🌟 Summary

The Ultralytics YOLO v8.2.90 release focuses on improving memory management on macOS devices and updating dependencies to ensure compatibility and stability.

📊 Key Changes

• Apple MPS Memory Optimization: Integrated torch.mps.empty_cache() for macOS to enhance memory management when using MPS devices.
• ONNXSlim Dependency Update: Reverted onnxslim to version 0.1.32 to address export issues with YOLOv10 for TFLite.
• Default Save Behavior: Adjusted behavior based on the interface: now defaults save to True for CLI and False for Python scripts.

🎯 Purpose & Impact

• Performance Enhancements: The memory optimization for macOS users can lead to significant training speed improvements, potentially reducing training time by up to 40%.
• Stability Improvements: Updating the onnxslim dependency ensures smoother model export processes and prevents compatibility errors with new versions.
• User Experience: The change in default saving behavior aims to align with user expectations, automatically saving results in CLI usage, reducing confusion, and providing flexibility in Python environments.

Overall, this release promotes better resource management on macOS, secures compatibility with dependency updates, and fine-tunes the user interface experience for more intuitive interactions. 🚀💡

What's Changed

• Revert to ONNXSlim 0.1.32 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/16084
• MPS unified memory cache empty by @Oil3 in https://github.com/ultralytics/ultralytics/pull/16078
• Fix Visualization Arguments docs table by @MatthewNoyce in https://github.com/ultralytics/ultralytics/pull/16061
• ultralytics 8.2.90 Apple MPS unified memory clearing by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/16085

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.2.89...v8.2.90