SDAPT

Domain Adaptive Pretraining in Split Learning
for Vision Foundation Models

Stan van Honk · Vasileios Tsouvalas · Nirvana Meratnia

Project Page ↗

Abstract

SDAPT adapts vision foundation models to distributed, unlabeled domain data. Clients retain a frozen image frontend and send features to a shared server, where self-distillation learns from a combined batch of client representations. LoRA adapters and fully trained final blocks update the server model; an exponential moving average teacher supplies the training targets. The adapted backbone then stays frozen while each client trains a private prediction head. The workflow covers satellite imagery, chest radiographs, and sketches, with FedAvg, MoCoSFL, and no-adaptation comparisons.

Setup

Create and activate the sdapt environment:

git clone https://github.com/Nousphera/SDAPT.git
cd SDAPT
python3.12 -m venv --prompt sdapt .venv
source .venv/bin/activate

From the activated sdapt environment, install the requirements:

python -m pip install -r requirements.txt

Data and Weights

Prepare one dataset, or omit --dataset to prepare all three:

python scripts/prepare_data.py --dataset quickdraw

Dataset	Source / Local Input
QuickDraw	DomainNet · automatic download
fMoW	WILDS · download or import with --fmow-dir /path/to/fmow_v1.1
CheXpert	Training images → data/chexpert/raw/CheXpert-v1.0-small/
CheXlocalize	Test images → data/chexpert/raw/CheXlocalize/CheXpert/

CheXpert needs train.csv and train/; CheXlocalize needs test_labels.csv and test/. Obtain access from their publishers, extract the folders above, then run preparation with --dataset chexpert.

Prepare Model

Download and verify your chosen pretrained encoder:

# Original SDAPT backbone
python scripts/prepare_model.py --model dinov2-vitb14

# DINOv3 Large
python scripts/prepare_model.py --model dinov3-vitl16

DINO: Small / Base, patch 8 or 16. DINOv2: Small / Base / Large / Giant, with or without register tokens. DINOv3: Small / Small+ / Base / Large / Huge+ / 7B, plus Large and 7B satellite weights. Use --list-models for exact names.

Run Experiments

From the repository root, set the shared options and run your chosen method:

common=(--dataset quickdraw --model dinov2-vitb14
        --nr_of_clients 10 --batch_size 15 --nr_of_epochs 100
        --steps_per_epoch 250 --accumulation_steps 7
        --shots 4 8 16 0 --head_epochs 25 --random_seed 2024 --device cuda)

# SDAPT (Ours)
python scripts/run_experiment.py "${common[@]}" --method sdapt

# Federated DINO adaptation
python scripts/run_experiment.py "${common[@]}" --method fedavg --local_steps 50

# Split MoCo-v3 adaptation
python scripts/run_experiment.py "${common[@]}" --method mocosfl

# Frozen pretrained encoder, without adaptation
python scripts/run_experiment.py "${common[@]}" --method no-dapt

Note

0 selects full-data linear probing. Shot budgets are capped by each client's available labels; saved metrics include the actual counts. All methods reuse the same IID partitions. The run simulates clients and server on one machine.

For DINOv3, choose --model dinov3-vitl16, dinov3-vith16plus, or dinov3-vit7b16. Use --parameter_dtype bfloat16 --gradient_checkpointing to reduce frozen-weight and activation memory. Trainable weights and EMA teachers stay in FP32; full-model baselines need more memory. These are additional backbones; the original experiments use DINOv2 ViT-B/14.

Metrics, private heads, and adapted teacher weights are saved under results/. To evaluate a saved teacher:

python scripts/run_experiment.py --dataset quickdraw --model dinov2-vitb14 \
  --method sdapt --stage evaluate \
  --checkpoint results/sdapt-dinov2-vitb14-quickdraw-seed2024/teacher.pt \
  --nr_of_clients 10 --shots 4 8 16 0 --device cuda

Code License · DINO · DINOv2 · DINOv3