14:45:09.763Z" },
- { url = "https://files.pythonhosted.org/packages/5e/26/d325f9f56c7e039034897e7380e9cc202b1e368bfd04d4cbe6a441f02885/pillow-12.2.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:9aba9a17b623ef750a4d11b742cbafffeb48a869821252b30ee21b5e91392c50", size = 6507628, upload-time = "2026-04-01T14:45:12.378Z" },
- { url = "https://files.pythonhosted.org/packages/5f/f7/769d5632ffb0988f1c5e7660b3e731e30f7f8ec4318e94d0a5d674eb65a4/pillow-12.2.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:deede7c263feb25dba4e82ea23058a235dcc2fe1f6021025dc71f2b618e26104", size = 7209321, upload-time = "2026-04-01T14:45:15.122Z" },
- { url = "https://files.pythonhosted.org/packages/6a/7a/c253e3c645cd47f1aceea6a8bacdba9991bf45bb7dfe927f7c893e89c93c/pillow-12.2.0-cp314-cp314-win32.whl", hash = "sha256:632ff19b2778e43162304d50da0181ce24ac5bb8180122cbe1bf4673428328c7", size = 6479723, upload-time = "2026-04-01T14:45:17.797Z" },
- { url = "https://files.pythonhosted.org/packages/cd/8b/601e6566b957ca50e28725cb6c355c59c2c8609751efbecd980db44e0349/pillow-12.2.0-cp314-cp314-win_amd64.whl", hash = "sha256:4e6c62e9d237e9b65fac06857d511e90d8461a32adcc1b9065ea0c0fa3a28150", size = 7217400, upload-time = "2026-04-01T14:45:20.529Z" },
- { url = "https://files.pythonhosted.org/packages/d6/94/220e46c73065c3e2951bb91c11a1fb636c8c9ad427ac3ce7d7f3359b9b2f/pillow-12.2.0-cp314-cp314-win_arm64.whl", hash = "sha256:b1c1fbd8a5a1af3412a0810d060a78b5136ec0836c8a4ef9aa11807f2a22f4e1", size = 2554835, upload-time = "2026-04-01T14:45:23.162Z" },
- { url = "https://files.pythonhosted.org/packages/b6/ab/1b426a3974cb0e7da5c29ccff4807871d48110933a57207b5a676cccc155/pillow-12.2.0-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:57850958fe9c751670e49b2cecf6294acc99e562531f4bd317fa5ddee2068463", size = 5314225, upload-time = "2026-04-01T14:45:25.637Z" },
- { url = "https://files.pythonhosted.org/packages/19/1e/dce46f371be2438eecfee2a1960ee2a243bbe5e961890146d2dee1ff0f12/pillow-12.2.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:d5d38f1411c0ed9f97bcb49b7bd59b6b7c314e0e27420e34d99d844b9ce3b6f3", size = 4698541, upload-time = "2026-04-01T14:45:28.355Z" },
- { url = "https://files.pythonhosted.org/packages/55/c3/7fbecf70adb3a0c33b77a300dc52e424dc22ad8cdc06557a2e49523b703d/pillow-12.2.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:5c0a9f29ca8e79f09de89293f82fc9b0270bb4af1d58bc98f540cc4aedf03166", size = 6322251, upload-time = "2026-04-01T14:45:30.924Z" },
- { url = "https://files.pythonhosted.org/packages/1c/3c/7fbc17cfb7e4fe0ef1642e0abc17fc6c94c9f7a16be41498e12e2ba60408/pillow-12.2.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1610dd6c61621ae1cf811bef44d77e149ce3f7b95afe66a4512f8c59f25d9ebe", size = 8127807, upload-time = "2026-04-01T14:45:33.908Z" },
- { url = "https://files.pythonhosted.org/packages/ff/c3/a8ae14d6defd2e448493ff512fae903b1e9bd40b72efb6ec55ce0048c8ce/pillow-12.2.0-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0a34329707af4f73cf1782a36cd2289c0368880654a2c11f027bcee9052d35dd", size = 6433935, upload-time = "2026-04-01T14:45:36.623Z" },
- { url = "https://files.pythonhosted.org/packages/6e/32/2880fb3a074847ac159d8f902cb43278a61e85f681661e7419e6596803ed/pillow-12.2.0-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8e9c4f5b3c546fa3458a29ab22646c1c6c787ea8f5ef51300e5a60300736905e", size = 7116720, upload-time = "2026-04-01T14:45:39.258Z" },
- { url = "https://files.pythonhosted.org/packages/46/87/495cc9c30e0129501643f24d320076f4cc54f718341df18cc70ec94c44e1/pillow-12.2.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:fb043ee2f06b41473269765c2feae53fc2e2fbf96e5e22ca94fb5ad677856f06", size = 6540498, upload-time = "2026-04-01T14:45:41.879Z" },
- { url = "https://files.pythonhosted.org/packages/18/53/773f5edca692009d883a72211b60fdaf8871cbef075eaa9d577f0a2f989e/pillow-12.2.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f278f034eb75b4e8a13a54a876cc4a5ab39173d2cdd93a638e1b467fc545ac43", size = 7239413, upload-time = "2026-04-01T14:45:44.705Z" },
- { url = "https://files.pythonhosted.org/packages/c9/e4/4b64a97d71b2a83158134abbb2f5bd3f8a2ea691361282f010998f339ec7/pillow-12.2.0-cp314-cp314t-win32.whl", hash = "sha256:6bb77b2dcb06b20f9f4b4a8454caa581cd4dd0643a08bacf821216a16d9c8354", size = 6482084, upload-time = "2026-04-01T14:45:47.568Z" },
- { url = "https://files.pythonhosted.org/packages/ba/13/306d275efd3a3453f72114b7431c877d10b1154014c1ebbedd067770d629/pillow-12.2.0-cp314-cp314t-win_amd64.whl", hash = "sha256:6562ace0d3fb5f20ed7290f1f929cae41b25ae29528f2af1722966a0a02e2aa1", size = 7225152, upload-time = "2026-04-01T14:45:50.032Z" },
- { url = "https://files.pythonhosted.org/packages/ff/6e/cf826fae916b8658848d7b9f38d88da6396895c676e8086fc0988073aaf8/pillow-12.2.0-cp314-cp314t-win_arm64.whl", hash = "sha256:aa88ccfe4e32d362816319ed727a004423aab09c5cea43c01a4b435643fa34eb", size = 2556579, upload-time = "2026-04-01T14:45:52.529Z" },
+version = "12.3.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/1c/3d/bb7fca845737cf9d7dbde16ed1843984665ff2e0a518f5db43e77ec540b9/pillow-12.3.0.tar.gz", hash = "sha256:3b8182a766685eaa002637e28b4ec8d6b18819a0c71f579bf0dbaa5830297cce", size = 47025035, upload-time = "2026-07-01T11:56:38.965Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/37/bf/fb3ebff8ddcb76aac5a01389251bbbb9519922a9b520d8247c1ca864a25d/pillow-12.3.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:ba09209fbe443b4acccebe845d8a138b89a8f4fbaeedd44953490b5315d5e965", size = 5345969, upload-time = "2026-07-01T11:54:06.397Z" },
+ { url = "https://files.pythonhosted.org/packages/d8/66/9a386a92561f402389a4fc70c18838bf6d35eb5eb5c6850b4b2dc64f5048/pillow-12.3.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:ffd0c5368496f41b0944be820fcb7a838aa6e623d250b01acf2643939c3f99d7", size = 4780323, upload-time = "2026-07-01T11:54:09.351Z" },
+ { url = "https://files.pythonhosted.org/packages/25/27/ac8f99618ffd3dde21db0f4d4b1d2ab00c0880595bfd17df103f7f39fd0c/pillow-12.3.0-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d9c7f76c0673154f044e9d78c8655fb4213f6ca31a836df48b40fe5d187717b9", size = 6266838, upload-time = "2026-07-01T11:54:11.71Z" },
+ { url = "https://files.pythonhosted.org/packages/84/21/a35af28dcc61f37ed850a2d64c65c701321dfbf25085e469d5559360cbbf/pillow-12.3.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:78cb2c6865a35ab8ff8b75fd122f6033b92a62c82801110e48ddd6c936a45d91", size = 6940830, upload-time = "2026-07-01T11:54:13.732Z" },
+ { url = "https://files.pythonhosted.org/packages/eb/51/8b08617af3ad95e33ce6d7dd2c99ed6c8298f7fb131636303956be022e25/pillow-12.3.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:e491916b378fba47242221bb9ead245211b70d504f495d105d17b14a24b4907c", size = 6344383, upload-time = "2026-07-01T11:54:15.756Z" },
+ { url = "https://files.pythonhosted.org/packages/1d/72/cf78ac9780bb93c28328f408973845a309d4d145041665f734572ced1b52/pillow-12.3.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:0dd2064cbc55aaec028ef5fbb60fa47bb6c3e7918e07ff17935284b227a9d2df", size = 7052934, upload-time = "2026-07-01T11:54:17.721Z" },
+ { url = "https://files.pythonhosted.org/packages/20/20/25e0f4dc178a6bc0696793720055519a0de89e7661dae886992decbd2f81/pillow-12.3.0-cp312-cp312-win32.whl", hash = "sha256:dbce0b29841537a2fa4a214c2bbf14de3587c9680caa9b4e217568472490b28f", size = 6472684, upload-time = "2026-07-01T11:54:19.839Z" },
+ { url = "https://files.pythonhosted.org/packages/45/89/da2f7971a317f83d807fdd4065c0af40208e59e692cc43d315a71a0e96d1/pillow-12.3.0-cp312-cp312-win_amd64.whl", hash = "sha256:a2b55dd6b2a4c4b7d87ffa56bdb33fdc5fdb9a462173861a7bc097f17d91cb09", size = 7227137, upload-time = "2026-07-01T11:54:22.025Z" },
+ { url = "https://files.pythonhosted.org/packages/de/47/4845a0a6c0dbf1db8456bd9fc791f13c5ced7ced20606d08a0aacfd25b49/pillow-12.3.0-cp312-cp312-win_arm64.whl", hash = "sha256:331b624368d4f1d069149002f25f44bc61c8919ce8ddb3c45bdad8f6e2d89510", size = 2568267, upload-time = "2026-07-01T11:54:24.051Z" },
+ { url = "https://files.pythonhosted.org/packages/9d/ac/31fb64e1e7efb5a4b50cd3d92049ba89ac6e4d8d3bb6a74e15048ca3353e/pillow-12.3.0-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:21900ce7ba264168cd50defae43cd75d25c833ad4ad6e73ffc5596d12e25ac89", size = 4161684, upload-time = "2026-07-01T11:54:25.934Z" },
+ { url = "https://files.pythonhosted.org/packages/87/b4/9805e23d2b4d77842b468513841fda254ee42f0289d25088340e4ff46e2d/pillow-12.3.0-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:4e8c2a84d977f50b9daed6eeaf3baef67d00d5d74d932288f02cb94518ee3ace", size = 4255487, upload-time = "2026-07-01T11:54:27.935Z" },
+ { url = "https://files.pythonhosted.org/packages/df/39/ecf519435a200c693fe053a6ee4d835b41cf963a4dfc2551c4e637cb2a71/pillow-12.3.0-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:ae26d61dfa7a47befdc7572b521024e8745f3d809bd95ca9505a7bba9ef849ec", size = 3696433, upload-time = "2026-07-01T11:54:29.813Z" },
+ { url = "https://files.pythonhosted.org/packages/42/92/2fc3ffad878ae8dd5469ec1bc8eb83b71f48e13efdf68f02709003982a32/pillow-12.3.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:7a743ff716f746fc19a9557f60dab1600d4613255f8a7aeb3cdde4db7eb15a66", size = 5345889, upload-time = "2026-07-01T11:54:31.97Z" },
+ { url = "https://files.pythonhosted.org/packages/10/76/8803c13605b763d33d156c4678fc77f8443389c0c51c8aef707bb02015f4/pillow-12.3.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:d69141514cc30b774ceea5e3ed3a6635c8d8a96edf664689b890f4089111fb35", size = 4780109, upload-time = "2026-07-01T11:54:34.026Z" },
+ { url = "https://files.pythonhosted.org/packages/1f/01/e18aff37cb0b4aac47ac90f016d347a49aca667ef97f190b06ac2aabc928/pillow-12.3.0-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f7401aebd7f581d7f83a439d87d474999317ee099218e5ad25d125290990ba65", size = 6263736, upload-time = "2026-07-01T11:54:36.131Z" },
+ { url = "https://files.pythonhosted.org/packages/f7/62/de5bdd77d935331f4f802edc11e4d82950f642caad6cb2f949837b8560e2/pillow-12.3.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0847a763afefb695bc912d7c131e7e0632d4edc1d8698f58ddabec8e46b8b6d3", size = 6937129, upload-time = "2026-07-01T11:54:38.216Z" },
+ { url = "https://files.pythonhosted.org/packages/70/4d/105627a13300c5e0df1d174230b32fd1273062c96f7745fd552b945d1e1d/pillow-12.3.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:571b9fcb07b97ef3a492028fb3d2dc0993ca23a06138b0315286566d29ef718a", size = 6339562, upload-time = "2026-07-01T11:54:40.354Z" },
+ { url = "https://files.pythonhosted.org/packages/6b/1d/f13de01a553988ab895ba1c722e06cf3144d4f57656fd5b81b6d881f1179/pillow-12.3.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:756c768d0c9c2955feb7a56c37ea24aea2e369f8d36a88da270b6a9f19e62b5e", size = 7049439, upload-time = "2026-07-01T11:54:42.489Z" },
+ { url = "https://files.pythonhosted.org/packages/c9/f9/066794cca041b969964f779ee5fa66a9498bbf34248ac39c5d7954e4198f/pillow-12.3.0-cp313-cp313-win32.whl", hash = "sha256:a876864214e136f0eb367788dbd7df045f4806801518e2cfe9e13229cfe06d8f", size = 6473287, upload-time = "2026-07-01T11:54:44.9Z" },
+ { url = "https://files.pythonhosted.org/packages/a6/9b/7a58e61d62be561da3a356fe2384d4059a6345fc130e23ef1c36a5b81d24/pillow-12.3.0-cp313-cp313-win_amd64.whl", hash = "sha256:1cca606cd25738df4ed873d5ad46bbdb3d83b5cbca291f6b4ff13a4df6b0bbe8", size = 7239691, upload-time = "2026-07-01T11:54:47.141Z" },
