Pretrained Model Release Roadmap

[lucasjinreal]

I planned to release pertained models in the next few month. It's not easy to train all supported models. I still encourage community can help train any version model you interested in and PR to main repo.

• DETR coco;
• AnchorDETR coco;
• YOLOX-s;
• DETR ONNX;
• DETR tensorrt;
• YOLOv7-Mask;
• more

[amirhp110]

Would you please some sample code on how to train pretrained model. The command option --opts MODEL.WEIGHTS [pre/trained-model-name.pth does not work for me.

Another question:
Training on my laptop with RTX 3080 is taking 3 days, expected, but training the same on machine with 8 x A100 40GB, is taking 1.8 days!
Do I need to change any parameters to take advantage of whole power? Only 5 GB of each 8 GPU is being utilized!

I am following Aarohi's post to train custom dataset, so let me know if any yaml file needs changes for better performance. I did increase the batches and didn't help.

[lucasjinreal]

@amirhp110 Just

python train_det.py --config-file configs/coco/exp/yolo2go_mobileone_exp1.yaml --num-gpus 2 --resume

[amirhp110]

Thanks for quick response
Don't we need to somehow provide the pre-trained model name?
Like in yolov7 (the other gir repo) the yolov7.pt is provided in the command or some yaml config.