AMP check improvements backup YOLOv5n pretrained

[glenn-jocher]

More robust and faster, removes tolerance parameter.

EDIT: will not work for non-pretrained models...

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced PyTorch Automatic Mixed Precision (AMP) testing for accuracy and compatibility.

📊 Key Changes

• 🔍 Improved AMP testing with a new helper function amp_allclose for more robust comparison.
• 🖼️ Added compatibility for checking AMP using a local image or a fallback to an online image if the local one isn't available.
• ✅ Introduced DetectMultiBackend into the AMP checking process to support more model types.
• ✂️ Removed unnecessary CPU device check resulting in more streamlined code.
• 🛠️ Wrapped the AMP testing code in try-except block to handle exceptions better and provide help via URL if needed.

🎯 Purpose & Impact

• 🚀 The modifications enable better verification of the AMP functionality, ensuring that models deliver consistent results when using mixed precision, which can speed up inference without significant loss of accuracy.
• 💻 By automatically falling back to an online image when a local test image is unavailable, users can test AMP functionality in more situations, guaranteeing a smoother user experience.
• 📈 The added support for DetectMultiBackend during AMP validation expands compatibility, which benefits users who work with different types of models, enhancing the utility of YOLOv5 across a wider range of environments.