+ { url = "https://files.pythonhosted.org/packages/aa/b0/c4ed4f0ef8f8fa5ee8351537db6650bb8189f7e118842978dd6589065692/pillow-12.3.0-cp313-cp313-win_arm64.whl", hash = "sha256:b629de27fda84b42cde7edef0d85f13b958b47f6e9bbcbba9b673c562a89bd8b", size = 2568185, upload-time = "2026-07-01T11:54:49.137Z" },
+ { url = "https://files.pythonhosted.org/packages/dc/01/001f65b68192f0228cc1dbbc8d2530ab5d58b61037ba0587f946fea607cd/pillow-12.3.0-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:9cf95fe4d0f84c82d282745d9bb08ad9f926efa00be4697e767b814ce40d4330", size = 4161736, upload-time = "2026-07-01T11:54:51.156Z" },
+ { url = "https://files.pythonhosted.org/packages/1a/d2/0219746d0fd16fc8a84498e79452375be3797d3ce4044596ce565164b84f/pillow-12.3.0-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:8728f216dcdb6e6d555cf971cb34076139ad74b31fc2c14da4fafc741c5f6217", size = 4255435, upload-time = "2026-07-01T11:54:53.414Z" },
+ { url = "https://files.pythonhosted.org/packages/c8/02/8d0bc62ef0302318c46ff2a512822d2610e81c7aa46c9b3abe6cbaca5ad0/pillow-12.3.0-cp314-cp314-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:a45650e8ce7fafffd731db8550230db6b0d306d181a90b67d3e6bca2f1990930", size = 3696262, upload-time = "2026-07-01T11:54:55.739Z" },
+ { url = "https://files.pythonhosted.org/packages/85/e2/73c77d218410b14f5f2d565e8a998d5317b7b9c75368d29985139f7a46f0/pillow-12.3.0-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:ba54cfebe86920a559a7c4d6b9050791c20513650a1952ebe3368c7dc70306f8", size = 5350344, upload-time = "2026-07-01T11:54:57.657Z" },
+ { url = "https://files.pythonhosted.org/packages/c7/da/32c752228ae345f489e3a42499d817b6c3996da7e8a3bc7a04fc806b243b/pillow-12.3.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:e158cb00350dc278f3b91551101aa7d12415a66ebf2c91d8d5ac14e56ddd3ad0", size = 4780131, upload-time = "2026-07-01T11:54:59.713Z" },
+ { url = "https://files.pythonhosted.org/packages/b1/9d/8b2c807dbef61a5197c047afe99823787eb66f63daf9fb2432f91d6f0462/pillow-12.3.0-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e9aeb04d6aef139de265b29683e119b638208f88cf73cdd1658aa07221165321", size = 6263757, upload-time = "2026-07-01T11:55:01.778Z" },
+ { url = "https://files.pythonhosted.org/packages/5c/44/c85361f65dbe00eea8576ee467c768d25129989efb76e94f205e9ca9bb46/pillow-12.3.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:251bf95b67017e27b13d82f5b326234ca62d70f9cf4c2b9032de2358a3b12c7b", size = 6936962, upload-time = "2026-07-01T11:55:03.93Z" },
+ { url = "https://files.pythonhosted.org/packages/18/7e/e483414b35800b86b6f08dbbc7803fb5cd52c4d6f897f47d53ea2c7e6f65/pillow-12.3.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:fe3cca2e4e8a592be0f269a1ca4835c25199d9f3ce815c8491048f785b0a0198", size = 6339171, upload-time = "2026-07-01T11:55:05.989Z" },
+ { url = "https://files.pythonhosted.org/packages/f0/f4/68c491844841ede6bed70189546b3ee9731cf9f2cbad396faff5e1ccba45/pillow-12.3.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:23aceaa007d6172b02c277f0cd359c79492bbb14f7072b4ede9fbcaf20648130", size = 7048116, upload-time = "2026-07-01T11:55:08.131Z" },
+ { url = "https://files.pythonhosted.org/packages/a3/34/77f3f793fed8efc7d243f21b33c5a3f0d1c97ee70346d3db855587e155ff/pillow-12.3.0-cp314-cp314-win32.whl", hash = "sha256:af8d94b0db561cf68b88a267c5c44b49e134f525d0dc2cb7ed413a66bc23559a", size = 6467209, upload-time = "2026-07-01T11:55:10.408Z" },
+ { url = "https://files.pythonhosted.org/packages/f1/e0/492879f69d94f91f60fc8cd05ba03650e9520afebb2fb7aa12777d7c7f38/pillow-12.3.0-cp314-cp314-win_amd64.whl", hash = "sha256:fdafc9cce40277e0f7a0feabce0ee50dd2fa1800f3b38015e51296b5e814048d", size = 7237707, upload-time = "2026-07-01T11:55:12.745Z" },
+ { url = "https://files.pythonhosted.org/packages/c9/ac/6b11f2875f1c2ac040d84e1bbf9cf22a88038f901ca1037898b280b38365/pillow-12.3.0-cp314-cp314-win_arm64.whl", hash = "sha256:e91206ee562682b51b98ef4b26a6ef48fd84e15fd4c4bc5ec768eb641d206838", size = 2565995, upload-time = "2026-07-01T11:55:14.736Z" },
+ { url = "https://files.pythonhosted.org/packages/52/69/c2208e56af9bfc1913afb24020297a691eb1d4ef688474c8a04913f65e04/pillow-12.3.0-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:164b31cd1a0490ab6efae01aa5df49da7061be0af1b30e035b6e9a1bfe34ee6e", size = 5352503, upload-time = "2026-07-01T11:55:17.076Z" },
+ { url = "https://files.pythonhosted.org/packages/07/70/e5686d753e898a45d778ff1718dba8516ead6ab6b95d85fc8c4b70650cf2/pillow-12.3.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:5afb51d599ea772b8365ae807ae557f18bccfe46ab261fd1c2a9ed700fc6eb17", size = 4782956, upload-time = "2026-07-01T11:55:19.448Z" },
+ { url = "https://files.pythonhosted.org/packages/d5/37/25c6692f06927ee973ff18c8d9ee98ad0b4d84ee67a09610c2dd1447958e/pillow-12.3.0-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3edce1d53195db527e0191f84b71d02022de0540bf43a16ed734ed7537b07385", size = 6322855, upload-time = "2026-07-01T11:55:21.613Z" },
+ { url = "https://files.pythonhosted.org/packages/cc/91/420637fcb8f1bc11029e403b4538e6694744428d8246118e45719f944556/pillow-12.3.0-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bf16ba1b4d0b6b7c8e534936632270cf70eb00dbe09005bc345b2677b726855c", size = 6989642, upload-time = "2026-07-01T11:55:24.006Z" },
+ { url = "https://files.pythonhosted.org/packages/10/08/b94d7811281ccf0d143a1cf768d1c49e1e54af63e7b708ab2ee3eb87face/pillow-12.3.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:24870b09b224f7ae3c39ed07d10e819d06f8720bc551847b1d623832b5b0e28d", size = 6391281, upload-time = "2026-07-01T11:55:26.252Z" },
+ { url = "https://files.pythonhosted.org/packages/d2/87/24233f785f55474dc02ce3e739c5528a77e3a862e9333d1dd7a25cc31f70/pillow-12.3.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:30f2aa603c41533cc25c05acd0da21636e84a315768feb631c937177db558931", size = 7096716, upload-time = "2026-07-01T11:55:28.318Z" },
+ { url = "https://files.pythonhosted.org/packages/23/26/fcb2f6e37175b04f53570b59937867e2b80ee1685e744023153028fc14f9/pillow-12.3.0-cp314-cp314t-win32.whl", hash = "sha256:4b0a7fe987b14c31ebda6083f74f22b561fd3739bc0ac51e019622e3d72668c7", size = 6474125, upload-time = "2026-07-01T11:55:30.956Z" },
+ { url = "https://files.pythonhosted.org/packages/90/de/3634abee5f1c9e13c56787b7d5517b0ba8d6de51700b95578cf338349c9f/pillow-12.3.0-cp314-cp314t-win_amd64.whl", hash = "sha256:962864dc93511324d51ddbb5b9f8731bf71675b93ca612a07441896f4688fb8c", size = 7242939, upload-time = "2026-07-01T11:55:34.044Z" },
+ { url = "https://files.pythonhosted.org/packages/ce/2a/fd13f8eb24de5714a6eb444a3d67e2842c6c576e159a43793adf23051351/pillow-12.3.0-cp314-cp314t-win_arm64.whl", hash = "sha256:0740a512dc522224c77d9aa5a8d70d8b7d73fb91f2c21125d8d025d3b8990e45", size = 2567506, upload-time = "2026-07-01T11:55:35.988Z" },
+ { url = "https://files.pythonhosted.org/packages/5d/dc/8fdce34ec725a33c81c6ba122b904d6b9024e50ea9ac7bede62fab54506c/pillow-12.3.0-cp315-cp315-ios_13_0_arm64_iphoneos.whl", hash = "sha256:0feb2e9d6ad6c9e3c06effe9d00f3f1e618a6643273576b016f591e9315a7139", size = 4162063, upload-time = "2026-07-01T11:55:37.941Z" },
+ { url = "https://files.pythonhosted.org/packages/76/66/2044b9a63d3b84ff048228dfcb7cd9bf0df983e8470971bf7d4c57b693de/pillow-12.3.0-cp315-cp315-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:9e881fca225083806662a5c43d627d215f258ff43c890f831966c7d7ba9c7402", size = 4255549, upload-time = "2026-07-01T11:55:40.022Z" },
+ { url = "https://files.pythonhosted.org/packages/52/7e/1f67e6f4ece6b582ee4b539decbcc9f848dc245a93ed8cd7338bafef72f1/pillow-12.3.0-cp315-cp315-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:4998562bf62a445225f22e07c896bb04b35b1b1f2eb6d760584c9c51d7a5f78c", size = 3696331, upload-time = "2026-07-01T11:55:41.98Z" },
+ { url = "https://files.pythonhosted.org/packages/12/40/d306fc2c8e4d45d7f175c77edca7063be7b86fe7fe6e68f4353bf71d808c/pillow-12.3.0-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:dc624f6bc473dacdf7ef7eb8678d0d08edf15cd94fad6ae5c7d6cc67a4e4902f", size = 5350370, upload-time = "2026-07-01T11:55:44.028Z" },
+ { url = "https://files.pythonhosted.org/packages/dd/44/668fb1437e8ce420f62d6106eb66e44a5971602a4d794615bdf79315d82d/pillow-12.3.0-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:71d6097b330eea8fd15097780c8e89cb1a8ce7838669f48c5bacd6f663dd4701", size = 4780147, upload-time = "2026-07-01T11:55:46.073Z" },
+ { url = "https://files.pythonhosted.org/packages/0c/08/93fa2e70e30a2d81547e481b6ee2bb9522117221fb1e0ce4b5df70967677/pillow-12.3.0-cp315-cp315-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:28ce87c5ab450a9dd970b52e5aca5fe63ed432d18a2eaddd1979a00a1ba24ace", size = 6273659, upload-time = "2026-07-01T11:55:48.264Z" },
+ { url = "https://files.pythonhosted.org/packages/f8/6d/043e96ff814fc31a33077e4cba86082167db520c93632afdf2042febbb0c/pillow-12.3.0-cp315-cp315-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6b02afb9b97f65fbca5f31db6a2a3ba21aa93030225f150fa3f249717e938fb4", size = 6947439, upload-time = "2026-07-01T11:55:50.503Z" },
+ { url = "https://files.pythonhosted.org/packages/af/92/ba71d2ee2ac0edf3fa33bd9d5ee9ee080da70b1766f3ca3934f9938ddac9/pillow-12.3.0-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:1182d52bc2d5e5d7d0949503aa7e36d12f42205dc287e4883f407b1988820d39", size = 6353577, upload-time = "2026-07-01T11:55:52.697Z" },
+ { url = "https://files.pythonhosted.org/packages/0f/ce/e63064e2122923ff687c8ad792d0d736a7b3920a56a46982e81a7fdd25d6/pillow-12.3.0-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:e795b7eb908249c4e43c7c99fac7c2c75dab0c43566e37db472a355f63693d71", size = 7060394, upload-time = "2026-07-01T11:55:55.149Z" },
+ { url = "https://files.pythonhosted.org/packages/54/76/a09cc3ccc8d773a7283d34c38bec1708f9e3cc932093cbc4c5e71ac4060b/pillow-12.3.0-cp315-cp315-win32.whl", hash = "sha256:57b3d78c95ba9059768b10e28b813002261d3f3dfc55cc48b0c988f625175827", size = 6467375, upload-time = "2026-07-01T11:55:57.769Z" },
+ { url = "https://files.pythonhosted.org/packages/3e/03/1846c49ba3b1d5550392a4bbd06d6fb4578e1cd91a803198b5c90f5f7d53/pillow-12.3.0-cp315-cp315-win_amd64.whl", hash = "sha256:fa4ecea169a355be7a3ade2c783e2ed12f0e40d2c5621cda8b3297faf7fbb9f5", size = 7237048, upload-time = "2026-07-01T11:55:59.975Z" },
+ { url = "https://files.pythonhosted.org/packages/fb/bb/89f35dcc79610423f9f195504d7def7f0d1416a711541b42867e25fe3412/pillow-12.3.0-cp315-cp315-win_arm64.whl", hash = "sha256:877c3f311ff35410f690861c4409e7ccbf0cd2f878e50628a28e5a0bb689e658", size = 2566006, upload-time = "2026-07-01T11:56:02.143Z" },
+ { url = "https://files.pythonhosted.org/packages/30/88/707027ba09942dfa2c28759b5c222d769290a41c6d20ea60ec250801941f/pillow-12.3.0-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:e9871b1ffbfa9656b60aeee92ed5136a5742696006fa322b29ea3d8da0ecc9cf", size = 5352509, upload-time = "2026-07-01T11:56:04.2Z" },
+ { url = "https://files.pythonhosted.org/packages/b0/6d/00352fa25332c2569cd387851f568cc5a4b75a9adbfb37ac4fbce4c02eec/pillow-12.3.0-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:53aa02d20d10c3d814d536aa4e5ac9b84ca0ff5a88377963b085ad6822f93e64", size = 4783167, upload-time = "2026-07-01T11:56:06.631Z" },
+ { url = "https://files.pythonhosted.org/packages/13/4f/9e049dfa21af7c22427275720e2490267ba8138120add5c4c574deb69782/pillow-12.3.0-cp315-cp315t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:446c34dcc4324b084a53b705127dc15717b22c5e140ae0a3c38349d4efec071e", size = 6329237, upload-time = "2026-07-01T11:56:08.868Z" },
