Ultralytics v8.3.141 Release Announcement
Hello Ultralytics community!
We’re excited to announce the release of Ultralytics v8.3.141, delivering powerful improvements that make model usage even more seamless, robust, and adaptable across devices. Here’s everything you need to know about this update.
Summary
This release introduces automatic detection and handling of RTDETR models, smarter GPU selection for flexible hardware compatibility, refined code and dataset management, and a range of fixes to make Ultralytics tools more accessible and reliable for everyone.
New Features
Automatic RTDETR Model Detection
- The
YOLOclass now effortlessly recognizes and initializes RT-DETR models directly from checkpoints. - No extra steps—using RTDETR is now as simple as loading any other YOLO model.
- Models can be re-initialized from existing instances without duplicating memory or data.
Smarter GPU Selection
- GPU selection dynamically uses a percentage of free memory, enhancing compatibility with various hardware setups.
Improvements
- YOLO Libtorch C++ Example:
Fixed CUDA device errors by specifying the device during model loading. (PR #16390 by @ssapsu) - Cleaner Code:
Simplified the TaskAlignedAssigner module and classification loss function for easier maintenance. - Legacy Transform Handling:
Improved classification predictions by managing legacy image transforms (PR #20715 by @Y-T-G). - Open Images V7 Dataset:
Streamlined the open-images-v7.yaml for faster and simpler dataset setup. - Enhanced Testing:
VisualAISearch tests now auto-download required images and gracefully skip unsupported environments (PR #20704 by @RizwanMunawar). - Colab Badge in Docs:
HomeObjects-3K dataset docs now feature a Colab badge for one-click model training.
Bug Fixes
- Prevented CUDA device errors in YOLOv8 C++ inference (PR #16390 by @ssapsu).
- Improved error handling in legacy image transforms for classification.
- Enhanced robustness across testing, prediction, and export modes.
- Various internal code cleanups and improvements for long-term maintainability.
Why This Matters
- Ease of Use: RTDETR and other models can now be loaded and used interchangeably with minimal setup.
- Adaptability: Dynamic GPU memory management allows Ultralytics tools to work better out of the box on a wide range of devices.
- Developer Friendliness: The ongoing code refinements make it easier for contributors to get involved and for users to prototype new ideas.
- Getting Started Made Simple: Colab integration and improved dataset scripts let newcomers and researchers hit the ground running.
Acknowledgements & Contributors
Special thanks to the many community members and contributors, especially @ssapsu for their first contribution tackling CUDA device issues in the C++ interface!
Get Started
Give the new version a try:
Download Ultralytics v8.3.141 Release
Explore the changes in detail via the v8.3.141 changelog.
We encourage you to experiment with the new features and improved workflows! Your feedback, questions, or issues help us make Ultralytics even better—please share your experience or suggestions by replying to this post.
Thank you for your continued support and contributions. Together with the amazing YOLO community, we’re building the future of computer vision!
—
The Ultralytics Team