ultralytics 8.3.200

v8.3.200 - `ultralytics 8.3.200` Refactor Sony IMX exports to `imx.py` (#22005)

on Python PyPI

🌟 Summary

A focused export refactor that adds a complete Sony IMX500 pipeline, cleans up IMX-specific logic in model heads, and introduces clearer export APIs—plus stability fixes for ONNX (OBB), TensorRT exports, and smoother multi-GPU training. 🚀

📊 Key Changes

• Export refactor and new IMX pipeline (PR #22005) ✨

  • New, clearer export APIs:
    • export_onnx → torch2onnx
    • export_engine → onnx2engine
  • New Sony IMX export:
    • torch2imx utility with quantization and IMX-friendly NMS wrapping
    • IMX-specific helpers (FXModel, Detect/Pose overrides, NMSWrapper)
  • Export utilities reorganized to ultralytics/utils/export/ with imx.py
  • Detect/Pose heads simplified by removing IMX-only branches
  • Removed FXModel from torch_utils
  • Docs updated to reflect the new structure

• IMX support clarity (PR #22082) 🧭

  • Error messages now explicitly state IMX exports support YOLOv8n and YOLO11n (detection only in benchmarks).

• ONNX + OBB reliability (PR #22079) 🧩

  • Fixes a RuntimeError for OBB models exported with NMS on CUDA by safeguarding empty-mask cases.

• TensorRT export stability (PR #22080) ⚙️

  • Corrects dynamic shape handling and re-enables TensorRT tests
  • Extends GPU CI timeout from 20 → 60 minutes for reliability

• DDP training improvements (PR #22073) 💼

  • Initializes world_size/ddp earlier, simplifies DDP APIs, and prevents duplicate wandb logs

• Tuner robustness (PR #22066) 🛡️

  • Avoids KeyError when close_mosaic isn’t in the hyperparameter search space

• UX polish (PRs #22071, #22083, #22046, #22063, #22085) ✍️

  • Aligns multi-GPU device printing
  • Clearer export error when trying to “export” to pt
  • Fixes a broken DeepStream link for Jetson users
  • Documentation typo and formatting cleanups
  • Updates docs authors metadata

🎯 Purpose & Impact

• Cleaner, more intuitive export APIs 🧭
  • The torch2onnx and onnx2engine names make conversion stages explicit and reduce confusion.
• First-class Sony IMX exports 📦
  • New torch2imx enables end-to-end IMX500 deployment with built-in quantization and NMS wrapping.
  • Clearer docs and package layout make maintenance and contributions easier.
• More reliable exports across backends ✅
  • ONNX OBB fix prevents runtime failures on CUDA.
  • TensorRT export is more dependable, with re-enabled coverage to catch regressions earlier.
• Smoother multi-GPU training 🧠
  • Early DDP setup avoids double logging and simplifies internals, improving stability for distributed runs.
• Fewer paper cuts for users ✂️
  • Robust tuner behavior, better error messages, aligned device output, and updated docs links all reduce friction.

Quick tips:

• Export to ONNX:
  • CLI: yolo export model=yolo11n.pt format=onnx
• Export to TensorRT:
  • CLI: yolo export model=yolo11n.pt format=engine dynamic
• Export to Sony IMX:
  • CLI: yolo export model=yolo11n.pt format=imx
  • Note: IMX benchmarks currently enforce detection-only and support YOLOv8n/YOLO11n.

What's Changed

• Fix KeyError when close_mosaic not part of search space by @Y-T-G in #22066
• Update yolo-data-augmentation.md by @miles-deans-ultralytics in #22063
• Update deepstream-nvidia-jetson.md by @miles-deans-ultralytics in #22046
• Revert "Re-enable TensorRT export in GPU tests" by @glenn-jocher in #22078
• Fix DDP device spacing by @glenn-jocher in #22071
• Add Onuralp and Miles to MkDocs GitHub authors by @glenn-jocher in #22085
• Improve export error message for pt format by @RizwanMunawar in #22083
• Update IMX export assert message to include YOLO11n by @RizwanMunawar in #22082
• Fix ONNX RuntimeError for OBB models exported with NMS by @Y-T-G in #22079
• Fix TensorRT dynamic shape calculation by @Y-T-G in #22080
• Fix wandb logging twice when running DDP training by @Laughing-q in #22073
• ultralytics 8.3.200 Refactor Sony IMX exports to imx.py by @Laughing-q in #22005

New Contributors

• @miles-deans-ultralytics made their first contribution in #22063

Full Changelog: v8.3.199...v8.3.200