+ { url = "https://files.pythonhosted.org/packages/36/16/cf6eeaae8d0fce8dd390a33437cf68c5d5bd73834a2bc6e2f14efda0ab45/pillow-12.3.0-cp315-cp315t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:cf1845d02ad822a369a49f2bb9345b1614744267682e7a03527dc3bf6eea1777", size = 6997047, upload-time = "2026-07-01T11:56:11.379Z" },
+ { url = "https://files.pythonhosted.org/packages/1e/69/dbf769bdd55f48bf5733cac28edc6364ffaa072ec9ba336266e4fe66be55/pillow-12.3.0-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:186941b6aef820ad110fb01fb06eb925374dc3a21b17e37ec9a53b250c6fe2d1", size = 6400440, upload-time = "2026-07-01T11:56:13.908Z" },
+ { url = "https://files.pythonhosted.org/packages/a0/e1/ffc9cfc2eea0d178da8018e18e959301ad9d6bc9f3edb7181e748a474b97/pillow-12.3.0-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:f13c32a3abd6079a66d9526e18dad9b6d280384d49d7c54040cd57b6424041d9", size = 7105895, upload-time = "2026-07-01T11:56:16.575Z" },
+ { url = "https://files.pythonhosted.org/packages/18/f0/a5595c1e8c3ae44b9828cb2f0fa8155e5095ef04d6327b8f61cf44a3df85/pillow-12.3.0-cp315-cp315t-win32.whl", hash = "sha256:1657923d2d45afb66526e5b933e5b3052e6bdea196c90d3abb2424e18c77dae8", size = 6474384, upload-time = "2026-07-01T11:56:18.855Z" },
+ { url = "https://files.pythonhosted.org/packages/e4/04/62bcd9f844984c5938d3b05264a61d797a29d3e0812341a8204af70bbdee/pillow-12.3.0-cp315-cp315t-win_amd64.whl", hash = "sha256:8cd2f7bdda092d99c9fc2fb7391354f306d01443d22785d0cbfafa2e2c8bb418", size = 7243537, upload-time = "2026-07-01T11:56:21.214Z" },
+ { url = "https://files.pythonhosted.org/packages/3d/68/1f3066acedf37673694a7141381d8f811ae97f30d34413d236abe7d489f1/pillow-12.3.0-cp315-cp315t-win_arm64.whl", hash = "sha256:06ff022112bc9cbf83b60f8e028d94ad87b60621706487e65f673de61610ab59", size = 2567491, upload-time = "2026-07-01T11:56:23.506Z" },
]
[[package]]
@@ -3316,95 +3658,18 @@ sdist = { url = "https://files.pythonhosted.org/packages/d6/f5/9be84ec8bcbe77d79
[[package]]
name = "pytest"
-version = "8.4.2"
-source = { registry = "https://pypi.org/simple" }
-resolution-markers = [
- "python_full_version == '3.13.*' and sys_platform == 'win32'",
- "python_full_version < '3.13' and sys_platform == 'win32'",
- "python_full_version == '3.13.*' and sys_platform != 'win32'",
- "python_full_version < '3.13' and sys_platform != 'win32'",
-]
-dependencies = [
- { name = "colorama", marker = "python_full_version < '3.14' and sys_platform == 'win32'" },
- { name = "iniconfig", marker = "python_full_version < '3.14'" },
- { name = "packaging", marker = "python_full_version < '3.14'" },
- { name = "pluggy", marker = "python_full_version < '3.14'" },
- { name = "pygments", marker = "python_full_version < '3.14'" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/a3/5c/00a0e072241553e1a7496d638deababa67c5058571567b92a7eaa258397c/pytest-8.4.2.tar.gz", hash = "sha256:86c0d0b93306b961d58d62a4db4879f27fe25513d4b969df351abdddb3c30e01", size = 1519618, upload-time = "2025-09-04T14:34:22.711Z" }
-wheels = [
- { url = "https://files.pythonhosted.org/packages/a8/a4/20da314d277121d6534b3a980b29035dcd51e6744bd79075a6ce8fa4eb8d/pytest-8.4.2-py3-none-any.whl", hash = "sha256:872f880de3fc3a5bdc88a11b39c9710c3497a547cfa9320bc3c5e62fbf272e79", size = 365750, upload-time = "2025-09-04T14:34:20.226Z" },
-]
-
-[[package]]
-name = "pytest"
-version = "9.0.3"
-source = { registry = "https://pypi.org/simple" }
-resolution-markers = [
- "python_full_version >= '3.14' and sys_platform == 'win32'",
- "python_full_version >= '3.14' and sys_platform != 'win32'",
-]
-dependencies = [
- { name = "colorama", marker = "python_full_version >= '3.14' and sys_platform == 'win32'" },
- { name = "iniconfig", marker = "python_full_version >= '3.14'" },
- { name = "packaging", marker = "python_full_version >= '3.14'" },
- { name = "pluggy", marker = "python_full_version >= '3.14'" },
- { name = "pygments", marker = "python_full_version >= '3.14'" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/7d/0d/549bd94f1a0a402dc8cf64563a117c0f3765662e2e668477624baeec44d5/pytest-9.0.3.tar.gz", hash = "sha256:b86ada508af81d19edeb213c681b1d48246c1a91d304c6c81a427674c17eb91c", size = 1572165, upload-time = "2026-04-07T17:16:18.027Z" }
-wheels = [
- { url = "https://files.pythonhosted.org/packages/d4/24/a372aaf5c9b7208e7112038812994107bc65a84cd00e0354a88c2c77a617/pytest-9.0.3-py3-none-any.whl", hash = "sha256:2c5efc453d45394fdd706ade797c0a81091eccd1d6e4bccfcd476e2b8e0ab5d9", size = 375249, upload-time = "2026-04-07T17:16:16.13Z" },
-]
-
-[[package]]
-name = "pytest-json-report"
-version = "1.5.0"
+version = "9.1.1"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "pytest", version = "8.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.14'" },
- { name = "pytest-metadata", marker = "python_full_version < '3.14'" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/4f/d3/765dae9712fcd68d820338908c1337e077d5fdadccd5cacf95b9b0bea278/pytest-json-report-1.5.0.tar.gz", hash = "sha256:2dde3c647851a19b5f3700729e8310a6e66efb2077d674f27ddea3d34dc615de", size = 21241, upload-time = "2022-03-15T21:03:10.2Z" }
-wheels = [
- { url = "https://files.pythonhosted.org/packages/81/35/d07400c715bf8a88aa0c1ee9c9eb6050ca7fe5b39981f0eea773feeb0681/pytest_json_report-1.5.0-py3-none-any.whl", hash = "sha256:9897b68c910b12a2e48dd849f9a284b2c79a732a8a9cb398452ddd23d3c8c325", size = 13222, upload-time = "2022-03-15T21:03:08.65Z" },
-]
-
-[[package]]
-name = "pytest-metadata"
-version = "3.1.1"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
- { name = "pytest", version = "8.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.14'" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/a6/85/8c969f8bec4e559f8f2b958a15229a35495f5b4ce499f6b865eac54b878d/pytest_metadata-3.1.1.tar.gz", hash = "sha256:d2a29b0355fbc03f168aa96d41ff88b1a3b44a3b02acbe491801c98a048017c8", size = 9952, upload-time = "2024-02-12T19:38:44.887Z" }
-wheels = [
- { url = "https://files.pythonhosted.org/packages/3e/43/7e7b2ec865caa92f67b8f0e9231a798d102724ca4c0e1f414316be1c1ef2/pytest_metadata-3.1.1-py3-none-any.whl", hash = "sha256:c8e0844db684ee1c798cfa38908d20d67d0463ecb6137c72e91f418558dd5f4b", size = 11428, upload-time = "2024-02-12T19:38:42.531Z" },
-]
-
-[[package]]
-name = "pytest-rerunfailures"
-version = "15.1"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
- { name = "packaging", marker = "python_full_version < '3.14'" },
- { name = "pytest", version = "8.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.14'" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/a0/78/e6e358545537a8e82c4dc91e72ec0d6f80546a3786dd27c76b06ca09db77/pytest_rerunfailures-15.1.tar.gz", hash = "sha256:c6040368abd7b8138c5b67288be17d6e5611b7368755ce0465dda0362c8ece80", size = 26981, upload-time = "2025-05-08T06:36:33.483Z" }
-wheels = [
- { url = "https://files.pythonhosted.org/packages/f3/30/11d836ff01c938969efa319b4ebe2374ed79d28043a12bfc908577aab9f3/pytest_rerunfailures-15.1-py3-none-any.whl", hash = "sha256:f674c3594845aba8b23c78e99b1ff8068556cc6a8b277f728071fdc4f4b0b355", size = 13274, upload-time = "2025-05-08T06:36:32.029Z" },
-]
-
-[[package]]
-name = "pytest-xdist"
-version = "3.8.0"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
- { name = "execnet", marker = "python_full_version < '3.14'" },
- { name = "pytest", version = "8.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.14'" },
+ { name = "colorama", marker = "sys_platform == 'win32'" },
+ { name = "iniconfig" },
+ { name = "packaging" },
+ { name = "pluggy" },
+ { name = "pygments" },
]
-sdist = { url = "https://files.pythonhosted.org/packages/78/b4/439b179d1ff526791eb921115fca8e44e596a13efeda518b9d845a619450/pytest_xdist-3.8.0.tar.gz", hash = "sha256:7e578125ec9bc6050861aa93f2d59f1d8d085595d6551c2c90b6f4fad8d3a9f1", size = 88069, upload-time = "2025-07-01T13:30:59.346Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/e4/47/b9efed96c114afcfa3c9d3fe98a76a1d14c74a9e266d397cf6eb64be5e01/pytest-9.1.1.tar.gz", hash = "sha256:1088fbde8f2b49d95a549a195707afa7a76a3ce9bcadc26b6d71f0ffda5fe313", size = 1636369, upload-time = "2026-06-19T10:58:32.857Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/ca/31/d4e37e9e550c2b92a9cbc2e4d0b7420a27224968580b5a447f420847c975/pytest_xdist-3.8.0-py3-none-any.whl", hash = "sha256:202ca578cfeb7370784a8c33d6d05bc6e13b4f25b5053c30a152269fd10f0b88", size = 46396, upload-time = "2025-07-01T13:30:56.632Z" },
+ { url = "https://files.pythonhosted.org/packages/24/25/1de2678b631f5a49215c6c96fff41ba892b0a34df68d6d80292b1b48aa7f/pytest-9.1.1-py3-none-any.whl", hash = "sha256:37a86b45efb9a47a61a36449063e8e18d0cab3161329fc099eb21783169c4f0c", size = 386536, upload-time = "2026-06-19T10:58:31.347Z" },
]
[[package]]
@@ -3451,16 +3716,16 @@ wheels = [
[[package]]
name = "python-multipart"
-version = "0.0.27"
+version = "0.0.32"
source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/69/9b/f23807317a113dc36e74e75eb265a02dd1a4d9082abc3c1064acd22997c4/python_multipart-0.0.27.tar.gz", hash = "sha256:9870a6a8c5a20a5bf4f07c017bd1489006ff8836cff097b6933355ee2b49b602", size = 44043, upload-time = "2026-04-27T10:51:26.649Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/5b/42/55c32bb9b12693c092ad250a0e82edb5b31ddeda6eb772de5f308b3804ad/python_multipart-0.0.32.tar.gz", hash = "sha256:be54b7f3fa167bb83e4fcd936b887b708f4e57fe75911c02aebf53efaf8d938e", size = 46881, upload-time = "2026-06-04T16:18:58.647Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/99/78/4126abcbdbd3c559d43e0db7f7b9173fc6befe45d39a2856cc0b8ec2a5a6/python_multipart-0.0.27-py3-none-any.whl", hash = "sha256:6fccfad17a27334bd0193681b369f476eda3409f17381a2d65aa7df3f7275645", size = 29254, upload-time = "2026-04-27T10:51:24.997Z" },
+ { url = "https://files.pythonhosted.org/packages/e1/04/e8135ebd1ad02c56ec633277529b2602ff99ff634be76cdba5744cf554fd/python_multipart-0.0.32-py3-none-any.whl", hash = "sha256:ff6d3f776f16878c894e52e107296ffc890e913c611b1a4ec6c44e2821fe2e23", size = 30042, upload-time = "2026-06-04T16:18:57.319Z" },
]
[[package]]
name = "pytorch-tokenizers"
-version = "1.2.0"
+version = "1.4.1"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "sentencepiece", marker = "python_full_version < '3.14'" },
@@ -3468,14 +3733,18 @@ dependencies = [
{ name = "tokenizers", marker = "python_full_version < '3.14'" },
]
wheels = [
- { url = "https://files.pythonhosted.org/packages/1e/cc/9d77d52ab1422cacfa0d4ebb144dc85edcf45c230cb89627721f1d357229/pytorch_tokenizers-1.2.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:a2893a30ad3db3e0d690dc7669b0b6da68dde2fd4d81dcad608e9e66ea8848ec", size = 1111123, upload-time = "2026-04-01T00:19:00.985Z" },
- { url = "https://files.pythonhosted.org/packages/1c/30/5279d5f6415596ae6b768867dd082e4a2af22b6bc9445b0fc9a205c77e1c/pytorch_tokenizers-1.2.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:7619cf3c9f39d0ddbdfb1e2f34cd2a1e4971d6ea6d5306e3a9efb1cfc73b8c7c", size = 1441086, upload-time = "2026-04-01T00:19:02.635Z" },
- { url = "https://files.pythonhosted.org/packages/99/cc/dd634211e2be38067b77d4dadd59c726f8cfc19bb1ce25fd048e4c8fa0d6/pytorch_tokenizers-1.2.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:251f959843fd6a7af204b58f71d11b73c39e6eeb0491720170301da2ad8c251b", size = 1556129, upload-time = "2026-04-01T00:19:04.004Z" },
