ultralytics 8.3.194

v8.3.194 - `ultralytics 8.3.194` Bump `onnxslim>=0.1.67` in Exporter (#21951)

on Python PyPI

🌟 Summary

Ultralytics 8.3.194 enhances model export stability and visibility, adds seamless CoreML NMS inference, and refactors telemetry for a cleaner, non-blocking experience—plus several quality-of-life fixes and docs updates. 🚀

📊 Key Changes

• Exporter dependency and logging (priority)
  • Bumped ONNX simplifier to onnxslim>=0.1.67 for ONNX and TensorFlow SavedModel exports and restored useful urllib3 logs while keeping sentry_sdk quiet. See PR ultralytics 8.3.194: bump onnxslim in Exporter. 🔧
• CoreML NMS support
  • You can now export YOLO models with NMS embedded and run inference directly in Python. See PR Support CoreML inference with NMS embedded models. 🍎
• TensorFlow export stability
  • Pinned TensorFlow and tf-keras during export: tensorflow>=2.0.0,<=2.19.0 and tf_keras<=2.19.0 to match onnx2tf requirements. See PR Pin tf_keras<=2.19.0. 🧩
• Telemetry refactor
  • Moved anonymous “Events” from Ultralytics HUB utils to a standalone ultralytics.utils.events module; sending runs in a background thread and no longer depends on HUB. See PR Refactor Events class. 🧠
• Segmentation robustness
  • Fixed cv2.resize channel limit (512) by splitting large mask tensors to prevent errors with many detections/classes. See PR Handle cv2.resize 512-channel limit. 🛡️
• SAM-2 API/doc consistency
  • Standardized predictor input to source and renamed img param to im in inference, updating examples accordingly. See PR Fix SAM2DynamicInteractivePredictor example. 📘
• Faster CI feedback
  • Reduced GPU CI timeout from 6 hours to 20 minutes to prevent runaway jobs. See PR Prevent long-running billed GPU CI. ⏱️
• Docs refresh
  • New Events reference page and updated Queue Management guide with a broader-use video. See Events utils reference and the updated Queue Management guide. 🎥

🎯 Purpose & Impact

• More reliable exports and fewer dependency conflicts
  • Newer onnxslim brings bug fixes and improved model simplification for both ONNX and TensorFlow exports. Restored urllib3 logs help troubleshoot network issues during export. ✅
• Smoother Apple workflows
  • Export with NMS and run CoreML models directly—no manual postprocessing required for supported detection use cases. Example:

    yolo export model=yolo11n.pt format=coreml nms=True
    yolo predict model=yolo11n.mlpackage

• Predictable TensorFlow SavedModel pipeline
  • Version pinning prevents breakages with newer TensorFlow releases; if you have TF > 2.19, you’ll be prompted to downgrade for export only. 🔒
• Cleaner, non-blocking telemetry
  • Events are now in ultralytics.utils.events and sent in the background—no HUB dependency and zero impact on training/export performance. If you import Events directly, update your import:

    from ultralytics.utils.events import events, Events

• Fewer runtime errors in dense segmentation
  • Avoids crashes from OpenCV’s 512-channel resize limit when handling many masks. 🧪
• Developer experience and CI efficiency
  • Faster feedback loops with tighter GPU timeouts; long tests may need adjustments. ⚡
• Minor API/docs alignment
  • SAM-2 examples now use source=... consistently, reducing confusion. 🧭

Enjoy improved stability and smoother workflows across exports, CoreML inference, and telemetry—with minimal breaking changes. 🎉

What's Changed

• Pin tf_keras<=2.19.0 by @Laughing-q in #21949
• Support CoreML inference with NMS embedded models by @Y-T-G in #21948
• fix: split large channel masks to handle cv2.resize 512 limitations by @ShuaiLYU in #21947
• Fix SAM2DynamicInteractivePredictor example in docs by @Y-T-G in #21955
• Prevent long-running billed GPU CI by @glenn-jocher in #21960
• Refactor Events class by @glenn-jocher in #21959
• Add https://youtu.be/Gxr9SpYPLh0 to docs by @RizwanMunawar in #21923
• ultralytics 8.3.194 Bump onnxslim>=0.1.67 in Exporter by @onuralpszr in #21951

Full Changelog: v8.3.193...v8.3.194