Default AutoBatch 0.8 fraction

[glenn-jocher]

Should be more stable, ran into CUDA OOM on validation when train utilization is too high.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced compatibility and optimization of batch size for memory usage in YOLOv5 models.

📊 Key Changes

• Adjusted warning for YOLOv5 v6.2 ClassificationModel to clarify its non-compatibility with AutoShape.
• Modified the default memory fraction used in the autobatch function from 0.9 to 0.8.

🎯 Purpose & Impact

• The update provides clearer guidance to users attempting to use AutoShape with the ClassificationModel, preventing confusion and potential errors. 🛑
• Reducing the default memory fraction allows for larger models or datasets by lowering the risk of running out of CUDA memory, leading to fewer crashes and enhanced performance on systems with limited resources. 📈🖥️