- { url = "https://files.pythonhosted.org/packages/68/a1/6fafe2ae2d09b05dea79b7696047f287bc229e366e476e9ea67dfa17ff9e/pytorch_tokenizers-1.2.0-cp312-cp312-win_amd64.whl", hash = "sha256:97a82db21665023b25afb869c1bf07c31047c442df15d33e036c4712d07cb109", size = 848594, upload-time = "2026-04-01T00:19:05.274Z" },
- { url = "https://files.pythonhosted.org/packages/92/8e/600bd39381e620ff8c004a1e1215cec0fc2e9f9dbdbe77275481168c069c/pytorch_tokenizers-1.2.0-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:9d1f4f2eb0f2ba5ce97c106c3af9ce7b5349028bf99229fae906cb836b249525", size = 1106909, upload-time = "2026-04-01T00:19:06.488Z" },
- { url = "https://files.pythonhosted.org/packages/46/3f/7d39b2ad4d1cfcdf2d37a9ac3caaaa9f5a4a87c75b179eee1d0c17f010c1/pytorch_tokenizers-1.2.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:a3d559e92145ce9d2d48b512beccd5cba8c9da920d23f4f373e376e571fe5647", size = 1442674, upload-time = "2026-04-01T00:19:08.366Z" },
- { url = "https://files.pythonhosted.org/packages/18/c7/4f6abd29eb68022458da29845caa12b525300afd885236ceb9a31358fb03/pytorch_tokenizers-1.2.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:27e954013155ebe38bb25d1f5e5279315e145cdb31b479e9eaf74b6a793fdaa4", size = 1556922, upload-time = "2026-04-01T00:19:09.88Z" },
- { url = "https://files.pythonhosted.org/packages/2b/d7/5a2a30613c73a613b03ebd1dd8ecf695e7a9490ede195ffb2862fd9471d6/pytorch_tokenizers-1.2.0-cp313-cp313-win_amd64.whl", hash = "sha256:f4304cbc66eb46c98ba21de46ae030ea08a65391ed18cde4ca3f5b48955313f0", size = 848532, upload-time = "2026-04-01T00:19:11.043Z" },
+ { url = "https://files.pythonhosted.org/packages/26/91/c3d2564c80587fa501c29652c317e8a4eec6f7b13c7b53f1de6bc0abddc0/pytorch_tokenizers-1.4.1-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:e3ad0a9c474396267e0f70273d74fec402cb88541afb6617f98c6042e71faedb", size = 1112739, upload-time = "2026-08-05T19:37:40.356Z" },
+ { url = "https://files.pythonhosted.org/packages/d6/cf/69d6d3c6fad65f0f1c63a8f0f0db85009a9db4e25fa50d573ead4e527cbb/pytorch_tokenizers-1.4.1-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:84ee954611ec869f875dd23618c66c15b718c2a4eed639fefb72e7806091ff76", size = 1451907, upload-time = "2026-08-05T19:37:41.622Z" },
+ { url = "https://files.pythonhosted.org/packages/ad/b3/4d38b4e5cd951df08511e0a11637baad56dec6ca82bca7b594c8d0c85e59/pytorch_tokenizers-1.4.1-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:3feb33e04dcaa2cea6dcbe99aba22da2ad81d1ed305bc0af52acab5b3f8ff094", size = 1568882, upload-time = "2026-08-05T19:37:43.108Z" },
+ { url = "https://files.pythonhosted.org/packages/44/58/8ab61ce7f666b4b06016a1517836c24f2dd06b7e6d7d5255200aeacd2656/pytorch_tokenizers-1.4.1-cp312-cp312-win_amd64.whl", hash = "sha256:40affc76c3f98ef5c26910b68b4dcd27b5d9209e2eccfd18e4a9cdc0f0048f47", size = 857910, upload-time = "2026-08-05T19:37:44.502Z" },
+ { url = "https://files.pythonhosted.org/packages/a4/c8/b811e0990964e82d0616707bf53bc6fce4692580aef42351ca1ae3041b8a/pytorch_tokenizers-1.4.1-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:2561c5187e028d906fc24c98288d183e05e8b2b27d565e5462cc2f5611801653", size = 1112791, upload-time = "2026-08-05T19:37:46.039Z" },
+ { url = "https://files.pythonhosted.org/packages/f7/51/3494837ac254666e3c4dfca14d2bf8a08bd8a3d454f066ba7701462f5928/pytorch_tokenizers-1.4.1-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:a647af855ec93485a6f97074a78a906b0c4d227896f2a8498dd674d5e8be14fd", size = 1451825, upload-time = "2026-08-05T19:37:47.32Z" },
+ { url = "https://files.pythonhosted.org/packages/ed/3f/ae7e9d6ed87b3089aa88e8b00c5b474ab5f29c28875a3f44156a5caaf2aa/pytorch_tokenizers-1.4.1-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:944d149ece8fe278d7aee6520e82bed14f6d14c890af1f45cc3399fa4c31ad44", size = 1568950, upload-time = "2026-08-05T19:37:48.752Z" },
+ { url = "https://files.pythonhosted.org/packages/c6/52/7e1bdab83e7c2a16846f2733c3f1b7ca386cb108e2f84c695ab26d9b1abe/pytorch_tokenizers-1.4.1-cp313-cp313-win_amd64.whl", hash = "sha256:75b16fe8bfb69686f4b5c5435deae6eeec859bd6aaf85b66bda743263db796bc", size = 857879, upload-time = "2026-08-05T19:37:50.593Z" },
+ { url = "https://files.pythonhosted.org/packages/18/23/162945a49f2c9147c2db24bd9a6c21fc550571e0f9071ad3f24eb4625c44/pytorch_tokenizers-1.4.1-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:97ee2d2ae991a0ca7e26b4bc6215293bb6498d5090efffb8300064d0e4f874cb", size = 1113125, upload-time = "2026-08-05T19:37:51.791Z" },
+ { url = "https://files.pythonhosted.org/packages/4a/e9/a999083759060840880d441ee0c8ec7859fdf0ae3c87d35ce825c976f6d5/pytorch_tokenizers-1.4.1-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:60cbe48d2b81b250f2c2fc85cfa89f5f4247e451ee968a68d958855447dbc855", size = 1451956, upload-time = "2026-08-05T19:37:53.323Z" },
+ { url = "https://files.pythonhosted.org/packages/58/70/61ca1ccc90b7efcd998fb0fa5f1d446a21c0d3978101c60f514e3b65802b/pytorch_tokenizers-1.4.1-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:b2d99c7db1d0033cf76818cbc786cdcc2f8582677da74cf60bdef504e5b9dad2", size = 1570276, upload-time = "2026-08-05T19:37:54.73Z" },
+ { url = "https://files.pythonhosted.org/packages/81/1d/558212519a4704baad64535ad14d568b37c723664736c3127715b58b0955/pytorch_tokenizers-1.4.1-cp314-cp314-win_amd64.whl", hash = "sha256:ebc91f9c010d1eea5291028e2e0352e63e2ea147e2a1b05e308134f247d016dd", size = 858339, upload-time = "2026-08-05T19:37:56.221Z" },
]
[[package]]
@@ -3712,24 +3981,26 @@ wheels = [
[[package]]
name = "safetensors"
-version = "0.7.0"
+version = "0.8.0"
source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/29/9c/6e74567782559a63bd040a236edca26fd71bc7ba88de2ef35d75df3bca5e/safetensors-0.7.0.tar.gz", hash = "sha256:07663963b67e8bd9f0b8ad15bb9163606cd27cc5a1b96235a50d8369803b96b0", size = 200878, upload-time = "2025-11-19T15:18:43.199Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/45/06/f955dbbb1859e3bd23c8ac6141af5106e7ad5fedec4a3a6e3d60f94b7001/safetensors-0.8.0.tar.gz", hash = "sha256:fabaf3e0f18a6618d9b36560682562157f77c2b71fcffc7b432be2baed9d753d", size = 325846, upload-time = "2026-06-09T07:52:25.563Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/fa/47/aef6c06649039accf914afef490268e1067ed82be62bcfa5b7e886ad15e8/safetensors-0.7.0-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:c82f4d474cf725255d9e6acf17252991c3c8aac038d6ef363a4bf8be2f6db517", size = 467781, upload-time = "2025-11-19T15:18:35.84Z" },
- { url = "https://files.pythonhosted.org/packages/e8/00/374c0c068e30cd31f1e1b46b4b5738168ec79e7689ca82ee93ddfea05109/safetensors-0.7.0-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:94fd4858284736bb67a897a41608b5b0c2496c9bdb3bf2af1fa3409127f20d57", size = 447058, upload-time = "2025-11-19T15:18:34.416Z" },
- { url = "https://files.pythonhosted.org/packages/f1/06/578ffed52c2296f93d7fd2d844cabfa92be51a587c38c8afbb8ae449ca89/safetensors-0.7.0-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e07d91d0c92a31200f25351f4acb2bc6aff7f48094e13ebb1d0fb995b54b6542", size = 491748, upload-time = "2025-11-19T15:18:09.79Z" },
- { url = "https://files.pythonhosted.org/packages/ae/33/1debbbb70e4791dde185edb9413d1fe01619255abb64b300157d7f15dddd/safetensors-0.7.0-cp38-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:8469155f4cb518bafb4acf4865e8bb9d6804110d2d9bdcaa78564b9fd841e104", size = 503881, upload-time = "2025-11-19T15:18:16.145Z" },
- { url = "https://files.pythonhosted.org/packages/8e/1c/40c2ca924d60792c3be509833df711b553c60effbd91da6f5284a83f7122/safetensors-0.7.0-cp38-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:54bef08bf00a2bff599982f6b08e8770e09cc012d7bba00783fc7ea38f1fb37d", size = 623463, upload-time = "2025-11-19T15:18:21.11Z" },
- { url = "https://files.pythonhosted.org/packages/9b/3a/13784a9364bd43b0d61eef4bea2845039bc2030458b16594a1bd787ae26e/safetensors-0.7.0-cp38-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:42cb091236206bb2016d245c377ed383aa7f78691748f3bb6ee1bfa51ae2ce6a", size = 532855, upload-time = "2025-11-19T15:18:25.719Z" },
- { url = "https://files.pythonhosted.org/packages/a0/60/429e9b1cb3fc651937727befe258ea24122d9663e4d5709a48c9cbfceecb/safetensors-0.7.0-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:dac7252938f0696ddea46f5e855dd3138444e82236e3be475f54929f0c510d48", size = 507152, upload-time = "2025-11-19T15:18:33.023Z" },
- { url = "https://files.pythonhosted.org/packages/3c/a8/4b45e4e059270d17af60359713ffd83f97900d45a6afa73aaa0d737d48b6/safetensors-0.7.0-cp38-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:1d060c70284127fa805085d8f10fbd0962792aed71879d00864acda69dbab981", size = 541856, upload-time = "2025-11-19T15:18:31.075Z" },
- { url = "https://files.pythonhosted.org/packages/06/87/d26d8407c44175d8ae164a95b5a62707fcc445f3c0c56108e37d98070a3d/safetensors-0.7.0-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:cdab83a366799fa730f90a4ebb563e494f28e9e92c4819e556152ad55e43591b", size = 674060, upload-time = "2025-11-19T15:18:37.211Z" },
- { url = "https://files.pythonhosted.org/packages/11/f5/57644a2ff08dc6325816ba7217e5095f17269dada2554b658442c66aed51/safetensors-0.7.0-cp38-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:672132907fcad9f2aedcb705b2d7b3b93354a2aec1b2f706c4db852abe338f85", size = 771715, upload-time = "2025-11-19T15:18:38.689Z" },
- { url = "https://files.pythonhosted.org/packages/86/31/17883e13a814bd278ae6e266b13282a01049b0c81341da7fd0e3e71a80a3/safetensors-0.7.0-cp38-abi3-musllinux_1_2_i686.whl", hash = "sha256:5d72abdb8a4d56d4020713724ba81dac065fedb7f3667151c4a637f1d3fb26c0", size = 714377, upload-time = "2025-11-19T15:18:40.162Z" },
- { url = "https://files.pythonhosted.org/packages/4a/d8/0c8a7dc9b41dcac53c4cbf9df2b9c83e0e0097203de8b37a712b345c0be5/safetensors-0.7.0-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:b0f6d66c1c538d5a94a73aa9ddca8ccc4227e6c9ff555322ea40bdd142391dd4", size = 677368, upload-time = "2025-11-19T15:18:41.627Z" },
- { url = "https://files.pythonhosted.org/packages/05/e5/cb4b713c8a93469e3c5be7c3f8d77d307e65fe89673e731f5c2bfd0a9237/safetensors-0.7.0-cp38-abi3-win32.whl", hash = "sha256:c74af94bf3ac15ac4d0f2a7c7b4663a15f8c2ab15ed0fc7531ca61d0835eccba", size = 326423, upload-time = "2025-11-19T15:18:45.74Z" },
- { url = "https://files.pythonhosted.org/packages/5d/e6/ec8471c8072382cb91233ba7267fd931219753bb43814cbc71757bfd4dab/safetensors-0.7.0-cp38-abi3-win_amd64.whl", hash = "sha256:d1239932053f56f3456f32eb9625590cc7582e905021f94636202a864d470755", size = 341380, upload-time = "2025-11-19T15:18:44.427Z" },
+ { url = "https://files.pythonhosted.org/packages/39/a0/f718cda65b05407d228f97602cf60dca269c979867aa5beb25410de26cd3/safetensors-0.8.0-cp310-abi3-macosx_10_12_x86_64.whl", hash = "sha256:c554f85858e05226d3c2828e32395e677434685d6d94594a41643361c5e837f0", size = 473568, upload-time = "2026-06-09T07:52:18.829Z" },
+ { url = "https://files.pythonhosted.org/packages/f5/b1/fa7c600e7dceae12e9606c7578cbc9ff1e1ed55844883ee5c92205e86226/safetensors-0.8.0-cp310-abi3-macosx_11_0_arm64.whl", hash = "sha256:c80201d22cbf405b80647a60ada77bba06c8fba2da2743ba1e89cdcc39a81f25", size = 484562, upload-time = "2026-06-09T07:52:17.518Z" },
+ { url = "https://files.pythonhosted.org/packages/09/7d/65a7de0af421317bb36a067241e4235fff194eed60b961ed6d3f59a3fc60/safetensors-0.8.0-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7a46e5ff292c356d6991e60942ba7f79817682d3a2cef0702136448cb9c4d235", size = 502844, upload-time = "2026-06-09T07:52:07.624Z" },
