🌟 Summary

The v8.3.31 release of Ultralytics introduces enhancements to automatic batch size estimation during model training, which aims to optimize memory usage and manage CUDA memory issues more effectively.

📊 Key Changes

• Batch Size Optimization: Implemented auto_batch functionality to determine the best batch size by evaluating memory consumption.
• Improved Profiling: The profiling tools have been updated to include a max_num_obj parameter for better batch size accuracy.
• Error Management: Introduced logging for CUDA out-of-memory warnings and an automatic switch to CPU computation when necessary.
• Documentation Updates: Removed the verbose argument from training documentation as it was deemed ineffective.

🎯 Purpose & Impact

• Efficient Memory Use: Automatically adjusting batch sizes helps prevent overloading GPU memory, resulting in more efficient and stable training sessions. This is particularly beneficial for preventing abrupt interruptions due to memory errors.
• Greater Reliability: By switching to CPU processing when encountering memory errors, the system maintains training continuity, avoiding crashes and ensuring an uninterrupted user experience.
• Simplified User Experience: Streamlining training configuration by removing unnecessary options enhances usability, making the setup less complex for users.

What's Changed

• Remove verbose arg from train docs. by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/17257
• ultralytics 8.3.31 add max_num_obj factor for AutoBatch by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/17514

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.30...v8.3.31