training-model.md

↤ BACK

Training the Model

Build Config File

Before we can train our model, we need to create a config fle.

python create_config.py

Start Training

Once everything is setup, you can start training the model:

python train.py nano

Size Options: nano | small | medium | large | extralarge

Size	Model	Depth Scale	Width Scale	Max Channels	mAPbox	mAPmask	Speed CPU ONNX	Speed A100 TensorRT	Params	FLOPs
nano	yolov8n-seg.pt	0.33	0.25	1024	36.7	30.5	96.1 ms	1.21 ms	3.4 M	12.6 B
small	yolov8s-seg.pt	0.33	0.50	1024	44.6	36.8	155.7 ms	1.47 ms	11.8 M	42.6 B
medium	yolov8m-seg.pt	0.67	0.75	768	49.9	40.8	317.0 ms	2.18 ms	27.3 M	110.2 B
large	yolov8l-seg.pt	1.00	1.00	512	52.3	42.6	572.4 ms	2.79 ms	46.0 M	220.5 B
extralarge	yolov8x-seg.pt	1.00	1.25	512	53.4	43.4	712.1 ms	4.02 ms	71.8 M	344.1 B

• mAP^val values are for single-model single-scale on COCO val2017 dataset.
• Speed averaged over COCO val images using an Amazon EC2 P4d instance.
• All Models are trained using an image size of 640px

Customizing:

Edit the following line in train.py with custom Training Configuration to tweak performance of our model.

args = dict(data='config.yaml', epochs=100, imgsz=640)

Resuming Training

If for some reason the training process gets interrupted, you can pick up from the last saved Epoch by running the following:

python resume.py nano

---