+ { url = "https://files.pythonhosted.org/packages/91/4f/3175c9d75634e0e0dda0082794193521035edd7c70a6f212bf33ca06ddf4/safetensors-0.8.0-cp310-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:4124502b78f03534117c848f87a39b8f31e577b15eff423bf8bfb95f2a8c30d0", size = 511823, upload-time = "2026-06-09T07:52:09.565Z" },
+ { url = "https://files.pythonhosted.org/packages/20/87/846c289e7aa2299eff406335717cf43ce8777194ece8aad75772e0411615/safetensors-0.8.0-cp310-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7bc0a787ba8a35be368ee3574edfa2b1ad389eebd0a72e482ae275490e3f6c98", size = 633461, upload-time = "2026-06-09T07:52:11.128Z" },
+ { url = "https://files.pythonhosted.org/packages/76/22/8d64d9df2c45d5ded401df889d0ad90882804ca172d79ec4f0df8f727fe0/safetensors-0.8.0-cp310-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:040070828e36dc8e122178bbbd5830ff9e97920affb84cbe0f46442497bed358", size = 545148, upload-time = "2026-06-09T07:52:13.603Z" },
+ { url = "https://files.pythonhosted.org/packages/28/50/f203ff3a3ddfe19308efc83c5a3a29ed02bf786732ec35e68bf9162f3365/safetensors-0.8.0-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fd6f3f93c9a0a7cc2788ee63fb763353d4bd2e89b0751bc78fcf7dda00bea774", size = 516040, upload-time = "2026-06-09T07:52:16.29Z" },
+ { url = "https://files.pythonhosted.org/packages/46/fb/cdaed17ceb2948784fd9c36b6fd3e951b608547cea81a48e8ee6f8cfdfcb/safetensors-0.8.0-cp310-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:fcdd41ec4628fee5799f807c73c353629130fbd942aa23d83c623dd6c9d52d78", size = 513832, upload-time = "2026-06-09T07:52:12.37Z" },
+ { url = "https://files.pythonhosted.org/packages/0d/49/1e15de264dcc3b77943d2d0c56a95809956883b1c2d6d585c792523f180b/safetensors-0.8.0-cp310-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:8e9f537aa183a38ace122d27303dcd986b26bd2a7591f9181d7f0c396f4677ca", size = 559930, upload-time = "2026-06-09T07:52:14.743Z" },
+ { url = "https://files.pythonhosted.org/packages/2a/43/bf38443278eab4b1be1fce2931e2b012ad9cb7df52ada751d0aab8f7659a/safetensors-0.8.0-cp310-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:87eec7ffed2b809f05a398a8becb7d013f19f7837cd15d9748580d6cf30dbaf4", size = 678670, upload-time = "2026-06-09T07:52:20.032Z" },
+ { url = "https://files.pythonhosted.org/packages/72/e3/68cd3fa5b48488e84add63e04cb12f3bc28ae4638c06d4508c6e88823d0e/safetensors-0.8.0-cp310-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:4a95ae2b05d7726d751da4ebf626a2ca782b706e101bd894c95bc2450b1cffcc", size = 786679, upload-time = "2026-06-09T07:52:21.322Z" },
+ { url = "https://files.pythonhosted.org/packages/29/4b/1c19c509d56e01f4fbb3d0a2e597450f6cc04d1d56cf52defb0a62dfd715/safetensors-0.8.0-cp310-abi3-musllinux_1_2_i686.whl", hash = "sha256:3ae091f16662658bdc019a4ff6cb4c085bb7d725eb5978b183ffd265863b6d2d", size = 765683, upload-time = "2026-06-09T07:52:22.594Z" },
+ { url = "https://files.pythonhosted.org/packages/27/43/41c1621732edd934d868a00d1b891584c892a7b62a9aab82ea5a0a5623ee/safetensors-0.8.0-cp310-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:8e080062fcde23be189565e1c3305d16751a218ecf9412c8601e64204eb6f846", size = 722361, upload-time = "2026-06-09T07:52:23.924Z" },
+ { url = "https://files.pythonhosted.org/packages/8e/3f/73ccf82579412b4a71c4ca673f10b5f1f888d7cf5af7fe24f27d30307be4/safetensors-0.8.0-cp310-abi3-win32.whl", hash = "sha256:2ddf52eac562eda224f99acfa7889d02968c1fd59a5b011ae7d8137c37e9c02d", size = 342401, upload-time = "2026-06-09T07:52:28.895Z" },
+ { url = "https://files.pythonhosted.org/packages/1b/6d/3fba214c1e5e0f69991677ec3bc17023f0421776975e1de0c682dca475e2/safetensors-0.8.0-cp310-abi3-win_amd64.whl", hash = "sha256:096ec1a98435df7beb08853bb5aa9081a84f23d0adc67ed1a0a10550f608373f", size = 355540, upload-time = "2026-06-09T07:52:27.832Z" },
+ { url = "https://files.pythonhosted.org/packages/8d/fc/7eedc3510d97878876e32774eebbeb61c43f148a96e915c84229a3e967aa/safetensors-0.8.0-cp310-abi3-win_arm64.whl", hash = "sha256:f7838e5135a406ad3e02efdcb8cf2e5397d368b0154537c4fec682dbc544d452", size = 340500, upload-time = "2026-06-09T07:52:26.745Z" },
]
[[package]]
@@ -3926,11 +4197,11 @@ wheels = [
[[package]]
name = "setuptools"
-version = "81.0.0"
+version = "84.0.0"
source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/0d/1c/73e719955c59b8e424d015ab450f51c0af856ae46ea2da83eba51cc88de1/setuptools-81.0.0.tar.gz", hash = "sha256:487b53915f52501f0a79ccfd0c02c165ffe06631443a886740b91af4b7a5845a", size = 1198299, upload-time = "2026-02-06T21:10:39.601Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/6d/44/f5da03a8ef95d369145c5bb53050e7877c9f3d312e128605fd9504829143/setuptools-84.0.0.tar.gz", hash = "sha256:f4695c21257f0d9b537ec2692c941d02ee143b7cc1276941349a546573b2ef73", size = 1168449, upload-time = "2026-08-08T18:27:58.365Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/e1/e3/c164c88b2e5ce7b24d667b9bd83589cf4f3520d97cad01534cd3c4f55fdb/setuptools-81.0.0-py3-none-any.whl", hash = "sha256:fdd925d5c5d9f62e4b74b30d6dd7828ce236fd6ed998a08d81de62ce5a6310d6", size = 1062021, upload-time = "2026-02-06T21:10:37.175Z" },
+ { url = "https://files.pythonhosted.org/packages/95/9c/c510029fc6ef33a6275cd2c5d3cecd6613dfd6aa401d57c54f1c18852ccf/setuptools-84.0.0-py3-none-any.whl", hash = "sha256:51a52592b3b99e102b609654876bd65f19f999935166d1352678931132b0c670", size = 818216, upload-time = "2026-08-08T18:27:56.719Z" },
]
[[package]]
@@ -3993,11 +4264,11 @@ wheels = [
[[package]]
name = "soupsieve"
-version = "2.8.3"
+version = "2.9.2"
source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/7b/ae/2d9c981590ed9999a0d91755b47fc74f74de286b0f5cee14c9269041e6c4/soupsieve-2.8.3.tar.gz", hash = "sha256:3267f1eeea4251fb42728b6dfb746edc9acaffc4a45b27e19450b676586e8349", size = 118627, upload-time = "2026-01-20T04:27:02.457Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/69/99/a6ca3beb3ccacb41fb3321d8a60e5566f9e6467601ef8eba6a17e1b89778/soupsieve-2.9.2.tar.gz", hash = "sha256:4a55d8cf158a9c2e587fa4922f1bbb91d68ac829e2d6f25403a85747c71daf74", size = 122445, upload-time = "2026-08-07T00:57:24.801Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/46/2c/1462b1d0a634697ae9e55b3cecdcb64788e8b7d63f54d923fcd0bb140aed/soupsieve-2.8.3-py3-none-any.whl", hash = "sha256:ed64f2ba4eebeab06cc4962affce381647455978ffc1e36bb79a545b91f45a95", size = 37016, upload-time = "2026-01-20T04:27:01.012Z" },
+ { url = "https://files.pythonhosted.org/packages/eb/dc/ad025c1ee131eba60c69f4dd5779b18fcf1e6b21a343e2162a84d5d133c7/soupsieve-2.9.2-py3-none-any.whl", hash = "sha256:8089a26fd974ca7a1f30276d3d8492ab266ab15af581642dfe8aa162e0c1c823", size = 37370, upload-time = "2026-08-07T00:57:23.524Z" },
]
[[package]]
@@ -4080,15 +4351,15 @@ wheels = [
[[package]]
name = "starlette"
-version = "1.0.0"
+version = "1.6.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "anyio" },
{ name = "typing-extensions", marker = "python_full_version < '3.13'" },
]
-sdist = { url = "https://files.pythonhosted.org/packages/81/69/17425771797c36cded50b7fe44e850315d039f28b15901ab44839e70b593/starlette-1.0.0.tar.gz", hash = "sha256:6a4beaf1f81bb472fd19ea9b918b50dc3a77a6f2e190a12954b25e6ed5eea149", size = 2655289, upload-time = "2026-03-22T18:29:46.779Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/b5/b4/205b0d5241d934e8add0c38aa924c4f9fb7330834ff11e5444db964ec3f9/starlette-1.6.0.tar.gz", hash = "sha256:d4e3ac5e546444960c710297a3c9fc3f7ebae1b7e963f3d36173b49da535be9b", size = 2716969, upload-time = "2026-08-08T18:27:57.512Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/0b/c9/584bc9651441b4ba60cc4d557d8a547b5aff901af35bda3a4ee30c819b82/starlette-1.0.0-py3-none-any.whl", hash = "sha256:d3ec55e0bb321692d275455ddfd3df75fff145d009685eb40dc91fc66b03d38b", size = 72651, upload-time = "2026-03-22T18:29:45.111Z" },
+ { url = "https://files.pythonhosted.org/packages/c8/cb/6a6a47d5b464bd08695d254f3da6e7986cc70c9fa5d778eda57538edfe56/starlette-1.6.0-py3-none-any.whl", hash = "sha256:a86dd39d14bb45f85a3d18525215a9ef0cfd1f192ac793220e72598c90335f0c", size = 75969, upload-time = "2026-08-08T18:27:56.196Z" },
]
[[package]]
@@ -4112,6 +4383,52 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/99/55/db07de81b5c630da5cbf5c7df646580ca26dfaefa593667fc6f2fe016d2e/tabulate-0.10.0-py3-none-any.whl", hash = "sha256:f0b0622e567335c8fabaaa659f1b33bcb6ddfe2e496071b743aa113f8774f2d3", size = 39814, upload-time = "2026-03-04T18:55:31.284Z" },
]
+[[package]]
+name = "tensorflow"
+version = "2.21.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "absl-py" },
+ { name = "astunparse" },
+ { name = "flatbuffers" },
+ { name = "gast" },
+ { name = "google-pasta" },
+ { name = "grpcio" },
+ { name = "h5py" },
+ { name = "keras" },
+ { name = "libclang" },
+ { name = "ml-dtypes" },
+ { name = "numpy" },
+ { name = "opt-einsum" },
+ { name = "packaging" },
+ { name = "protobuf" },
+ { name = "requests" },
+ { name = "setuptools" },
+ { name = "six" },
+ { name = "termcolor" },
+ { name = "typing-extensions" },
+ { name = "wrapt" },
+]
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/39/59/27adba20bbd1088b00fc0e9232aa21493b4800af2299eb6005017a6053f4/tensorflow-2.21.0-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:56ecd7d47429acbe1df2694d50b75bf9fc3995ac92cb367cd9af6c4780ead712", size = 223488205, upload-time = "2026-03-06T17:24:03.48Z" },
+ { url = "https://files.pythonhosted.org/packages/20/a5/f139fbbd4e81a2e556170718698acf518a71a84927fa1dc1f5935b6ab3d2/tensorflow-2.21.0-cp312-cp312-manylinux_2_27_aarch64.whl", hash = "sha256:b07d15737406533898e36c7ef7944b1be18e9028dc521b9229952c537bbc5552", size = 281946471, upload-time = "2026-03-06T17:24:11.896Z" },
+ { url = "https://files.pythonhosted.org/packages/ce/d7/6e71b4ded8ce99cd21add95611ca14af6e9ad4f2baeabdeb79a4a6b3cb1f/tensorflow-2.21.0-cp312-cp312-manylinux_2_27_x86_64.whl", hash = "sha256:b3b95643c4e70eb925839938fb35cbe142f317ec84af6844ee61513713bb13c0", size = 572611111, upload-time = "2026-03-06T17:24:26.209Z" },
+ { url = "https://files.pythonhosted.org/packages/7d/0d/4ee4bc074597b41c9c00dc97b4418ef1eb8736fe9186ffdb3961efdfb730/tensorflow-2.21.0-cp312-cp312-win_amd64.whl", hash = "sha256:27ba0682572b1e50a0db1ee74cfb787eafcfdb1b751bbbf401fed62fe32a92e0", size = 350945509, upload-time = "2026-03-06T17:24:41.65Z" },
+ { url = "https://files.pythonhosted.org/packages/40/09/268b45a61be2bce136dabf3a3cd7099c8a984ae398198f71920b4c60c502/tensorflow-2.21.0-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:a145ed46c58192b7c3f9916d070caf4f6afc6dc7e5511f83dd97677f4c4947f4", size = 223766900, upload-time = "2026-03-06T17:24:51.349Z" },
+ { url = "https://files.pythonhosted.org/packages/72/72/343b86b4c9bfe28e81f749439f908c1e26aeac73d9f12b8dcdb996eb8ecb/tensorflow-2.21.0-cp313-cp313-manylinux_2_27_aarch64.whl", hash = "sha256:a10abdfb8b1189210c251021a3f153b7ccc52a8e6521351f3dc3331e8ba593e0", size = 282213003, upload-time = "2026-03-06T17:24:59.859Z" },
+ { url = "https://files.pythonhosted.org/packages/86/6c/10d075ffc09754c7f10e749ba3c9d46dd809fb007990c7f788128044180c/tensorflow-2.21.0-cp313-cp313-manylinux_2_27_x86_64.whl", hash = "sha256:e9d8da8dcab9650efb45f032ba70af2f016f907e6e0c6bda29dd101bba945406", size = 572881074, upload-time = "2026-03-06T17:25:14.453Z" },
+ { url = "https://files.pythonhosted.org/packages/86/91/dedad8403e7b0036d99be4878987693b7b7f62097eb8537fa6ce62ea131c/tensorflow-2.21.0-cp313-cp313-win_amd64.whl", hash = "sha256:76cccbe0a95d9392dee1ae501ae0656b6c73c1cac29a7f8f32d570e0670863f7", size = 351205371, upload-time = "2026-03-06T17:25:33.144Z" },
