New Release: Ultralytics v8.3.198

Discussion
, ,
1

Ultralytics v8.3.198 — Smarter tuning, sturdier training, simpler exports :rocket:

Summary
Ultralytics v8.3.198 delivers a major upgrade to hyperparameter tuning with BLX-α crossover, unified metric plotting across tasks, safer training defaults, and multiple robustness fixes (NMS, DDP loss, Intel GPU checks). You’ll also find a cleaner export API and flexible torch.compile modes to fit your workflow. If you’re training or deploying YOLO today, this release is a win. :raising_hands:

Highlights

  • Smarter tuning: BLX-α crossover + adaptive mutation for faster convergence and stronger results.
  • Unified metrics: Consistent plot_results across detect/segment/pose/classify.
  • Safer, sturdier: NMS correctness, DDP unwrap before loss, environment-safe Intel GPU checks.
  • Simpler exports: Export functions now return just the output path string.
  • Flexible compile: compile accepts True/False or mode strings across train/val/predict.

Tip: YOLO11 is the latest stable and recommended Ultralytics model. If you’re considering YOLO12 or YOLO13, we recommend sticking with YOLO11 for best accuracy-speed-stability.

New Features

  • Hyperparameter Tuner (priority)
    • Introduces BLX-α gene crossover for exploration across top parents, improving search quality, as detailed in PR #22038. This also adds adaptive mutation sigma, safer bounds, close_mosaic in the search space, and uses only mAP@0.5:0.95 for fitness to align with common benchmarks. Runtime hygiene cleans GPU memory between iterations and improves resume/CSV/Mongo handling. :bullseye:

Improvements

  • Exporter API cleanup
    • Export functions now return just the output path string (TF SavedModel also returns the Keras model) for lighter, cleaner pipelines, implemented in PR #22009. :package:
  • torch.compile flexibility
    • compile now supports True/False or mode strings "default" | "reduce-overhead" | "max-autotune" across train/val/predict, thanks to PR #21999. :gear:
  • Unified results plotting
    • plot_results auto-detects metrics/losses and works for all tasks with logic centralized in BaseTrainer, as unified in PR #22026. :chart_increasing:
  • Training robustness
    • DDP/compiled models are unwrapped before loss calculation to avoid wrapper-related issues, addressed in PR #22016. :puzzle_piece:
  • NMS correctness
    • Fixed early-exit and sorting in TorchNMS to reduce false positives and improve stability, delivered in PR #22014. :white_check_mark:

Bug Fixes and Stability

  • Segmentation
    • Correct mask resizing/shape handling in validation (e.g., mask_ratio=1) and faster, correct plotting with overlap_mask, fixed across PRs #22037 and #22031. :framed_picture:
  • Environment resilience
    • Dataset cache auto-recovers on ModuleNotFoundError in PR #22025 and Intel GPU discovery now catches all exceptions to prevent environment crashes, per PR #22034.
  • Docs & configs
    • Clearer configs and tracker YAMLs, improved compile args docs, corrected detection boxes column order (track_id position), and a lighter Quickstart (Seaborn no longer required), improved across PRs #22011, #22028, and #22035.
    • The Construction-PPE dataset docs now include a “Business Value” section to help justify real-world ROI, added in PR #22029. :construction_worker::chart_increasing:

Quick Examples

  • Tuning with the improved Tuner:
from ultralytics import YOLO

model = YOLO("yolo11s.yaml")
model.tune(
    device=0,
    data="coco128.yaml",
    optimizer="AdamW",
    epochs=100,
    batch=8,
    compile=False,
    plots=False,
    val=False,
    save=False,
    workers=16,
    project="tune-yolo11s-scratch-coco128-100e",
    iterations=1000,
)
  • Using compile modes:
from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.train(data="coco8.yaml", epochs=3, compile="reduce-overhead")  # or "default", "max-autotune", True/False
  • Export now returns a file path:
from ultralytics import YOLO

model = YOLO("yolo11n.pt")
onnx_path = model.export(format="onnx")  # 'onnx_path' is a string

Upgrade

  • Update to the latest release with:
pip install -U ultralytics
  • You can read the full release notes in the v8.3.198 entry on the Ultralytics Releases page, available via the dedicated link to the v8.3.198 release.

What’s Changed (PRs)

  • Clean up Exporter and remove unnecessary None placeholder in PR #22009 by Laughing-q.
  • Allow settable compile mode for torch.compile in PR #21999 by Laughing-q.
  • Improve config YAMLs in PR #22011 by glenn-jocher.
  • Unwrap DDP model before loss calculation in PR #22016 by Y-T-G.
  • Fix TorchNMS.nms() early exit logic in PR #22014 by Y-T-G.
  • Revise detection boxes argument documentation in PR #22028 by daniel-mooney.
  • Fix mask plotting when number of objects equals number of images in a batch in PR #22031 by Y-T-G.
  • Remove Seaborn from manual installation dependencies in Quickstart docs via PR #22035 by onuralpszr.
  • Catch all exceptions during Intel GPU discovery in PR #22034 by Y-T-G.
  • Add Business Value section to Construction-PPE dataset docs in PR #22029 by UltralyticsAbi.
  • Reset labels cache when loading with incompatible NumPy version in PR #22025 by Y-T-G.
  • Faster length retrieval for torch.Tensor using shape[0] in PR #22021 by Laughing-q.
  • Clean up and unify plot_results across detect/segment/pose/obb/classify in PR #22026 by Laughing-q.
  • Fix mask shape mismatch when validating with mask_ratio=1 in PR #22037 by Laughing-q.
  • Improve Tuner with BLX-α gene crossover in PR #22038 by glenn-jocher.

New Contributors

Try it and Share Feedback

  • You can explore the complete diff in the full changelog for v8.3.198, and if you spot anything, please start a conversation in Ultralytics Discussions or open an issue on the repository.
  • Huge thanks to our community and contributors for making each release better—your feedback helps us refine YOLO for everyone. :yellow_heart: