New Release: Ultralytics v8.3.156

[glenn-jocher]

Ultralytics v8.3.156 Release: Enhanced Export, Smarter Data, and More!

Summary

We’re excited to announce Ultralytics v8.3.156, bringing improved export workflows, smarter data handling, clearer training visualizations, streamlined examples, and upgraded docs to help you work faster and more efficiently. This release is especially valuable for users working with TensorRT, INT8 quantization, or those seeking to build custom pipelines with clear, practical guidance!

---

New Features & Major Improvements

TensorRT INT8 Export Improvements

• More Flexible Export:
  No longer requires dynamic=True for INT8 export, giving you greater flexibility (PR #20989 by @Y-T-G).
• Calibration Made Robust:
  INcomplete batches are always dropped during INT8 calibration, eliminating common sources of error.
• Enhanced Hardware Support:
  The best calibration algorithm is now auto-selected for DLA hardware for optimal compatibility.

Dataloader Flexibility

• drop_last Option:
  The dataloader now offers an option to drop incomplete batches during both training and export, promoting consistent results and preventing subtle bugs.

RT-DETR ONNXRuntime Example Revamp

• Standalone Example:
  Example runs out-of-the-box with a new download utility and dedicated requirements.txt for easy setup (PR #21059 by @onuralpszr).

Visualization & Documentation

• Classification Training Visualization:
  Batch plots are now more informative and easier to interpret for classification tasks (PR #21072 by @RizwanMunawar).
• Documentation Upgrades:
  • Improved FastSAM and utility function docs (PR #21086 by @glenn-jocher).
  • Expanded classification custom example to demonstrate how to use custom Trainer and Validator classes (PR #21076 by @Laughing-q).
  • Added a new community contributor to the documentation author list.

Performance Fixes

• Cache Optimization:
  Resolved an issue with double-caching during auto-batch estimation, reducing memory usage and potential slowdowns (PR #20891 by @XBastille).

---

Why This Matters

• More Reliable Model Export: Smoother TensorRT INT8 export with fewer errors and better device compatibility.
• Flexible Data Handling: Dropping incomplete batches means more robust training, calibration, and export.
• Easier Customization: Improved examples and documentation accelerate both beginner and advanced workflows.
• Better Insights: Clearer visualizations and improved docs make it easier to monitor and understand your training.
• Optimized Performance: Reduced unnecessary memory usage—train and deploy faster.

---

Full List of Changes

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

New Contributor:
Welcome @XBastille for their first contribution in PR #20891!

See the complete changelog for all updates: Ultralytics v8.3.156 full changelog

---

Try It Out & Share Feedback!

Ultralytics v8.3.156 is ready now—download the latest release, test the new features, and let us know how it improves your workflows. Your experience, insights, and suggestions directly shape the evolution of YOLO and the Ultralytics ecosystem.

Thank you for being part of the journey. As always, credit for these advancements goes to the incredible YOLO community and the dedicated Ultralytics team!