ultralytics 8.3.223

v8.3.223 - `ultralytics 8.3.223` Sort `scores` inplace in `FastNMS` with `use_triu=False` (#22537)

on Python PyPI

🌟 Summary

Ultralytics 8.3.223 improves post-processing correctness in FastNMS, stabilizes RT-DETR validation, and refactors TensorFlow/TFLite/Edge TPU/TFJS export for cleaner APIs and more reliable exports—plus better FLOPs accuracy via a dependency bump. 🚀

📊 Key Changes

• FastNMS fix (priority) 🔧

  • Sorts and writes back scores in-place when use_triu=False, ensuring alignment with sorted indices for consistent outputs in downstream consumers like NMSModel.
  • PR: Sort scores in FastNMS by @Y-T-G (#22537).

• RT-DETR validation stability ✅

  • Ensures ratio_pad is correctly formatted during validation to prevent crashes and stabilize evaluation.
  • PR: RTDETRDataset ratio_pad fix by @Y-T-G (#22525).

• Export refactor: TensorFlow/TFLite/Edge TPU/TFJS ♻️

  • Modular export utilities split into ultralytics.utils.export.engine and ultralytics.utils.export.tensorflow with wrappers to avoid fragile ops and improve reliability.
  • Cleaner exporter code paths for SavedModel, .pb, TFLite (INT8), Edge TPU, and TFJS.
  • PR: Refactor exports by @Laughing-q (#22115).

• Accurate FLOPs reporting 📈

  • Dependency bump: ultralytics-thop>=2.0.18 for better FLOPs calculations, improving benchmarking across models (e.g., YOLO11/YOLO26).
  • PR: Update dependency by @glenn-jocher (#22526).

• IMX export message clarity 🧭

  • Error message now correctly states IMX export supports detection, pose, and classification.
  • PR: Update IMX message by @lakshanthad (#22533).

• Standardized dataset class names 🇺🇸

  • American English naming in dataset YAMLs (e.g., “doughnut” → “donut”, “armour” → “armor”). Display-only; no impact on training indices.
  • PR: Remove British class names by @glenn-jocher (#22517).

🎯 Purpose & Impact

• More correct and stable NMS outputs 🛡️

  • The FastNMS fix eliminates mismatches between sorted indices and scores, preventing subtle ranking issues and improving result consistency for all users relying on FastNMS.

• Fewer validation crashes for RT-DETR 📏

  • Proper ratio_pad handling avoids runtime errors and increases reliability during validation and metrics computation.

• Easier, more reliable exports across ecosystems 🌐

  • Export refactor reduces edge-case failures in SavedModel/.pb/TFLite/Edge TPU/TFJS flows and centralizes TF-specific logic for cleaner maintenance and future improvements.

• Better, more consistent benchmarks 🔬

  • Updated ultralytics-thop improves FLOPs accuracy and compatibility, leading to more trustworthy comparisons across models and environments.

• Clearer developer experience ✨

  • IMX messaging prevents confusion; dataset naming standardization improves consistency in visualizations and docs.

Quick tip to upgrade:

pip install -U ultralytics

Thanks to contributors @Y-T-G, @Laughing-q, @glenn-jocher, and @lakshanthad! 🙌

What's Changed

• Remove British class names by @glenn-jocher in #22517
• Update IMX export error message with classification models support by @lakshanthad in #22533
• Update ultralytics-thop>=2.0.18 by @glenn-jocher in #22526
• Refactor exports for saved_model/pb/edgetpu/tfjs formats by @Laughing-q in #22115
• Return correctly formatted ratio_pad in RTDETRDataset by @Y-T-G in #22525
• ultralytics 8.3.223 Sort scores inplace in FastNMS with use_triu=False by @Y-T-G in #22537

Full Changelog: v8.3.222...v8.3.223