Checkpoint from PyTorch-trained model?

[nirbenz]

Thank you for this awesome repository.
Due to some changes in the training scheme v.s. original darknet code, I wonder if the provided PyTorch weights are converted from the original .weights file - or rather, results of a fresh training session in PyTorch.
If it's the former, it would be wonderful of the results of PyTorch training could be provided as well!

Thanks!

[glenn-jocher]

@nirbenz these are the official yolov3 weights trained with darknet. Pytorch training is still under development with outstanding issues (see Issue #22).

• PyTorch format (darknet trained): https://storage.googleapis.com/ultralytics/yolov3.pt
• Darknet format (darknet trained): https://pjreddie.com/media/files/yolov3.weights

[nirbenz]

Thanks for the quick answer. I might join the training effort real soon :)

[glenn-jocher]

@nirbenz no problem. I forgot to mention, pytorch training can be done from scratch or starting from the official yolov3 weights by doing -resume = True. In this case you can train all the parameters or transfer learn on a subset of the parameters (like the yolo layers only, or all the layers after darknet-53 for example).

If your application is similar to coco then it should be faster to begin from the official weights, however we see quick drops in mAP when doing this, indicating a potential problem somewhere in the training code. This is where things remain stuck at at the moment unfortunately.