ultralytics 8.3.141

v8.3.141 - `ultralytics 8.3.141` Automatically detect RTDETR models (#20578)

on Python PyPI

🌟 Summary

This update introduces automatic detection and seamless handling of RTDETR models, along with several improvements for GPU selection, code clarity, dataset accessibility, and overall robustness. 🚀

📊 Key Changes

• Automatic RTDETR Model Detection:
  • The YOLO class now automatically recognizes and initializes RTDETR models from checkpoints, making it effortless to use RTDETR alongside other Ultralytics models.
  • You can now re-initialize a model from an existing model instance without duplicating memory or data.
• Smarter GPU Selection:
  • GPU selection now uses a percentage of free memory (e.g., 20%) instead of a fixed amount, improving compatibility across different hardware.
• YOLO Libtorch C++ Fix:
  • Fixed CUDA device errors in the YOLOv8 C++ example by specifying the device during model loading.
• Cleaner Code & Maintenance:
  • Simplified the TaskAlignedAssigner module and classification loss function for better readability and maintainability.
  • Improved handling of legacy image transforms in classification prediction to prevent errors.
  • Streamlined the Open Images V7 dataset download script for easier setup.
• Enhanced Testing & Docs:
  • VisualAISearch tests now auto-download required images and skip unsupported environments for more reliable testing.
  • Added a Colab badge to the HomeObjects-3K dataset docs for one-click model training in Google Colab.

🎯 Purpose & Impact

• Effortless Model Usage:
  • Users can now load RTDETR models just like any other YOLO model, with no extra steps or manual configuration.
  • Developers can pass model instances directly to the YOLO class, avoiding unnecessary duplication and saving memory.
• Better Hardware Compatibility:
  • GPU selection adapts to different devices, making it easier to run Ultralytics tools on a wide range of systems without manual tuning.
• Improved Reliability:
  • Fixes and enhancements across C++ inference, testing, and prediction make the platform more robust for both developers and end users.
• Faster Onboarding & Prototyping:
  • The new Colab badge and simplified dataset scripts lower the barrier for beginners and speed up experimentation for researchers.
• Cleaner, More Maintainable Codebase:
  • Ongoing code quality improvements help ensure long-term stability and ease of contribution.

✨ This release makes Ultralytics models—especially RTDETR—even easier to use, more reliable, and more accessible for everyone!

What's Changed

• Fix CUDA Error in YOLO Libtorch C++ Implementation by @ssapsu in #16390
• Clean up TaskAlignedAssigner module by @Laughing-q in #20725
• Update Autodevice memory check by using percentage form by @Laughing-q in #20714
• Speed up semantic search test with 4 images for fast embedding by @RizwanMunawar in #20704
• Eliminate loss.detach() temp return variable by @genji970 in #20721
• Add HomeObjects-3K notebook in docs by @RizwanMunawar in #20732
• Handle legacy CenterCrop in Classify prediction by @Y-T-G in #20715
• Simplify open-images-v7.yaml by @glenn-jocher in #20733
• ultralytics 8.3.141 Automatically detect RTDETR models by @Y-T-G in #20578

New Contributors

• @ssapsu made their first contribution in #16390

Full Changelog: v8.3.140...v8.3.141