New Release: Ultralytics v8.3.137

:rocket: Ultralytics v8.3.137 Release – Faster Text Embeddings & ONNX Export!

We’re excited to announce the release of Ultralytics v8.3.137, bringing major speed and efficiency upgrades for users of YOLOWorld, YOLOE, and text-based models, plus improved ONNX export reliability! Whether you’re training large open-vocabulary models or deploying workflows, this release offers significant gains in speed, maintainability, and usability.


:glowing_star: Summary

This update vastly improves training speed and resource utilization for YOLOWorld and YOLOE by introducing smarter caching of text embeddings and optimizations in embedding computations. Exporting to ONNX and TensorFlow is now even more robust. These changes mean a smoother, faster experience for everyone working with text-rich category models and exporting for deployment.


:new_button: New Features & Major Improvements

YOLOWorld & YOLOE Text Feature Optimization

  • Category Text Embedding Caching:
    Training is faster than ever thanks to new caching mechanisms for text embeddings, removing repeated computations across epochs.
  • Unified & Improved Embedding Generation:
    YOLOWorld and YOLOE trainers now follow a consistent and enhanced process for embedding storage and regeneration.
  • Modular Text Model Utility:
    The new build_text_model utility streamlines text model handling and compatibility.
  • Efficient Preprocessing:
    Batch preparation integrates text features more efficiently, reducing bottlenecks.
  • Model-Specific Embedding Saving:
    Embeddings are uniquely saved and only rebuilt when needed, cutting down redundancy.

PR: YOLO-World text features cache optimization by @h13-0
Contributor: @h13-0


Embedding Computation Optimization

  • Streamlined Index Handling:
    Embedding extraction is quicker due to smarter handling of embedding indices, especially helpful for large label sets.

PR: Optimization for embedding computation by @genji970
Contributor: @genji970


ONNX Export Improvements

  • Dependency Update:
    Upgraded onnxslim to version 0.1.53 for more reliable ONNX and TensorFlow exports.
  • Cleaner Test Logic:
    Simplified test conditions make the export pipeline easier to debug and maintain.
  • Jetson Export Support:
    Jetson-specific test logic is now re-enabled.

PR: Use onnxslim>=0.1.53 and simplify when dynamic=True by @inisis
Contributor: @inisis


:bullseye: Purpose & Impact

  • :high_voltage: Substantially Faster Text-based Model Training:
    Spend less time waiting on training and more time iterating.
  • :brain: Smarter Resource Usage:
    Reduced computational waste helps in both research and production environments.
  • :counterclockwise_arrows_button: Reliable Model Exports:
    Robust ONNX and TensorFlow exports mean stress-free deployment.
  • :hammer_and_wrench: Easier Debugging:
    Clean, clear test outputs help speed up development for everyone.

:raising_hands: New Contributors

Big thanks to our newest contributors for making this release stronger!


:link: Useful Links


:test_tube: Try It Out & Share Your Feedback!

We encourage everyone to update to v8.3.137, put these new features to the test, and let us know what you think. Your feedback is essential to the continuous improvement of Ultralytics – together, the community drives the progress!

Happy training and deploying! For any questions, suggestions, or feedback, drop a comment below or open an issue on Ultralytics GitHub.

— Thanks from the entire YOLO community and Ultralytics team!