PyTorch 1.6.0 update with native AMP

[Lornatang]

PyTorch have Automatic Mixed Precision (AMP) Training.

fix #555 and #557

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Refinement of mixed precision training in YOLOv5 with native PyTorch support.

📊 Key Changes

• 🧹 Removed old APEX-based mixed precision code.
• ➕ Added native PyTorch mixed precision via torch.cuda.amp.
• 🛠 Fixed memory reporting to use torch.cuda.memory_reserved instead of torch.cuda.memory_cached.
• 🧱 Replaced manual mixed precision with amp.GradScaler for gradient scaling.
• 🎚 Removed conditional for CPU in DDP (Distributed Data Parallel) setup.
• 🧹 Code cleanup for readability and simplicity.
• 🔄 Changed some if conditions to reflect new CUDA boolean variable usage.

🎯 Purpose & Impact

• 💪 Increased Efficiency: Use of the latest native PyTorch functionalities for mixed precision training should improve the training efficiency and speed.
• 🧠 Improved Clarity and Maintenance: The removal of external dependencies like NVIDIA's APEX and the transition to built-in methods makes the code cleaner and easier to understand and maintain.
• 🐞 Bug Fixes: Fixes the memory reporting method which should provide more accurate gpu memory usage information.
• ▶️ User Experience: Users with the latest PyTorch version will experience a more streamlined setup process without the need for additional APEX installation.
• 📈 Consistency and Stability: The refactor should make training behavior more consistent and stable across different hardware setups.