+]
+
+[[package]]
+name = "termcolor"
+version = "3.3.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/46/79/cf31d7a93a8fdc6aa0fbb665be84426a8c5a557d9240b6239e9e11e35fc5/termcolor-3.3.0.tar.gz", hash = "sha256:348871ca648ec6a9a983a13ab626c0acce02f515b9e1983332b17af7979521c5", size = 14434, upload-time = "2025-12-29T12:55:21.882Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/33/d1/8bb87d21e9aeb323cc03034f5eaf2c8f69841e40e4853c2627edf8111ed3/termcolor-3.3.0-py3-none-any.whl", hash = "sha256:cf642efadaf0a8ebbbf4bc7a31cec2f9b5f21a9f726f4ccbb08192c9c26f43a5", size = 7734, upload-time = "2025-12-29T12:55:20.718Z" },
+]
+
[[package]]
name = "threadpoolctl"
version = "3.6.0"
@@ -4205,10 +4522,10 @@ wheels = [
[[package]]
name = "torch"
-version = "2.11.0"
+version = "2.13.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "cuda-bindings", marker = "sys_platform == 'linux'" },
+ { name = "cuda-bindings", marker = "python_full_version < '3.15' and sys_platform == 'linux'" },
{ name = "cuda-toolkit", extra = ["cublas", "cudart", "cufft", "cufile", "cupti", "curand", "cusolver", "cusparse", "nvjitlink", "nvrtc", "nvtx"], marker = "sys_platform == 'linux'" },
{ name = "filelock" },
{ name = "fsspec" },
@@ -4220,39 +4537,35 @@ dependencies = [
{ name = "nvidia-nvshmem-cu13", marker = "sys_platform == 'linux'" },
{ name = "setuptools" },
{ name = "sympy" },
- { name = "triton", marker = "sys_platform == 'linux'" },
+ { name = "triton", version = "3.7.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.15' and sys_platform == 'linux'" },
{ name = "typing-extensions" },
]
wheels = [
- { url = "https://files.pythonhosted.org/packages/6f/8b/69e3008d78e5cee2b30183340cc425081b78afc5eff3d080daab0adda9aa/torch-2.11.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:4b5866312ee6e52ea625cd211dcb97d6a2cdc1131a5f15cc0d87eec948f6dd34", size = 80606338, upload-time = "2026-03-23T18:11:34.781Z" },
- { url = "https://files.pythonhosted.org/packages/13/16/42e5915ebe4868caa6bac83a8ed59db57f12e9a61b7d749d584776ed53d5/torch-2.11.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:f99924682ef0aa6a4ab3b1b76f40dc6e273fca09f367d15a524266db100a723f", size = 419731115, upload-time = "2026-03-23T18:11:06.944Z" },
- { url = "https://files.pythonhosted.org/packages/1a/c9/82638ef24d7877510f83baf821f5619a61b45568ce21c0a87a91576510aa/torch-2.11.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:0f68f4ac6d95d12e896c3b7a912b5871619542ec54d3649cf48cc1edd4dd2756", size = 530712279, upload-time = "2026-03-23T18:10:31.481Z" },
- { url = "https://files.pythonhosted.org/packages/1c/ff/6756f1c7ee302f6d202120e0f4f05b432b839908f9071157302cedfc5232/torch-2.11.0-cp312-cp312-win_amd64.whl", hash = "sha256:fbf39280699d1b869f55eac536deceaa1b60bd6788ba74f399cc67e60a5fab10", size = 114556047, upload-time = "2026-03-23T18:10:55.931Z" },
- { url = "https://files.pythonhosted.org/packages/87/89/5ea6722763acee56b045435fb84258db7375c48165ec8be7880ab2b281c5/torch-2.11.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:1e6debd97ccd3205bbb37eb806a9d8219e1139d15419982c09e23ef7d4369d18", size = 80606801, upload-time = "2026-03-23T18:10:18.649Z" },
- { url = "https://files.pythonhosted.org/packages/32/d1/8ed2173589cbfe744ed54e5a73efc107c0085ba5777ee93a5f4c1ab90553/torch-2.11.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:63a68fa59de8f87acc7e85a5478bb2dddbb3392b7593ec3e78827c793c4b73fd", size = 419732382, upload-time = "2026-03-23T18:08:30.835Z" },
- { url = "https://files.pythonhosted.org/packages/3d/e1/b73f7c575a4b8f87a5928f50a1e35416b5e27295d8be9397d5293e7e8d4c/torch-2.11.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:cc89b9b173d9adfab59fd227f0ab5e5516d9a52b658ae41d64e59d2e55a418db", size = 530711509, upload-time = "2026-03-23T18:08:47.213Z" },
- { url = "https://files.pythonhosted.org/packages/66/82/3e3fcdd388fbe54e29fd3f991f36846ff4ac90b0d0181e9c8f7236565f82/torch-2.11.0-cp313-cp313-win_amd64.whl", hash = "sha256:4dda3b3f52d121063a731ddb835f010dc137b920d7fec2778e52f60d8e4bf0cd", size = 114555842, upload-time = "2026-03-23T18:09:52.111Z" },
- { url = "https://files.pythonhosted.org/packages/db/38/8ac78069621b8c2b4979c2f96dc8409ef5e9c4189f6aac629189a78677ca/torch-2.11.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:8b394322f49af4362d4f80e424bcaca7efcd049619af03a4cf4501520bdf0fb4", size = 80959574, upload-time = "2026-03-23T18:10:14.214Z" },
- { url = "https://files.pythonhosted.org/packages/6d/6c/56bfb37073e7136e6dd86bfc6af7339946dd684e0ecf2155ac0eee687ae1/torch-2.11.0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:2658f34ce7e2dabf4ec73b45e2ca68aedad7a5be87ea756ad656eaf32bf1e1ea", size = 419732324, upload-time = "2026-03-23T18:09:36.604Z" },
- { url = "https://files.pythonhosted.org/packages/07/f4/1b666b6d61d3394cca306ea543ed03a64aad0a201b6cd159f1d41010aeb1/torch-2.11.0-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:98bb213c3084cfe176302949bdc360074b18a9da7ab59ef2edc9d9f742504778", size = 530596026, upload-time = "2026-03-23T18:09:20.842Z" },
- { url = "https://files.pythonhosted.org/packages/48/6b/30d1459fa7e4b67e9e3fe1685ca1d8bb4ce7c62ef436c3a615963c6c866c/torch-2.11.0-cp313-cp313t-win_amd64.whl", hash = "sha256:a97b94bbf62992949b4730c6cd2cc9aee7b335921ee8dc207d930f2ed09ae2db", size = 114793702, upload-time = "2026-03-23T18:09:47.304Z" },
- { url = "https://files.pythonhosted.org/packages/26/0d/8603382f61abd0db35841148ddc1ffd607bf3100b11c6e1dab6d2fc44e72/torch-2.11.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:01018087326984a33b64e04c8cb5c2795f9120e0d775ada1f6638840227b04d7", size = 80573442, upload-time = "2026-03-23T18:09:10.117Z" },
- { url = "https://files.pythonhosted.org/packages/c7/86/7cd7c66cb9cec6be330fff36db5bd0eef386d80c031b581ec81be1d4b26c/torch-2.11.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:2bb3cc54bd0dea126b0060bb1ec9de0f9c7f7342d93d436646516b0330cd5be7", size = 419749385, upload-time = "2026-03-23T18:07:33.77Z" },
- { url = "https://files.pythonhosted.org/packages/47/e8/b98ca2d39b2e0e4730c0ee52537e488e7008025bc77ca89552ff91021f7c/torch-2.11.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:4dc8b3809469b6c30b411bb8c4cad3828efd26236153d9beb6a3ec500f211a60", size = 530716756, upload-time = "2026-03-23T18:07:50.02Z" },
- { url = "https://files.pythonhosted.org/packages/78/88/d4a4cda8362f8a30d1ed428564878c3cafb0d87971fbd3947d4c84552095/torch-2.11.0-cp314-cp314-win_amd64.whl", hash = "sha256:2b4e811728bd0cc58fb2b0948fe939a1ee2bf1422f6025be2fca4c7bd9d79718", size = 114552300, upload-time = "2026-03-23T18:09:05.617Z" },
- { url = "https://files.pythonhosted.org/packages/bf/46/4419098ed6d801750f26567b478fc185c3432e11e2cad712bc6b4c2ab0d0/torch-2.11.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:8245477871c3700d4370352ffec94b103cfcb737229445cf9946cddb7b2ca7cd", size = 80959460, upload-time = "2026-03-23T18:09:00.818Z" },
- { url = "https://files.pythonhosted.org/packages/fd/66/54a56a4a6ceaffb567231994a9745821d3af922a854ed33b0b3a278e0a99/torch-2.11.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:ab9a8482f475f9ba20e12db84b0e55e2f58784bdca43a854a6ccd3fd4b9f75e6", size = 419735835, upload-time = "2026-03-23T18:07:18.974Z" },
- { url = "https://files.pythonhosted.org/packages/b1/e7/0b6665f533aa9e337662dc190425abc0af1fe3234088f4454c52393ded61/torch-2.11.0-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:563ed3d25542d7e7bbc5b235ccfacfeb97fb470c7fee257eae599adb8005c8a2", size = 530613405, upload-time = "2026-03-23T18:08:07.014Z" },
- { url = "https://files.pythonhosted.org/packages/cf/bf/c8d12a2c86dbfd7f40fb2f56fbf5a505ccf2d9ce131eb559dfc7c51e1a04/torch-2.11.0-cp314-cp314t-win_amd64.whl", hash = "sha256:b2a43985ff5ef6ddd923bbcf99943e5f58059805787c5c9a2622bf05ca2965b0", size = 114792991, upload-time = "2026-03-23T18:08:19.216Z" },
+ { url = "https://files.pythonhosted.org/packages/c4/3a/ed0f4d4d1dcde03bced7aac9a28e800abcdc0cbd06b6775044c9fbd877b7/torch-2.13.0-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:2fe228aba290d14b9f31b049be550dbd469c3fd3013d7a19705b30454da97027", size = 111213045, upload-time = "2026-07-08T16:05:22.997Z" },
+ { url = "https://files.pythonhosted.org/packages/df/a9/f6a2a4d763ff1df02e9a64c477029db614295bc9367f4131223791ccc243/torch-2.13.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:572df8be8ffb4599c88cbd6a0726f1f854f4da65d2e3c09f0e2c2283333cd6d4", size = 427210998, upload-time = "2026-07-08T16:04:37.708Z" },
+ { url = "https://files.pythonhosted.org/packages/f3/82/fea946351658e6534db52d2cc12bc53087cbf87f9440c5f180f367c1950b/torch-2.13.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:796633c4cdf0fe2cdced72d8f88f22e73dbcfce83132763162f6d4bff13b820b", size = 526605292, upload-time = "2026-07-08T16:04:22.81Z" },
+ { url = "https://files.pythonhosted.org/packages/21/d6/e8f3c6f7e01f626f77259de9860d2a78bc84c40539e28e79b7e98b0bb659/torch-2.13.0-cp312-cp312-win_amd64.whl", hash = "sha256:024c6cc0c1b085f2f91f20a3dc27b0471d021c31ce84b81be3afdc39f791fd9d", size = 122057313, upload-time = "2026-07-08T16:03:53.43Z" },
+ { url = "https://files.pythonhosted.org/packages/0d/fa/c1c10b7aff4a9a3e8956d4f0a5f468fa6db7abc3208805719076772b4833/torch-2.13.0-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:33449899ce5496c1b84b4853179d94fd102028ae1407314d9fb956bb79e70d09", size = 111213743, upload-time = "2026-07-08T16:03:28.579Z" },
+ { url = "https://files.pythonhosted.org/packages/11/18/9ecb37b56293a0be8d80f810bf672a72fe7e02f8b475d5ef1b9bf8a0d748/torch-2.13.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:1e09d6a722504957c694faceca843acde562786df1144ebcc5a74075ec7f6005", size = 427213008, upload-time = "2026-07-08T16:03:44.106Z" },
+ { url = "https://files.pythonhosted.org/packages/d4/5a/7c50ba1b7b713d71d34669c6d13dab0a11531a3eceb0307a5162dbfec0f7/torch-2.13.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:a3a9a21312872af8a26950b2c15680335a386a1f56ed03e780653d78b9607e9e", size = 526602329, upload-time = "2026-07-08T16:03:12.649Z" },
+ { url = "https://files.pythonhosted.org/packages/91/3d/e7adcc6aaf36961cd18f56cf8ad0f3058c3a5c84ccf391762176c94581b8/torch-2.13.0-cp313-cp313-win_amd64.whl", hash = "sha256:49b58f1e2c52440abb6f17c28f0335fe6c6d01ad1a7f55b0183b81e4b34d64e6", size = 122057920, upload-time = "2026-07-08T16:03:01.808Z" },
+ { url = "https://files.pythonhosted.org/packages/36/76/6dcc7f0c07052102dd36f83cbc5800842a909c8c3fbf1a7f8a5844954de9/torch-2.13.0-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:d849b390e07d8d333ce8ecaf91b273c656c598379a19c9acf1318a883f6b391c", size = 111227066, upload-time = "2026-07-08T16:03:33.6Z" },
+ { url = "https://files.pythonhosted.org/packages/e9/09/2c10e8cd0e00fa5d23c052df6ce467eaa7182399f5e0f824f1e4ff42ccae/torch-2.13.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:a3893dc2da0a972a8ca5d698c85a9f967559ac5f8ee1797b77408aa8734d073c", size = 427226309, upload-time = "2026-07-08T16:02:53.127Z" },
+ { url = "https://files.pythonhosted.org/packages/76/c6/22c2102bbef14ca6a6cb4c20e42f088e49c5f812be4e160ae57502e325f9/torch-2.13.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:49f1ea385c754e54919408a9bb3b5a72b0b755bbe2c916c1d6f70afbec4908a2", size = 526614507, upload-time = "2026-07-08T16:02:16.441Z" },
+ { url = "https://files.pythonhosted.org/packages/2b/0c/7d1deb6bce5bc3e6042caf39100ac768eba3b9a098e1dddd16f75bd6489b/torch-2.13.0-cp314-cp314-win_amd64.whl", hash = "sha256:4f8573e3ce9ebcd53fe922f01077a6085ccdfbe5f12fd215883a9d87d7a744fd", size = 122051871, upload-time = "2026-07-08T16:03:23.521Z" },
