Fix for "RuntimeError: result type Float can't be cast to the desired output type long int"

[ThibaultCastells]

Hello,
I tried this repo and got the following message when experimenting with yolov5 (I am using Pytorch 1.12.1 and python 3.10.6):

Traceback (most recent call last):
File "/*****/yoloair/train.py", line 695, in
main(opt)
File "/*****/yoloair/train.py", line 591, in main
train(opt.hyp, opt, device, callbacks)
File "/*****/yoloair/train.py", line 376, in train
loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size
File "/*****/yoloair/utils/loss.py", line 123, in call
tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets
File "/*****/yoloair/utils/loss.py", line 222, in build_targets
indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices
RuntimeError: result type Float can't be cast to the desired output type long int

I fixed it by casting from Float to Int:
indices.append((b, a, gj.clamp_(0, int(gain[3] - 1)), gi.clamp_(0, int(gain[2] - 1))))

I figured that this could save some time to people getting the same issue.

[iscyy]

hi, can you tell me what training command you used to cause this error, I did not get a similar error using this repo (Python 3.9.6, Pytorch1.9.0)
Also do you get this error using the yolov5 repo?
https://github.com/ultralytics/yolov5/blob/v6.1/utils/loss.py#L217
This is consistent with the official yolov5_v6.1.

[ThibaultCastells]

Yes, my command was:
python3 -m torch.distributed.run --nproc_per_node 8 train.py --batch 64 --data data/coco.yaml --weights yolov5s.pt --device 0,1,2,3,4,5,6,7
But I think it is related to the Pytorch version rather than the command.

I did not try with yolov5 but noticed that some people got the exact same issue with yolov5 or other repo based on yolov5:
Yolov5: ultralytics/yolov5#8405
Yolor: WongKinYiu/yolor#270

Apparently it is a compatibility issue with Pytorch 1.12, but I thought that fixing the issue is a better solution than downgrading Pytorch.

If you think my fix is not good, this is another solution suggested in the yolor issue: WongKinYiu/yolor@1102f52

[iscyy]

hi, thank you for your reply, I read these links, one of the solutions: WongKinYiu/yolor@1102f52, it is consistent with the latest official code of yolov5, can you resubmit a PR?

[ThibaultCastells]

Okay no problem, then I am closing this issue. Here is the new pull request.