AutoBatch cudnn.benchmark=True fix

[glenn-jocher]

May resolve #9287

Signed-off-by: Glenn Jocher glenn.jocher@ultralytics.com

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Redefined conditions for batch size determination and disabled CUDNN benchmark by default.

📊 Key Changes

• Added a condition to return the default batch size when torch.backends.cudnn.benchmark is enabled.
• Commented out the line that sets torch.backends.cudnn.benchmark to True by default, referencing an issue in the YOLOv5 repository.

🎯 Purpose & Impact

• The added condition helps to avoid altering the batch size if CUDNN's benchmarking feature is activated. This can ensure stability and avoid performance degradation in certain cases.
• Disabling the CUDNN benchmark by default aims to bypass potential problems with the AutoBatch feature, as reported in an issue. This change prioritizes model and training stability over the possible but not guaranteed speed improvements benchmarking might provide.
• Users may notice a change in training speed, but with a likely increase in predictability and stability during model training, especially when reproducing results across different runs.

[glenn-jocher]

Fix successfully tested in Colab on T4 GPU: