Fix TensorRT --dynamic excess outputs bug

[glenn-jocher]

Potential fix for #8790

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved TensorRT dynamic input handling in YOLOv5 🚀

📊 Key Changes

• Renamed dynamic_input variable to dynamic to simplify naming.
• Modified the logic for reshaping the input if the input image shape does not match the expected model shape.
• Enhanced error message clarity when an image shape mismatch occurs.

🎯 Purpose & Impact

• 🔨 Code Maintainability: Renaming the variable makes the code more readable and maintainable.
• 💡 Better Error Handling: Clear error messages help to diagnose issues related to input sizes quickly, making it easier for users to debug.
• ✨ Dynamic Resizing: Enhanced support for dynamic input sizes provides more flexibility for users wanting to process images of various dimensions without manual resizing.

[glenn-jocher]

@democat3457 confirm fix for #8866. Can you please review this PR and see if this approach makes sense?

[democat3457]

I have a better solution, should I make a PR?

Add this line:

...
self.context.set_binding_shape(self.model.get_binding_index('images'), img.shape)
self.bindings['images'] = self.bindings['images']._replace(shape=img.shape)
+ self.bindings['output'].data.resize_(tuple(self.context.get_binding_shape(self.model.get_binding_index('output'))))
...

[glenn-jocher]

@democat3457 got it! I'll add your line and remove my fix.

[glenn-jocher]

@democat3457 stupid question, is the previous line not modifying inplace? Does it need a reasignment like you have here?

self.bindings['images'] = self.bindings['images']._replace(shape=im.shape)

i.e. could this just be:

self.bindings['images']._replace(shape=im.shape)

[glenn-jocher]

@democat3457 how does this look?

[democat3457]

@democat3457 stupid question, is the previous line not modifying inplace? Does it need a reasignment like you have here?

self.bindings['images'] = self.bindings['images']._replace(shape=im.shape)

i.e. could this just be:

self.bindings['images']._replace(shape=im.shape)

It does not modify in place, the underscore just stops it from getting mixed up with a possible field of the namedtuple. (Also, I looked - there doesn't exist an in-place replacement, sadly)

[democat3457]

@democat3457 how does this look?

Looks good to me!

[glenn-jocher]

@democat3457 PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