ultralytics 8.3.156

v8.3.156 - `ultralytics 8.3.156` Eliminate `dynamic=True` enforcement for TensorRT INT8 export (#20989)

on Python PyPI

🌟 Summary

This release enhances model export reliability—especially for TensorRT INT8 quantization—improves data handling for training and calibration, and brings documentation and example updates for a smoother user experience. 🚀

📊 Key Changes

• TensorRT INT8 Export Improvements:
  • Removed the strict requirement for dynamic=True during INT8 export, making the process more flexible.
  • INT8 calibration now always drops incomplete batches, preventing errors and ensuring consistent calibration.
  • Automatically selects the best calibration algorithm for DLA hardware, boosting compatibility.
• Dataloader Flexibility:
  • Added a drop_last option, letting users drop incomplete batches during data loading for both training and export workflows.
• RT-DETR ONNXRuntime Example Revamp:
  • The RT-DETR ONNXRuntime Python example is now fully standalone, with unnecessary dependencies removed and a new file download utility added.
  • Introduced a dedicated requirements.txt for easier setup.
• Classification Training Visualization:
  • Improved how training samples are visualized for classification models, making batch visualizations clearer and more informative.
• Documentation Enhancements:
  • FastSAM and utility function docs updated for clarity and easier navigation.
  • Expanded classification custom example to show how to use custom Trainer and Validator classes.
  • Added a new community contributor to the documentation authors list.
• Performance Fixes:
  • Resolved an issue with double caching during auto-batch estimation, reducing unnecessary memory usage.

🎯 Purpose & Impact

• More Reliable Model Export:
  • Users exporting INT8 models with TensorRT will experience fewer calibration errors and smoother hardware compatibility, especially on advanced devices.
• Flexible & Robust Data Handling:
  • Dropping incomplete batches ensures consistent training and calibration, reducing the risk of subtle bugs or crashes.
• Easier Customization & Setup:
  • The improved RT-DETR example and expanded classification docs make it easier for both beginners and advanced users to adapt workflows to their needs.
• Better User Experience:
  • Enhanced visualizations and clearer documentation help users understand their data and available features, leading to more effective model development.
• Performance Optimizations:
  • Fixes to caching and data loading streamline training, saving resources and time.

This update is recommended for all users who export models, work with INT8 quantization, or want improved examples and documentation. 🚀🛠️

What's Changed

• Add RTDETR-ONNXRuntime-Python file download and remove unnecessary dependencies by @onuralpszr in #21059
• Fix plot_images for classify training by @RizwanMunawar in #21072
• Fix: Cache optimization for auto-batch training by @XBastille in #20891
• Fix fast-sam.md links by @glenn-jocher in #21086
• Update classification custom example by @Laughing-q in #21076
• ultralytics 8.3.156 Eliminate dynamic=True enforcement for TensorRT INT8 export by @Y-T-G in #20989

New Contributors

• @XBastille made their first contribution in #20891

Full Changelog: v8.3.155...v8.3.156