ultralytics 8.3.123

v8.3.123 - `ultralytics 8.3.123` Rockchip RKNN export INT8 quantization support (#20450)

on Python PyPI

🌟 Summary

This release adds INT8 quantization support for Rockchip RKNN model exports, improves result reporting, enhances TensorFlow export compatibility, and makes several codebase refinements for clarity and robustness. 🚀

📊 Key Changes

• RKNN INT8 Quantization: You can now export models for Rockchip devices in INT8 format, reducing model size and boosting inference speed.
• Export Options Updated: Documentation and export tables now include the new int8 option for RKNN exports, and exported model filenames clearly indicate INT8 or FP16 format.
• Profiling Functions Renamed: Profiling methods have been renamed for better clarity (profile → profile_ops, profile() method → run()).
• Result Summaries Improved: The verbose() method for results now provides clearer, more consistent output for both detection and classification tasks.
• TensorFlow Export Fixes: Improved compatibility and performance for TensorFlow SavedModel exports, especially with Attention blocks and GPU acceleration.
• Tracking Robustness: Feature extraction and tracking are now more reliable, especially when no objects are detected in some images.

🎯 Purpose & Impact

• Faster, Smaller Models on Rockchip: INT8 quantization allows users to deploy lighter, faster models on Rockchip hardware, making edge AI applications more efficient. ⚡
• Easier Model Exporting: Clearer export options and filenames help users quickly identify and select the right model format for their needs.
• Smoother Developer Experience: Renamed functions and improved documentation reduce confusion and make the codebase easier to use and maintain.
• Better Results Analysis: Enhanced result summaries make it simpler for users to interpret model outputs.
• Broader Compatibility: TensorFlow export improvements ensure models work reliably across more scenarios, including advanced architectures and GPU deployments.
• More Robust Tracking: Fixes prevent errors in tracking workflows, especially with challenging or sparse datasets, leading to more stable deployments.

Overall, this update brings valuable new features and refinements for both developers and end users, especially those working with Rockchip devices or deploying models in diverse environments. 🎉

What's Changed

• Rename 'profile' methods to avoid conflicts by @glenn-jocher in #20442
• Optimize Result.verbose() by @glenn-jocher in #20445
• Fix TFLite GPU delegate on Android by @Y-T-G in #20436
• Fix IndexError with empty predictions when using ReID by @Y-T-G in #20449
• ultralytics 8.3.123 Rockchip RKNN export INT8 quantization support by @oDestroyeRo in #20450

New Contributors

• @oDestroyeRo made their first contribution in #20450

Full Changelog: v8.3.122...v8.3.123