ultralytics 8.3.147

v8.3.147 - `ultralytics 8.3.147` Confusion Matrix export to CSV, XML, HTML, JSON, and SQL formats (#20834)

on Python PyPI

🌟 Summary

This release brings powerful new ways to export confusion matrix results, improved YOLOv7 ONNX/TensorRT inference support, enhanced OpenVINO documentation for YOLO11, and several usability and documentation updates. 📊🚀

📊 Key Changes

• Confusion Matrix Export Enhancements

  • You can now export confusion matrix results from model validation in multiple formats: CSV, XML, HTML, JSON, and SQL.
  • ConfusionMatrix class improved to support class names and easier export.
  • Documentation and tests updated to reflect these new export options.

• YOLOv7 ONNX & TensorRT Inference Support

  • Added detailed guides and scripts to help users export YOLOv7 models to ONNX/TensorRT and run inference with Ultralytics.
  • Clarified that only inference (not training) is supported for YOLOv7 in Ultralytics.

• OpenVINO Documentation Update for YOLO11

  • All OpenVINO integration docs now reference YOLO11, with updated export instructions, usage examples, and benchmarks for Intel CPUs, GPUs, and NPUs.
  • Added new hardware compatibility tips and troubleshooting guidance.

• Prediction Arguments & OBB Documentation

  • Added documentation for the new rect argument, explaining its effect on image padding and inference speed.
  • Fixed code example for accessing oriented bounding box (OBB) results.

• BatchNorm Initialization Fix

  • Prevents unintended changes to BatchNorm statistics when initializing models from YAML files, ensuring more stable model behavior.

• Training Parameter Docstring Update

  • Updated training function documentation to use the correct parameter name batch instead of batch_size.

🎯 Purpose & Impact

• Easier, Flexible Analysis

  • Exporting confusion matrices in various formats makes it simple to analyze, share, or report results using your preferred tools. Great for research, presentations, or audits.

• Broader Model Compatibility

  • YOLOv7 ONNX/TensorRT inference support lets users leverage more model types within Ultralytics, expanding deployment options.

• Up-to-date Hardware Guidance

  • OpenVINO documentation now fully supports YOLO11 and the latest Intel hardware, helping users achieve optimal performance.

• Improved Usability

  • Clearer documentation and parameter naming reduce confusion and help both new and experienced users get the most out of Ultralytics.

• More Reliable Training

  • Fixes to BatchNorm initialization ensure consistent model behavior, especially when loading custom architectures.

In summary:
This update makes validation results more accessible, expands model and hardware support, and improves the overall user experience for both developers and non-experts. 🎉

What's Changed

• Add guide for YOLOv7 ONNX inference by @Y-T-G in #18453
• OpenVINO docs update by @ambitious-octopus in #20731
• Fix OBB predict example by @Y-T-G in #20855
• Add rect to prediction arguments by @Y-T-G in #20857
• Prevent BatchNorm stats from being changed on model initialization from yaml by @Y-T-G in #18149
• Update train method docstrings to reflect correct batch parameter by @erfan-zekri in #20888
• ultralytics 8.3.147 Confusion Matrix export to CSV, XML, HTML, JSON, and SQL formats by @RizwanMunawar in #20834

Full Changelog: v8.3.146...v8.3.147