Replace model.half() with amp.autocast()

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced GitHub Actions and model loading, improved inference precision, and added Kaggle support.

📊 Key Changes

• GitHub Actions now trigger on develop branch as well as master.
• Rebase instructions in greetings.yml changed to use the develop branch.
• Model loading uses mixed precision (FP16/FP32) during inference for efficiency.
• The training script train.py now uses attempt_download() for cleaner code.
• detect.py script adjustments for better GPU utilization with AMP (Automatic Mixed Precision).
• Improvements in hubconf.py and experimental.py to streamline model checkpoint loading.
• Kaggle badge added to tutorial.ipynb for easier access.
• utils/google_utils.py now includes a safe_download function to handle downloads robustly.
• Image loading now supports all channels in BGR color space in utils/datasets.py.

🎯 Purpose & Impact

• 🚀 Improve Continuous Integration by extending workflow checks to more branches.
• 🛠️ Provide cleaner code with better error handling for downloads.
• 💨 Enhance inference efficiency using mixed precision to reduce memory and speed up computations.
• ⛓️ Streamline model loading process, reducing potential for errors and redundancy.
• 🌐 Improve user experience by enabling easier access from Kaggle.
• 📉 Potentially reduce bugs and boost user confidence in codebase reliability.
• 👁️ Make the model better suited for various color channel formats, improving image handling.

[glenn-jocher]

Master branch appears faster. Autocast adds about 2 ms per batch :(

[glenn-jocher]

/rebase