Not all dynamic dimensions are correctly set during model expert about ultralytics

[glenn-jocher]

Hello! It looks like you're on the right track with setting up dynamic dimensions for your custom YOLOv8 model export to TensorRT. However, the issue you're encountering with the first dimension not being set to -1 for certain tensors is likely due to how TensorRT handles dynamic shapes during the export process.

For TensorRT, when you specify dynamic axes, you need to ensure that the model's architecture and the export settings are correctly aligned to handle these dynamic dimensions. From your description, it seems that only the last two output tensors are correctly recognizing the dynamic batch size.

Here's a couple of things you might want to check or adjust:

1. Ensure Consistency in Model Definition: Double-check your model's forward function and make sure that all tensors that should have dynamic dimensions are correctly defined to handle varying batch sizes.
2. Adjust Export Settings: Sometimes, explicitly setting the dynamic axes for all outputs as you've done might not be recognized due to internal handling by the export function. You might need to adjust or simplify how you define these settings.

If these steps don't resolve the issue, consider providing a minimal reproducible example when you seek further help. This way, the community or the support team can replicate the issue exactly and provide more targeted assistance.

Here's a simplified dynamic setting you might try adjusting to:

dynamic_axes = {'input': {0: 'batch'}, 'output': {0: 'batch'}}
model.export(format='engine', imgsz=1920, batch=1, half=True, dynamic_axes=dynamic_axes, simplify=False, int8=False)

Hope this helps! Let us know how it goes. 🚀

[zvant]

@glenn-jocher
My forward function is quite complicated, involving multiple stages of scaling, NMS, sorting, indexing, matching, etc. It uses a for loop to iterate the samples in each batch. If TensorRT requires the dynamic dimension to be aligned through all these steps, it is not surprising it cannot work.

And yes, a simpler example does work with dynamic batch size.

My application does not really require dynamic batch size, so I might just set it to 1 for now. Thanks for the advice!

[glenn-jocher]

Hi there! 👋

It sounds like you've got a complex forward function with multiple operations that might indeed complicate the handling of dynamic dimensions in TensorRT. Sticking to a fixed batch size, especially if dynamic batching isn't a requirement for your application, is a practical approach. Setting the batch size to 1, as you mentioned, should help maintain consistency and stability in your model's performance.

If you ever need to revisit dynamic batching or encounter any other issues, feel free to reach out. Happy coding! 😊