New smart_inference_mode() conditional decorator

[glenn-jocher]

Applies torch.inference_mode() decorator if torch>=1.9.0 else torch.no_grad() decorator. Material speed improvements observed in detect.py and val.py.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Integration of a new smart_inference_mode decorator across various Ultralytics YOLOv5 files for improved handling of inference operations.

📊 Key Changes

• smart_inference_mode Decorator: A new decorator function, smart_inference_mode, has been created to wrap functions that previously used the torch.no_grad() decorator.
• Updates in Critical Files: Modifications made in detect.py, export.py, val.py, models/common.py, and models/yolo.py to use the new decorator.
• EMA Update Function: The Exponential Moving Average (EMA) update function in utils/torch_utils.py now uses smart_inference_mode instead of torch.no_grad().
• Minor Refactoring: The _make_grid function in models/yolo.py now accepts an additional parameter to handle torch version compatibility.

🎯 Purpose & Impact

• Enhanced Inference Performance: The change transitions to torch.inference_mode for PyTorch versions >=1.9.0, which provides a more performant context manager for inference operations, potentially speeding up the process.
• Backward Compatibility: The new decorator maintains compatibility with older versions of PyTorch by falling back to torch.no_grad() when necessary.
• Simplified Codebase: By replacing individual torch.no_grad() calls with a decorator that checks the version and decides the context, the codebase becomes cleaner and more maintainable.

These updates aim to ensure that the YOLOv5 codebase remains modern and efficient, aligning with the latest best practices for PyTorch inference, while still supporting older versions of the framework seamlessly. Users should benefit from potential performance improvements during model inference without needing to alter their own usage of the library. 🚀📈