Bug: utils/export_yolo11.py fails with ModuleNotFoundError: No module named 'onnxscript'

[feli-santos]

When running the utils/export_yolo11.py script, the export process fails with a ModuleNotFoundError because recent versions of PyTorch (torch) have a new, unlisted dependency for their ONNX exporter.

The torch.onnx.export() function now defaults to using the dynamo exporter, which internally requires the onnxscript library. Since this library isn't listed as a dependency, the script crashes.

How to Reproduce

This bug can be reliably reproduced in any clean Python environment (like a Docker container) that installs the latest versions of the required libraries.

1. Set up an environment with python3-pip, git, and torch.
2. Clone this repository: git clone https://github.com/marcoslucianops/DeepStream-Yolo.git
3. Install the other dependencies: pip3 install ultralytics onnx onnxslim onnxruntime
4. Get a model file (e.g., yolo11x.pt)
5. Try to run the export script: python3 DeepStream-Yolo/utils/export_yolo11.py -w yolo11x.pt
6. The script will fail with the following traceback:

   Traceback (most recent call last):
     File ".../DeepStream-Yolo/utils/export_yolo11.py", line 141, in <module>
       main(args)
     File ".../DeepStream-Yolo/utils/export_yolo11.py", line 107, in main
       torch.onnx.export(
     ...
     File "/usr/local/lib/python3.10/dist-packages/torch/onnx/_internal/exporter/_core.py", line 18, in <module>
       import onnxscript
   ModuleNotFoundError: No module named 'onnxscript'
   

Proposed Fix

Force the Legacy Exporter (Code Change)

A code-level fix is to modify utils/export_yolo11.py to force PyTorch to use the legacy (TorchScript-based) exporter, which does not require onnxscript.

This is done by adding dynamo=False to the torch.onnx.export() call (around line 107):

Current Code:

    torch.onnx.export(
        model, onnx_input_im, onnx_output_file, verbose=False, opset_version=args.opset, do_constant_folding=True,
        input_names=['input'], output_names=['output'], dynamic_axes=dynamic_axes if args.dynamic else None
    )

Fixed Code:

    torch.onnx.export(
        model, onnx_input_im, onnx_output_file, verbose=False, opset_version=args.opset, do_constant_folding=True,
        input_names=['input'], output_names=['output'], dynamic_axes=dynamic_axes if args.dynamic else None,
        dynamo=False  # <-- Add this argument
    )

This makes the script compatible with newer PyTorch versions without requiring an extra dependency.