+ { url = "https://files.pythonhosted.org/packages/f4/ce/aa8b7f9949d32e0f2f624f342bc3b48112c1b8a130288465938bc83bcbf9/torch-2.13.0-cp314-cp314t-macosx_14_0_arm64.whl", hash = "sha256:c28def70706c2f9ecc752574766e8ae4da9b810ab6676b611166761a78a9f1e1", size = 111537025, upload-time = "2026-07-08T16:02:44.28Z" },
+ { url = "https://files.pythonhosted.org/packages/69/d1/491e3a0389430946145888b0203f2b6a759ce2a61481b96a85c2da4f2ced/torch-2.13.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:31061ff56ed8fbf26c749806905aeb749ebeb819810fd5d52508aa5afd90dddc", size = 427219769, upload-time = "2026-07-08T16:02:31.18Z" },
+ { url = "https://files.pythonhosted.org/packages/9a/1d/38006e045bf0a1fc28ef01e757c554e59e59a8770c284bc4f47b14e60441/torch-2.13.0-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:cc26eead4cf51d0b544e31e364dcf000846549c273bd148936fe9d24d29acb92", size = 526571320, upload-time = "2026-07-08T16:01:59.348Z" },
+ { url = "https://files.pythonhosted.org/packages/56/94/655c91992a882bd5071aa0b5d22a07dbb130d801e872be97c0b627a7c693/torch-2.13.0-cp314-cp314t-win_amd64.whl", hash = "sha256:a7de8a313090dc5c7d7ba4bfe5c3be222528f9a4dba1acc83bddb1157360c4b8", size = 122306773, upload-time = "2026-07-08T16:02:39.832Z" },
]
[[package]]
name = "torchao"
-version = "0.17.0"
+version = "0.18.0"
source = { registry = "https://pypi.org/simple" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/32/fe/a4036a8e80fa800c92dbcbf75f541cd4c106248b6b579db6dab1800f616a/torchao-0.17.0-cp310-abi3-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:87a418ce0ec064a821ceab83c921b501acef0ce9a6ccd1be358fcd16c3ae8c58", size = 3206172, upload-time = "2026-03-30T22:25:52.974Z" },
- { url = "https://files.pythonhosted.org/packages/c9/37/ef37ca885265e5f79a168616767dd416a3cea1cc3b28bb6b503ce4a5b652/torchao-0.17.0-py3-none-any.whl", hash = "sha256:02eba449036715b9ae784fbaa1a6f97994bb7b0421ce92d1d5d1c08e5bd6d349", size = 1200680, upload-time = "2026-03-30T22:25:54.457Z" },
+ { url = "https://files.pythonhosted.org/packages/19/55/ed9ad98f0f09d5a1124d09830043d13a39e63539f9590d2bdb6d71cbc4a4/torchao-0.18.0-cp310-abi3-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6540b148e40ba81cbd4de86392225a076a1591146e9cebb099b3b234ba9feebe", size = 3372585, upload-time = "2026-08-03T19:43:10.993Z" },
+ { url = "https://files.pythonhosted.org/packages/c4/4d/485477bb8f05bd501016059c6d8abd742f830cb1b24ab7704e086c7cc35a/torchao-0.18.0-py3-none-any.whl", hash = "sha256:5c2b4485341bf28b7fed2c4fc95b9f298e209f41685350f067de85527a05585e", size = 1369798, upload-time = "2026-08-03T19:43:12.649Z" },
]
[[package]]
@@ -4284,7 +4597,7 @@ wheels = [
[[package]]
name = "torchvision"
-version = "0.26.0"
+version = "0.28.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "numpy" },
@@ -4292,26 +4605,22 @@ dependencies = [
{ name = "torch" },
]
wheels = [
- { url = "https://files.pythonhosted.org/packages/ae/e7/56b47cc3b132aea90ccce22bcb8975dec688b002150012acc842846039d0/torchvision-0.26.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:c409e1c3fdebec7a3834465086dbda8bf7680eff79abf7fd2f10c6b59520a7a4", size = 1863502, upload-time = "2026-03-23T18:12:57.326Z" },
- { url = "https://files.pythonhosted.org/packages/f4/ec/5c31c92c08b65662fe9604a4067ae8232582805949f11ddc042cebe818ed/torchvision-0.26.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:406557718e62fdf10f5706e88d8a5ec000f872da913bf629aab9297622585547", size = 7767944, upload-time = "2026-03-23T18:12:42.805Z" },
- { url = "https://files.pythonhosted.org/packages/f5/d8/cb6ccda1a1f35a6597645818641701207b3e8e13553e75fce5d86bac74b2/torchvision-0.26.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:d61a5abb6b42a0c0c311996c2ac4b83a94418a97182c83b055a2a4ae985e05aa", size = 7522205, upload-time = "2026-03-23T18:12:54.654Z" },
- { url = "https://files.pythonhosted.org/packages/1c/a9/c272623a0f735c35f0f6cd6dc74784d4f970e800cf063bb76687895a2ab9/torchvision-0.26.0-cp312-cp312-win_amd64.whl", hash = "sha256:7993c01648e7c61d191b018e84d38fe0825c8fcb2720cd0f37caf7ba14404aa1", size = 4255155, upload-time = "2026-03-23T18:12:32.652Z" },
- { url = "https://files.pythonhosted.org/packages/da/80/0762f77f53605d10c9477be39bb47722cc8e383bbbc2531471ce0e396c07/torchvision-0.26.0-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:5d63dd43162691258b1b3529b9041bac7d54caa37eae0925f997108268cbf7c4", size = 1860809, upload-time = "2026-03-23T18:12:47.629Z" },
- { url = "https://files.pythonhosted.org/packages/e6/81/0b3e58d1478c660a5af4268713486b2df7203f35abd9195fea87348a5178/torchvision-0.26.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:a39c7a26538c41fda453f9a9692b5ff9b35a5437db1d94f3027f6f509c160eac", size = 7727494, upload-time = "2026-03-23T18:12:46.062Z" },
- { url = "https://files.pythonhosted.org/packages/b6/dc/d9ab5d29115aa05e12e30f1397a3eeae1d88a511241dc3bce48dc4342675/torchvision-0.26.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:b7e6213620bbf97742e5f79832f9e9d769e6cf0f744c5b53dad80b76db633691", size = 7521747, upload-time = "2026-03-23T18:12:36.815Z" },
- { url = "https://files.pythonhosted.org/packages/a9/1b/f1bc86a918c5f6feab1eeff11982e2060f4704332e96185463d27855bdf5/torchvision-0.26.0-cp313-cp313-win_amd64.whl", hash = "sha256:4280c35ec8cba1fcc8294fb87e136924708726864c379e4c54494797d86bc474", size = 4319880, upload-time = "2026-03-23T18:12:38.168Z" },
- { url = "https://files.pythonhosted.org/packages/66/28/b4ad0a723ed95b003454caffcc41894b34bd8379df340848cae2c33871de/torchvision-0.26.0-cp313-cp313t-macosx_12_0_arm64.whl", hash = "sha256:358fc4726d0c08615b6d83b3149854f11efb2a564ed1acb6fce882e151412d23", size = 1951973, upload-time = "2026-03-23T18:12:48.781Z" },
- { url = "https://files.pythonhosted.org/packages/71/e2/7a89096e6cf2f3336353b5338ba925e0addf9d8601920340e6bdf47e8eb3/torchvision-0.26.0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:3daf9cc149cf3cdcbd4df9c59dae69ffca86c6823250442c3bbfd63fc2e26c61", size = 7728679, upload-time = "2026-03-23T18:12:26.196Z" },
- { url = "https://files.pythonhosted.org/packages/69/1d/4e1eebc17d18ce080a11dcf3df3f8f717f0efdfa00983f06e8ba79259f61/torchvision-0.26.0-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:82c3965eca27e86a316e31e4c3e5a16d353e0bcbe0ef8efa2e66502c54493c4b", size = 7609138, upload-time = "2026-03-23T18:12:35.327Z" },
- { url = "https://files.pythonhosted.org/packages/f3/a4/f1155e943ae5b32400d7000adc81c79bb0392b16ceb33bcf13e02e48cced/torchvision-0.26.0-cp313-cp313t-win_amd64.whl", hash = "sha256:ebc043cc5a4f0bf22e7680806dbba37ffb19e70f6953bbb44ed1a90aeb5c9bea", size = 4248202, upload-time = "2026-03-23T18:12:41.423Z" },
- { url = "https://files.pythonhosted.org/packages/7f/c8/9bffa9c7f7bdf95b2a0a2dc535c290b9f1cc580c3fb3033ab1246ffffdeb/torchvision-0.26.0-cp314-cp314-macosx_12_0_arm64.whl", hash = "sha256:eb61804eb9dbe88c5a2a6c4da8dec1d80d2d0a6f18c999c524e32266cb1ebcd3", size = 1860813, upload-time = "2026-03-23T18:12:39.636Z" },
- { url = "https://files.pythonhosted.org/packages/7b/ac/48f28ffd227991f2e14f4392dde7e8dc14352bb9428c1ef4a4bbf5f7ed85/torchvision-0.26.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:9a904f2131cbfadab4df828088a9f66291ad33f49ff853872aed1f86848ef776", size = 7727777, upload-time = "2026-03-23T18:12:22.549Z" },
- { url = "https://files.pythonhosted.org/packages/a4/21/a2266f7f1b0e58e624ff15fd6f01041f59182c49551ece0db9a183071329/torchvision-0.26.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:0f3e572efe62ad645017ea847e0b5e4f2f638d4e39f05bc011d1eb9ac68d4806", size = 7522174, upload-time = "2026-03-23T18:12:29.565Z" },
- { url = "https://files.pythonhosted.org/packages/fc/ba/1666f90bc0bdd77aaa11dcc42bb9f621a9c3668819c32430452e3d404730/torchvision-0.26.0-cp314-cp314-win_amd64.whl", hash = "sha256:114bec0c0e98aa4ba446f63e2fe7a2cbca37b39ac933987ee4804f65de121800", size = 4348469, upload-time = "2026-03-23T18:12:24.44Z" },
- { url = "https://files.pythonhosted.org/packages/45/8f/1f0402ac55c2ae15651ff831957d083fe70b2d12282e72612a30ba601512/torchvision-0.26.0-cp314-cp314t-macosx_12_0_arm64.whl", hash = "sha256:b7d3e295624a28b3b1769228ce1345d94cf4d390dd31136766f76f2d20f718da", size = 1860826, upload-time = "2026-03-23T18:12:34.1Z" },
- { url = "https://files.pythonhosted.org/packages/d2/6a/18a582fe3c5ee26f49b5c9fb21ad8016b4d1c06d10178894a58653946fda/torchvision-0.26.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:7058c5878262937e876f20c25867b33724586aa4499e2853b2d52b99a5e51953", size = 7729089, upload-time = "2026-03-23T18:12:31.394Z" },
- { url = "https://files.pythonhosted.org/packages/c5/9b/f7e119b59499edc00c55c03adc9ec3bd96144d9b81c46852c431f9c64a9a/torchvision-0.26.0-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:8008474855623c6ba52876589dc52df0aa66e518c25eca841445348e5f79844c", size = 7522704, upload-time = "2026-03-23T18:12:20.301Z" },
- { url = "https://files.pythonhosted.org/packages/d0/6a/09f3844c10643f6c0de5d95abc863420cfaf194c88c7dffd0ac523e2015f/torchvision-0.26.0-cp314-cp314t-win_amd64.whl", hash = "sha256:e9d0e022c19a78552fb055d0414d47fecb4a649309b9968573daea160ba6869c", size = 4454275, upload-time = "2026-03-23T18:12:27.487Z" },
+ { url = "https://files.pythonhosted.org/packages/15/49/c1cab1ecbb3ff1a380a3f99283db1dee61b8afe354f6352c643b65937130/torchvision-0.28.0-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:e9f54c30cd52e3ef7fd034cc69b7bb7e0964e1c8f8743e018ab92e95b40f9eee", size = 1856020, upload-time = "2026-07-08T16:07:52.182Z" },
+ { url = "https://files.pythonhosted.org/packages/f0/4c/95233776e2def960e5abb7a07931230a545f43717a56a1e1140162033598/torchvision-0.28.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:5cf78ebc401ce64ae19b8c55de866bb836797d559a4de9c25ccbe74cfa642d3a", size = 7842127, upload-time = "2026-07-08T16:07:53.446Z" },
+ { url = "https://files.pythonhosted.org/packages/93/e4/e9b2495d0d57b9f60d63c57d0a910410a81b4b073bf70917bef815291119/torchvision-0.28.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:028a3d481b37d785605620d7cdad897064c5a55bae2aa1f2658766333e291940", size = 7675040, upload-time = "2026-07-08T16:07:58.017Z" },
+ { url = "https://files.pythonhosted.org/packages/7c/9c/55ed9cb6dfe3ee9c837df5cd0e758372e5829aa38b8dd71343aa632cc4e2/torchvision-0.28.0-cp312-cp312-win_amd64.whl", hash = "sha256:87dc16b2df427c1318ad335f1e2be2b3b15b2cf20f7934c83b0505a48425ee5d", size = 4085785, upload-time = "2026-07-08T16:07:50.928Z" },
+ { url = "https://files.pythonhosted.org/packages/20/55/08a726c14c67b37c8aca04b077766909f1c7ed23f76116884fe63b9bd033/torchvision-0.28.0-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:d483b4aa3f5237569053f749cd1a2b5bb548ca456e40461a5dd087f21149d123", size = 1856021, upload-time = "2026-07-08T16:07:45.386Z" },
+ { url = "https://files.pythonhosted.org/packages/db/8f/40beacd53809194f5259e590d1afaeaa8ad57da15f77c646e6560bcc4616/torchvision-0.28.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:bb6dd6918460ed89cc7644adcc2402991474d6933cf1ce92b390641cb233fddf", size = 7797014, upload-time = "2026-07-08T16:07:43.04Z" },
+ { url = "https://files.pythonhosted.org/packages/32/db/062cdb5a84380a60439775311fff34d89229760d2a50680393dc18699956/torchvision-0.28.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:ad7b3a439265cc3739a4ab5b4c998c0e38ea99c0ee7ca4dea35c5d0b099ec237", size = 7674669, upload-time = "2026-07-08T16:07:38.91Z" },
