Fix Flask REST API

[Zengyf-CVer]

@glenn-jocher
I fixed this bug.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Update and fix in the Flask REST API for YOLOv5 model deployment.

📊 Key Changes

• Clarified the description by changing "yolov5s object detection model" to "Flask REST API exposing a YOLOv5s model".
• Changed variable name from args to opt for argument parsing consistency.
• Added a workaround to bypass a known issue with rate limiting when using torch.hub.

🎯 Purpose & Impact

• The description change improves the readability and understanding of what the API does.
• The variable renaming is a minor change for internal code consistency and readability.
• The workaround for the torch.hub rate limit issue ensures that developers won't run into HTTP 403 errors when trying to load models, leading to smoother development and deployment experiences. This could be particularly impactful for users working with YOLOv5 model in cloud environments or with frequent reloads.

[glenn-jocher]

@Zengyf-CVer thanks for the PR! I looked into this and seems like PyTorch > 1.10 should have this bug fixed:

https://stackoverflow.com/questions/68901236/urllib-error-httperror-http-error-403-rate-limit-exceeded-when-loading-resnet1

Pytorch released a bugfix for 1.9 (i.e. 1.9.1) so updating to 1.9.1 should rectify this issue for good.

[Zengyf-CVer]

@glenn-jocher My configuration is torch1.10, but this bug appears.

[glenn-jocher]

@Zengyf-CVer sorry to hear that! That's very strange. Could you test 1.11 please? If there error is still present then I'll go ahead and merge with an explanation for that line, because that's the strangest lambda function I've ever seen.

[Zengyf-CVer]

@glenn-jocher
I've experimented with torch 1.11 and no similar problem occurs. But this problem still exists on torch 1.10. I checked it, and it is probably related to the reloading of github. I changed the parameter value of force_reload to False, and it can run on torch 1.10, and there is no need to add the line of lamda code.

# torch 1.10
    model = torch.hub.load("ultralytics/yolov5", "yolov5s", force_reload=False)  # force_reload to recache
# torch 1.11
# No need to do any operation, it can be used normally.

[glenn-jocher]

@Zengyf-CVer ok, thanks for the tests! I think we should leave the code alone then as the issue seems to be fixed in latest PyTorch release.

[Zengyf-CVer]

@glenn-jocher
I think so too, but it would be better to have some version hints in the code section.

[glenn-jocher]

@Zengyf-CVer yes, we could do an if torch<1.11 statement. Could you do this using this example?

yolov5/models/yolo.py

Lines 79 to 82 in 71621df

	if check_version(torch.__version__, '1.10.0'): # torch>=1.10.0 meshgrid workaround for torch>=0.7 compatibility
	yv, xv = torch.meshgrid(torch.arange(ny, device=d), torch.arange(nx, device=d), indexing='ij')
	else:
	yv, xv = torch.meshgrid(torch.arange(ny, device=d), torch.arange(nx, device=d))

[Zengyf-CVer]

@glenn-jocher
I did a few more tests and found a problem:When I delete both .cache/torch/hub/ultralytics_yolov5_master and yolov5s.pt in the root directory of the workspace, I found that whether it is torch 1.10.2 or torch 1.11.0, there will be urllib.error.HTTPError: HTTP Error 403: rate limit exceeded this problem. So I wrote the lambda code again and found that the program can be used normally.So torch.hub._validate_not_a_forked_repo = lambda a,b,c: True is useful to solve the current problem.

[glenn-jocher]

@Zengyf-CVer understood, thanks for the detailed debugging! PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