ultralytics 8.3.222

v8.3.222 - `ultralytics 8.3.222` Enable TFLite INT8 export for end2end models (#22503)

on Python PyPI

🌟 Summary

TFLite INT8 export is now enabled for end2end models, plus improved quantized NMS accuracy and full IMX500 Classification support—expanding fast, small-footprint deployments across mobile and edge devices. 🚀📱

📊 Key Changes

• TFLite INT8 for end2end models ✅
  • Enables exporting quantized TFLite INT8 models even when model.end2end=True (Enable TFLite INT8 export for end2end models – PR #22503).
  • Works with simple commands like:
    • yolo export model=yolov10n.pt format=tflite int8
    • yolo val task=detect model=yolov10n_saved_model/yolov10n_int8.tflite data=coco128.yaml
• More accurate NMS with quantization 🧠
  • Improves NMS stability and accuracy for quantized exports, especially with TFLite INT8 (Improve accuracy of NMS export with quantization – PR #22487).
• IMX500: Classification export + inference 🎛️
  • Adds full Classification task support for IMX500, docs, examples, and benchmarks (Support IMX export and inference for classification – PR #21405).
• Documentation and DX polish ✨
  • Consistent US English across docs, logs, tests (e.g., grey→gray, labelled→labeled) (Standardize American English – PR #22508).
  • Clearer hyperparameter tuning guidance (Clarify tuning limitations – PR #22513).
  • Fixed a segmentation docs snippet variable name (Fix variable name – PR #22501).
  • Added a Construction-PPE training video to the dataset page (Add PPE video – PR #22485).
  • Shorter docs banner copy (Update Banner Text – PR #22486).
• CI and tooling upgrades 🧰
  • Expanded Ruff rules in docs CI (Update docs.yml with RUF and FA – PR #22480).
  • Cleaner Codespell config to reduce false positives (Update codespell settings – PR #22509).
  • GitHub Actions artifact steps bumped to latest (download-artifact v6 – PR #22489, upload-artifact v5 – PR #22490).

🎯 Purpose & Impact

• Broader edge deployment options 🌍
  • INT8 TFLite for end2end models unlocks smaller, faster models on mobile and embedded devices with minimal friction.
• Better accuracy under quantization 📈
  • Improved NMS logic helps retain performance when using INT8 exports—meaning more reliable results on constrained hardware.
• IMX500 across tasks 🧩
  • With Classification now supported alongside Detection and Pose, teams can cover more use cases on the Raspberry Pi AI Camera and similar devices.
• Smoother onboarding and contributor experience 🧑‍💻
  • Clearer docs, fewer CI false positives, and corrected examples reduce confusion and speed up development.

Happy exporting and deploying! ⚡

What's Changed

• Update docs.yml with RUF and FA ruff formatting by @glenn-jocher in #22480
• Update Banner Text by @sergiuwaxmann in #22486
• Bump actions/download-artifact from 5 to 6 in /.github/workflows by @dependabot[bot] in #22489
• Bump actions/upload-artifact from 4 to 5 in /.github/workflows by @dependabot[bot] in #22490
• Add https://youtu.be/lFaVnrhMmaE to docs by @RizwanMunawar in #22485
• Improve accuracy of NMS export with quantization by @Y-T-G in #22487
• Fix British english by @glenn-jocher in #22508
• Update pyproject.toml codespell settings by @glenn-jocher in #22509
• Support IMX export and inference for classification by @ambitious-octopus in #21405
• Clarify hyperparameter tuning limitations in guide by @amm1111 in #22513
• Fix variable name from res to results by @RizwanMunawar in #22501
• ultralytics 8.3.222 Enable TFLite INT8 export for end2end models by @Y-T-G in #22503

Full Changelog: v8.3.221...v8.3.222