pytorch 1.6 compatibility

[rafale77]

I just tried loading the yolov5 implementation and ran into this exact issue:
ultralytics/yolov5#561

Could the models be re-trained on pytorch1.6 to avoid this?

[WongKinYiu]

yes, you can train the models from scratch to avoid this issue.

[rafale77]

I found a way to make work on pytorch1.6 without retraining:

ultralytics/yolov5@43a616a#diff-4f7ffd108bb4c62b490d3debfd3bc264.

Do you want me to submit a PR?

[WongKinYiu]

do i just need to add this line?

m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatability

[rafale77]

Yep!

[rafale77]

By the way, I see that the base of your implementation is from the ultralytic release version 2.0. You could probably move to their latest version 3.0.