New Release: Ultralytics v8.3.109

[glenn-jocher]

Ultralytics v8.3.109 Release: Improved Evaluation, Memory Management, and Smoother Exports!

Hello Ultralytics Community!

We’re excited to announce the release of Ultralytics v8.3.109, bringing you enhanced reliability, greater control, and performance improvements across the board. This update focuses on making your model evaluation, training, and export workflows even smoother—while refining some key areas like GPU memory management and RT-DETR validation.

---

Quick Summary

• Accurate model evaluation with new RT-DETR options.
• Simplified CoreML exports for classification models.
• Improved training stability with robust batch sizing and memory clearing.
• Greater documentation clarity and reliable CI behavior.
• Smoother development for both users and contributors!

Explore the full release here:
Ultralytics v8.3.109 Release Notes

---

New Features & Major Improvements

RT-DETR Validator Update

• Confidence Threshold Filtering: Added a user-configurable confidence filter to the RT-DETR model validator for truly representative evaluation scores.
  • PR by @Y-T-G: Add conf filter to RT-DETR validator

CoreML Export Simplification

• Cleaner Classification Model Exports: CoreML export is now more streamlined and robust for classification, removing unnecessary complexities.
  • PR by @rromanchuk: Remove unreachable CoreML Classify model logic

AutoBatch Logic Refinement

• Improved Automatic Batch Size: The logic for automatically selecting the optimal batch size during training is more stable and avoids overfitting to outliers.
  • PR by @Laughing-q: Revert polynomial fitting logic

Better GPU Memory Management

• Proactive OOM Prevention: The threshold for clearing GPU memory is now 50% (down from 90%), helping to safeguard against memory issues, especially on hardware with tight constraints.
  • PR by @Y-T-G: Reduce trigger threshold to clear GPU memory to 50%

---

Other Notable Changes

• Testing Improvements: Solution tests are now conditionally skipped on Linux (Python 3.11) and Raspberry Pi to prevent issues and keep your dev flow smooth.

  • PR by @RizwanMunawar: Enable Solutions tests except on problematic platforms

• CI & Documentation Upgrades: Fixed broken VS Code doc links and made our CI system even more resilient.

  • PR by @glenn-jocher: Update links.yml with set -o pipefail

---

Why This Matters

• More control, less friction: Fine-tune your evaluation process and export models with confidence.
• Fewer interruptions: Lower memory-clearing thresholds and robust batch sizing keep your hardware happy—goodbye to unexpected crashes!
• Documentation you can trust: Enjoy clearer guidance and robust, always-on testing.

---

Try It Out and Share Your Feedback!

Ready to explore the latest improvements?
Update with:

pip install -U ultralytics

Full details, discussions, and the complete changelog are available at the Ultralytics v8.3.109 Release Page.

Your feedback drives the roadmap—let us know how it works for you, share your experience, or open an issue if you spot a bug. Every insight helps us (and the community!) build even better tools.

A huge thank you to all contributors and users for your ongoing support and collaboration—the achievements are only possible because of you and this incredible community.

Happy experimenting!
The Ultralytics Team