+ { url = "https://files.pythonhosted.org/packages/f3/a6/b4081e2d04e1541abf82785ac9e5178a494c19330391f551356c8c18b7b3/torchvision-0.28.0-cp313-cp313-win_amd64.whl", hash = "sha256:7e9dd6f60d6e15f8dc27d4f877fdb6002fc70d70272412135f1c2ff9cfa08d3b", size = 4157380, upload-time = "2026-07-08T16:07:40.22Z" },
+ { url = "https://files.pythonhosted.org/packages/c5/b9/da40eca5bbe9596c12ae9899ab7abaf887f5e20f29d08b924b4633714821/torchvision-0.28.0-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:3bd9dba55224a9db4a2d77f6feaa5651770d8c8e86d3d0ddb0fa6bec54c8712b", size = 1856014, upload-time = "2026-07-08T16:07:44.282Z" },
+ { url = "https://files.pythonhosted.org/packages/06/d6/313aafd3df4eaf5f330211bd4e75b7598bddbfee4f55580d3b58536e1b20/torchvision-0.28.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:89f90e29b0966352811b12589f3a3c61943bf2bb9487b9d7bbec10efb1096bb5", size = 7796873, upload-time = "2026-07-08T16:07:30.907Z" },
+ { url = "https://files.pythonhosted.org/packages/b3/41/31f8e959ab8f942600b6357f8999c21d779d5fd3304b0fd204ff4b518239/torchvision-0.28.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:36beb0782976906069ca03d4c9aacaf4b6b838b06ed6c20960ea9c51cce7acdd", size = 7674634, upload-time = "2026-07-08T16:07:29.657Z" },
+ { url = "https://files.pythonhosted.org/packages/15/15/4c5115253fd470672cdac0a1cf139e06b4f3e29d041238a2b255937f63be/torchvision-0.28.0-cp314-cp314-win_amd64.whl", hash = "sha256:3557cc7b539f46dabcda2b6f2b14017ccbeef024de466d4fc5835fc3f287f769", size = 4184005, upload-time = "2026-07-08T16:07:35.805Z" },
+ { url = "https://files.pythonhosted.org/packages/6a/80/822a6163da716f8a78141cf6678d74e26a572285d4ea866ef8aa657bb307/torchvision-0.28.0-cp314-cp314t-macosx_14_0_arm64.whl", hash = "sha256:09ce8f56e81f19b9c378ae7bb109f83f6659fd8bc3cd14241a48e4af46e9ed49", size = 1856011, upload-time = "2026-07-08T16:07:33.404Z" },
+ { url = "https://files.pythonhosted.org/packages/7f/d1/cd3f9463b39a790ec8c0c2f6e6c8061edb1562114d04fcdfa786ed889345/torchvision-0.28.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:62c7d110f86a039245b587e4fae60278c649f3bd42ff79cfbc1178eca4e72542", size = 7796742, upload-time = "2026-07-08T16:07:28.339Z" },
+ { url = "https://files.pythonhosted.org/packages/d9/82/3e0a7ad18e99831e2d7f4713d3be717b7159ff5a920862dd5c23c454aa71/torchvision-0.28.0-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:904cf89af220f8c6b2ed0296bb5065b474ce43b77558e48b2bf9de8b0ba17204", size = 7675526, upload-time = "2026-07-08T16:07:34.572Z" },
+ { url = "https://files.pythonhosted.org/packages/18/d4/23aea03b28297bc66a4461f55ae4296368a9d85fa9a454bafcb2a5348bd7/torchvision-0.28.0-cp314-cp314t-win_amd64.whl", hash = "sha256:46f581979c010ad6da6bd85ee602aa707e1ff44312670223b7a0ee517ad06d47", size = 4291452, upload-time = "2026-07-08T16:07:32.236Z" },
]
[[package]]
@@ -4350,17 +4659,30 @@ wheels = [
name = "triton"
version = "3.6.0"
source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+ "python_full_version >= '3.14' and sys_platform == 'win32'",
+ "python_full_version == '3.13.*' and sys_platform == 'win32'",
+ "python_full_version < '3.13' and sys_platform == 'win32'",
+]
+
+[[package]]
+name = "triton"
+version = "3.7.1"
+source = { registry = "https://pypi.org/simple" }
+resolution-markers = [
+ "python_full_version >= '3.14' and sys_platform != 'win32'",
+ "python_full_version == '3.13.*' and sys_platform != 'win32'",
+ "python_full_version < '3.13' and sys_platform != 'win32'",
+]
wheels = [
- { url = "https://files.pythonhosted.org/packages/17/5d/08201db32823bdf77a0e2b9039540080b2e5c23a20706ddba942924ebcd6/triton-3.6.0-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:374f52c11a711fd062b4bfbb201fd9ac0a5febd28a96fb41b4a0f51dde3157f4", size = 176128243, upload-time = "2026-01-20T16:16:07.857Z" },
- { url = "https://files.pythonhosted.org/packages/ab/a8/cdf8b3e4c98132f965f88c2313a4b493266832ad47fb52f23d14d4f86bb5/triton-3.6.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:74caf5e34b66d9f3a429af689c1c7128daba1d8208df60e81106b115c00d6fca", size = 188266850, upload-time = "2026-01-20T16:00:43.041Z" },
- { url = "https://files.pythonhosted.org/packages/3c/12/34d71b350e89a204c2c7777a9bba0dcf2f19a5bfdd70b57c4dbc5ffd7154/triton-3.6.0-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:448e02fe6dc898e9e5aa89cf0ee5c371e99df5aa5e8ad976a80b93334f3494fd", size = 176133521, upload-time = "2026-01-20T16:16:13.321Z" },
- { url = "https://files.pythonhosted.org/packages/f9/0b/37d991d8c130ce81a8728ae3c25b6e60935838e9be1b58791f5997b24a54/triton-3.6.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:10c7f76c6e72d2ef08df639e3d0d30729112f47a56b0c81672edc05ee5116ac9", size = 188289450, upload-time = "2026-01-20T16:00:49.136Z" },
- { url = "https://files.pythonhosted.org/packages/ce/4e/41b0c8033b503fd3cfcd12392cdd256945026a91ff02452bef40ec34bee7/triton-3.6.0-cp313-cp313t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1722e172d34e32abc3eb7711d0025bb69d7959ebea84e3b7f7a341cd7ed694d6", size = 176276087, upload-time = "2026-01-20T16:16:18.989Z" },
- { url = "https://files.pythonhosted.org/packages/35/f8/9c66bfc55361ec6d0e4040a0337fb5924ceb23de4648b8a81ae9d33b2b38/triton-3.6.0-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d002e07d7180fd65e622134fbd980c9a3d4211fb85224b56a0a0efbd422ab72f", size = 188400296, upload-time = "2026-01-20T16:00:56.042Z" },
- { url = "https://files.pythonhosted.org/packages/49/55/5ecf0dcaa0f2fbbd4420f7ef227ee3cb172e91e5fede9d0ecaddc43363b4/triton-3.6.0-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ef5523241e7d1abca00f1d240949eebdd7c673b005edbbce0aca95b8191f1d43", size = 176138577, upload-time = "2026-01-20T16:16:25.426Z" },
- { url = "https://files.pythonhosted.org/packages/df/3d/9e7eee57b37c80cec63322c0231bb6da3cfe535a91d7a4d64896fcb89357/triton-3.6.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a17a5d5985f0ac494ed8a8e54568f092f7057ef60e1b0fa09d3fd1512064e803", size = 188273063, upload-time = "2026-01-20T16:01:07.278Z" },
- { url = "https://files.pythonhosted.org/packages/48/db/56ee649cab5eaff4757541325aca81f52d02d4a7cd3506776cad2451e060/triton-3.6.0-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0b3a97e8ed304dfa9bd23bb41ca04cdf6b2e617d5e782a8653d616037a5d537d", size = 176274804, upload-time = "2026-01-20T16:16:31.528Z" },
- { url = "https://files.pythonhosted.org/packages/f6/56/6113c23ff46c00aae423333eb58b3e60bdfe9179d542781955a5e1514cb3/triton-3.6.0-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:46bd1c1af4b6704e554cad2eeb3b0a6513a980d470ccfa63189737340c7746a7", size = 188397994, upload-time = "2026-01-20T16:01:14.236Z" },
+ { url = "https://files.pythonhosted.org/packages/94/fa/f856e24deb462d5f18bd4b5a746957862ab9b6ee5834bda60605ec348366/triton-3.7.1-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9497f2e696ee368862a181a90b2dcc03ca978cc4f602abd67c7d81022a6988e1", size = 184692359, upload-time = "2026-06-17T20:03:48.288Z" },
+ { url = "https://files.pythonhosted.org/packages/c4/6f/fb96d15db6f36d6eae4cafb998c2e0353bf59d7c4ea1662d7497f269134a/triton-3.7.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7e40869937a68206ec70d7f25bb7ec6433cb083f9135e1f36dbd318dc449a728", size = 197719725, upload-time = "2026-06-17T19:53:20.419Z" },
+ { url = "https://files.pythonhosted.org/packages/00/42/c5089d4d9327fcd1e862c599cc2927f39418f84dd11a84cb2ccff9d4787a/triton-3.7.1-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:cdbfc09d9ec58bc5e68321525653220de7515c199e7a8097a97c85e62b52cd0a", size = 184694629, upload-time = "2026-06-17T20:03:53.444Z" },
+ { url = "https://files.pythonhosted.org/packages/07/42/2c3ac59253ae8892b6f307875263dd23dc875cdf732d3aea40d6d41fb7cb/triton-3.7.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:58c0e131da05134a2a4788ccbcc0c1105cf0f54c8e98f19e34cd465396dc15eb", size = 197729241, upload-time = "2026-06-17T19:53:27.801Z" },
+ { url = "https://files.pythonhosted.org/packages/40/71/e01aa7ad573883ed9456f130226babdec70b005e098c4d6226a6238e761b/triton-3.7.1-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fe4ea396a06171f1f1f58cbd39c70b09294398f7dd7c620939bab54ad6f934fa", size = 184705764, upload-time = "2026-06-17T20:03:59.064Z" },
+ { url = "https://files.pythonhosted.org/packages/a4/09/5683146fda6a2b569deb78ccfd8fbfea8bfe55f726b081c0a6bb18dd6f28/triton-3.7.1-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2020153b08280415ec0da6607834e79166442147e78e144df06b508c75b186d2", size = 197729537, upload-time = "2026-06-17T19:53:35.516Z" },
+ { url = "https://files.pythonhosted.org/packages/e9/f8/448220c3092019f9fdfab39ec47985968181d67da34b44f6a7f6280a5cbb/triton-3.7.1-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c58e4c61f0c73b5dba3b5d19b4a7093c32f90dc18b2a7f121a7c16ccd31107b7", size = 184814760, upload-time = "2026-06-17T20:04:04.984Z" },
+ { url = "https://files.pythonhosted.org/packages/f0/ac/229b7d4589d2e5937310e72c6d46e89599d16a4a12b479ffa1499fee8eb8/triton-3.7.1-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:10ba85fa2cca4a2fbdeb36bf1cb082f2c252bda55bf9fccd74f65ec5bc647e68", size = 197824404, upload-time = "2026-06-17T19:53:42.772Z" },
]
[[package]]
@@ -4434,11 +4756,11 @@ wheels = [
[[package]]
name = "urllib3"
-version = "2.6.3"
+version = "2.7.0"
source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/c7/24/5f1b3bdffd70275f6661c76461e25f024d5a38a46f04aaca912426a2b1d3/urllib3-2.6.3.tar.gz", hash = "sha256:1b62b6884944a57dbe321509ab94fd4d3b307075e0c2eae991ac71ee15ad38ed", size = 435556, upload-time = "2026-01-07T16:24:43.925Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/53/0c/06f8b233b8fd13b9e5ee11424ef85419ba0d8ba0b3138bf360be2ff56953/urllib3-2.7.0.tar.gz", hash = "sha256:231e0ec3b63ceb14667c67be60f2f2c40a518cb38b03af60abc813da26505f4c", size = 433602, upload-time = "2026-05-07T16:13:18.596Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/39/08/aaaad47bc4e9dc8c725e68f9d04865dbcb2052843ff09c97b08904852d84/urllib3-2.6.3-py3-none-any.whl", hash = "sha256:bf272323e553dfb2e87d9bfd225ca7b0f467b919d7bbd355436d3fd37cb0acd4", size = 131584, upload-time = "2026-01-07T16:24:42.685Z" },
+ { url = "https://files.pythonhosted.org/packages/7f/3e/5db95bcf282c52709639744ca2a8b149baccf648e39c8cc87553df9eae0c/urllib3-2.7.0-py3-none-any.whl", hash = "sha256:9fb4c81ebbb1ce9531cce37674bbc6f1360472bc18ca9a553ede278ef7276897", size = 131087, upload-time = "2026-05-07T16:13:17.151Z" },
]
[[package]]
@@ -4645,6 +4967,18 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/6f/28/258ebab549c2bf3e64d2b0217b973467394a9cea8c42f70418ca2c5d0d2e/websockets-16.0-py3-none-any.whl", hash = "sha256:1637db62fad1dc833276dded54215f2c7fa46912301a24bd94d45d46a011ceec", size = 171598, upload-time = "2026-01-10T09:23:45.395Z" },
]
+[[package]]
+name = "wheel"
+version = "0.47.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "packaging" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/39/62/75f18a0f03b4219c456652c7780e4d749b929eb605c098ce3a5b6b6bc081/wheel-0.47.0.tar.gz", hash = "sha256:cc72bd1009ba0cf63922e28f94d9d83b920aa2bb28f798a31d0691b02fa3c9b3", size = 63854, upload-time = "2026-04-22T15:51:27.727Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/87/1b/9e33c09813d65e248f7f773119148a612516a4bea93e9c6f545f78455b7c/wheel-0.47.0-py3-none-any.whl", hash = "sha256:212281cab4dff978f6cedd499cd893e1f620791ca6ff7107cf270781e587eced", size = 32218, upload-time = "2026-04-22T15:51:26.296Z" },
+]
+
[[package]]
name = "whisperspeech"
version = "0.7"