ultralytics 8.3.190

v8.3.190 - `ultralytics 8.3.190` Autobackend `torch`-native NMS (#21862)

on Python PyPI

🌟 Summary

ultralytics 8.3.190 delivers a faster, torchvision-free NMS pipeline powered by a new PyTorch-native implementation optimized for YOLO11, plus stability and UX upgrades across CoreML, progress bars, macOS handling, and reproducibility. 🚀

📊 Key Changes

• Autobackend torch-native NMS (most important)

  • Introduces a new NMS module ultralytics/utils/nms.py with:
    • TorchNMS.nms (exact-match, optimized NMS) and TorchNMS.fast_nms (vectorized FastNMS)
    • non_max_suppression refactored here; rotated NMS supported via fast_nms + batch_probiou
  • Removes hard dependency on torchvision.ops.nms for prediction-time use
  • YOLO11-optimized, works on CPU and GPU, improves post-processing latency ⚡
  • Updated imports across the codebase (predict, val, examples) to from ultralytics.utils import nms
  • OBB flows and exporter now use TorchNMS.fast_nms with batch_probiou for rotated boxes

• AutoBackend API cleanup

  • Renamed parameter weights= to model= for consistency and to avoid duplicate memory usage (@glenn-jocher)
  • Unified handling for in-memory torch.nn.Module and .pt loading; gradients disabled by default for safer, faster inference
  • Adds clear CoreML logging and a version gate

• CoreML stability

  • Enforces coremltools>=8.0 for CoreML inference to prevent runtime issues; shows a helpful message before loading

• Performance and stability

  • Reintroduces platform-safe in-place ops (sigmoid_(), clamp_()) for non-macOS 14 systems to boost speed
  • Adds MACOS14/NOT_MACOS14 flags to handle Apple MPS quirks safely
  • Faster internet checks using DNS lookups, improving startup reliability

• Import-time and CI improvements

  • Scopes heavy imports (requests/psutil) to runtime, reducing import overhead and avoiding circular deps, including Ultralytics HUB utils
  • CI workflows now run for PRs targeting any branch, not just main

• Geometric ops robustness

  • Refactors core ops to better handle both HW and HWC shapes, clarifies padding logic, and reduces redundant indexing

• Reproducibility

  • Trainer saves git_commit in checkpoints; new utils.get_git_commit() with caching

• UX and docs

  • TQDM progress bar now shows ETA, stable rate, cleaner units, and improved final-state display ⏳
  • Dataset/scripts/docs switch to subprocess.run(...) for safer command execution
  • Tutorial notebook updated for latest versions and simpler dataset download APIs

🎯 Purpose & Impact

• Faster, leaner inference 🏃‍♂️💨

  • PyTorch-native NMS reduces reliance on torchvision, speeds up post-processing, and lowers latency, especially helpful in deployment and environments where torchvision is heavy or unavailable.

• Better platform reliability 🛡️

  • macOS 14 handling prevents MPS-related crashes with in-place ops.
  • CoreML gating avoids version pitfalls and smooths model execution on Apple devices.

• Improved developer experience 👩‍💻

  • Cleaner AutoBackend(model=...) API, unified PyTorch paths, and no-grad defaults reduce boilerplate and risks.
  • Progress bar is more informative with ETA and better formatting.

• Stronger reproducibility and CI coverage 📦

  • Git commit embedded in checkpoints makes experiments traceable.
  • CI runs for all PR branches, speeding up feedback across workflows.

• Faster startup and “just works” behavior ⚙️

  • Scoped imports and DNS-based connectivity checks improve responsiveness and robustness.

Quick usage examples:

• Use the new NMS utilities:

  from ultralytics.utils import nms
  
  # preds: [B, C+4(+M), N]
  out = nms.non_max_suppression(preds, conf_thres=0.25, iou_thres=0.45)

• Updated AutoBackend parameter:

  from ultralytics.nn.autobackend import AutoBackend
  
  model = AutoBackend(model="yolo11n.pt", device="cuda")

• Ensure CoreML tools are up-to-date:

  pip install -U "coremltools>=8.0"
  

What's Changed

• Add coremltools>=8.0 check to Autobackend by @glenn-jocher in #21861
• Update TQDM progress bar to display remaining time by @Laughing-q in #21868
• Notebook updates by @glenn-jocher in #21870
• Fix prediction logging info for obb by @Laughing-q in #21871
• Scope requests imports by @glenn-jocher in #21876
• Target CI to all PR branches by @glenn-jocher in #21879
• Eliminate duplicate ops indexing by @glenn-jocher in #21877
• MACOS14 check by @glenn-jocher in #21883
• Rename Autobackend() weights to model by @glenn-jocher in #21881
• Reintroduce inplace .clamp_() and .sigmoid_() ops by @glenn-jocher in #21878
• Add git commit hash to checkpoints by @Bovey0809 in #21836
• Standardize docs build subprocess calls by @glenn-jocher in #21887
• Use subprocess.run() for improved dataset autodownloads by @glenn-jocher in #21888
• Faster online check using DNS resolution by @glenn-jocher in #21886
• ultralytics 8.3.190 Autobackend torch-native NMS by @glenn-jocher in #21862

New Contributors

• @Bovey0809 made their first contribution in #21836

Full Changelog: v8.3.189...v8.3.190