ultralytics 8.3.193

v8.3.193 - `ultralytics 8.3.193` Preload `torchvision` for long-sequence inference (#21926)

on Python PyPI

🌟 Summary

Ultralytics 8.3.193 boosts long-video and large-batch inference performance by auto-enabling TorchVision NMS, streamlines checkpoint loading with a single API, and delivers cleaner, more informative progress bars and logging. 🚀

📊 Key Changes

• Long-sequence inference optimization (PR #21926) ⚡
  • Detects long sequences (streams, screenshots, >1000 images, or any video) and preloads torchvision to trigger TorchVision-based NMS automatically.
  • Keeps warnings targeted: only warns when not in stream mode.
• Unified checkpoint loading API (PR #21933) 🔄
  • Replaces attempt_load_one_weight/attempt_load_weights with a single load_checkpoint function across the codebase.
  • Docs updated to reference ultralytics.nn.tasks.load_checkpoint.
  • Minimal example:

    from ultralytics.nn.tasks import load_checkpoint
    model, ckpt = load_checkpoint("yolo11n.pt", device="cpu", fuse=False)

• TQDM progress bar improvements (PRs #21944, #21934, #21932) 📈
  • Faster, simpler formatting via f-strings, smarter/smoother rate calculations.
  • Human-friendly rate scaling by default (K/M/G for all units), plus “G” tier for very high rates.
  • Cleaner completion display for byte downloads (e.g., “12.3 MB” instead of “12.3/12.3 MB”).
  • Note: output strings changed; update any scripts that parse exact progress text.
• Clearer save paths (PR #21931) 🗂️
  • get_save_dir() now returns absolute paths for unambiguous run locations in logs.
• Stable exports via dependency pin (PR #21941) 🔒
  • Pins onnxslim==0.1.65 for ONNX and TF SavedModel exports to avoid unexpected breakages.
• Robust PyTorch version handling (PR #21921) 🧰
  • Normalizes TORCH_VERSION usage to prevent issues with non-string version objects and align all version checks/logging.
• Better segmentation validation (PR #21503) 🧪
  • Corrects mask preparation for overlap_mask=True and simplifies IoU computation for more accurate, efficient val results.
• Dataset and docs fixes (PRs #21939, #21942, #21929, #21930) 📚
  • Fixes autosplit import in xView.yaml (ultralytics.data.split.autosplit).
  • Clarifies recommended dataset folder structure in docs.
  • Updates docs CI to target Python 3.9; improves a label error message to mention negative coordinates.

🎯 Purpose & Impact

• Faster, more reliable inference on long videos/large image sets with automatic TorchVision NMS preload—no user action required. 🎬⚡
• Simpler model loading for developers via a single load_checkpoint entry point; fewer APIs to remember. 🧠
• Cleaner, more readable progress bars and absolute save paths improve day-to-day usability and logs across CLI and Ultralytics HUB. 🧭
• More stable export workflows thanks to dependency pinning; fewer surprise failures in ONNX/TF pipelines. 📦
• Improved compatibility and diagnostics across PyTorch versions; fewer version-related edge cases. 🛡️
• More accurate segmentation validation and smoother dataset setup with fixed imports and clarified docs. ✅

Tip: If any tooling relies on exact progress bar strings, review and adjust parsers due to the new TQDM formatting. ✍️

What's Changed

• Update docs.yml for Python 3.9 ruff format target by @glenn-jocher in #21929
• Display resolved save_dir by @glenn-jocher in #21931
• Remove TQDM duplicate filesize info on completion by @glenn-jocher in #21932
• TQDM smart rate scaling by @glenn-jocher in #21934
• Clarify negative coordinate assertion by @samrere in #21930
• Pin onnxslim==0.1.65 by @Laughing-q in #21941
• Fix autosplit import in xView.yaml by @Laughing-q in #21939
• Fix torch._version_ check for Torch>1.9 by @RizwanMunawar in #21921
• Scope batch mask process to _prepare_batch by @Y-T-G in #21503
• Consolidate checkpoint loading functions by @glenn-jocher in #21933
• Faster isinstance() checks by @glenn-jocher in #21928
• Fix dataset directory structure information in docs by @RizwanMunawar in #21942
• TQDM f-strings for speed by @glenn-jocher in #21944
• ultralytics 8.3.193 Preload torchvision for long-sequence inference by @Laughing-q in #21926

New Contributors

• @samrere made their first contribution in #21930

Full Changelog: v8.3.192...v8.3.193