Remove .train() mode exports

[glenn-jocher]

No common use cases.

Signed-off-by: Glenn Jocher glenn.jocher@ultralytics.com

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved ONNX export functionality for YOLOv5 models.

📊 Key Changes

• Removed the train parameter from export_onnx function signature.
• Set do_constant_folding=True by default in ONNX export.
• The model is set to evaluation mode (model.eval()) by default during export, instead of toggling between training and evaluation modes.
• Removed train parameter and related conditional logic in the run function and parse_opt function signature for export script.

🎯 Purpose & Impact

• ⚙️ Consistency: The ONNX export will consistently perform constant folding, which often results in a smaller and more efficient model.
• 👁️ Simplification: Simplifies the export process by removing the need for the user to specify the mode (train vs. eval) during export.
• 🔄 Usability: This change aids both new and existing users by reducing the number of parameters and potential for confusion.
• 🚀 Performance and Reliability: By always exporting the model in the evaluation mode, exported models are optimized for inference, which is the common use case, lending to potential performance improvements in downstream tasks.