AutoBatch checks against failed solutions

[glenn-jocher]

@kalenmike this is a simple improvement to AutoBatch to verify that returned solutions have not already failed, i.e. return batch-size 8 when 8 already produced CUDA out of memory.

This is a halfway fix until I can implement a 'final solution' that will actively verify the solved-for batch size rather than passively assume it works.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved GPU memory management for batch size determination in YOLOv5.

📊 Key Changes

• Added emojis import to autobatch.py for enriched logging.
• Included device check before profiling to handle situations without CUDA.
• Enhanced display of CUDA memory statistics including name, total, reserved, allocated, and free memory.
• Revised profiling of batch sizes to store results more effectively.
• Implemented a polynomial fit to determine optimal batch size based on available memory.
• Added a catch for profiling failures to ensure chosen batch size has successfully passed memory requirement tests.

🎯 Purpose & Impact

• 🔍 Clearer insights into memory usage: Users gain better visibility about their GPU memory, aiding in troubleshooting and performance optimization.
• 💡 Better batch size predictions: The system more accurately predicts batch sizes for training, helping avoid memory-related crashes and improving utilization.
• 🛡 Increased robustness: Fallbacks when certain batch sizes fail during profiling leads to more stable and reliable model training sessions.