I can see that in ultralytics, uploading a test image works, and the model works fine. Once the training is complete I download the PyTorch file and then used this command to convert to an .mlpackage ...

Shouldn't this (operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK) fix it? Why not? How can I know which torch operator is causing the problems and how can I fix this? Versions.