Export with official nn.SiLU()

[glenn-jocher]

Remove utils.activations.SiLU replacement for nn.SiLU on export. This was in place for export support on some formats that did not recognize PyTorch's nn.SiLU module, but now all export formats fully support this. Note may require coremltools>=5.1 for CoreML exports, not tested with coremltools==4.1 support.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Simplified model activation functions and updated non-max suppression timing parameters in YOLOv5 export script.

📊 Key Changes

• Removed specific activation function handling (SiLU) during model export.
• Removed unused imports (torch.nn, Conv, and SiLU modules).
• Increased the base non-max suppression time limit from dynamic-only to a static base time plus a dynamic component based on batch size.

🎯 Purpose & Impact

• Simplifying activation functions helps streamline the export process and removes potential obstacles when converting models for mobile or other platforms. 🚀
• Cleaning up imports reduces clutter and slightly improves code maintainability. 🧹
• Adjusting the non-max suppression timing ensures the algorithm has enough time to process the results even with larger batch sizes, which could improve detection reliability at the cost of potential slight latency increase. ⏱️💡

These changes collectively work towards optimizing YOLOv5's performance and usability across different deployment scenarios. Users can expect a smoother export process for their trained models with these updates.