ultralytics 8.3.198

v8.3.198 - `ultralytics 8.3.198` Improve `Tuner` with BLX-α gene crossover (#22038)

on Python PyPI

🌟 Summary

Smarter hyperparameter tuning and sturdier training/inference across the board. v8.3.198 introduces a much stronger Tuner with BLX-α crossover, unified metric plotting, safer defaults, and multiple robustness fixes (NMS, DDP loss, Intel GPU checks), plus simpler export APIs and flexible torch.compile modes. 🚀

📊 Key Changes

• Hyperparameter Tuner (priority)
  • BLX-α gene crossover for smarter exploration across top parents, not just single-parent mutation. See PR #22038 🔬
  • Adaptive mutation sigma (decays early for stable convergence), safer mutation bounds, and added close_mosaic to the search space.
  • Search space tweaks (e.g., lower bound for cls gain → 0.1) and consistent type casting.
  • Fitness now uses only mAP@0.5:0.95 for selection, aligning with common benchmarks. 🎯
  • Runtime hygiene: GPU memory cleanup between iterations to reduce OOM/fragmentation.
  • More robust resume logic and CSV/MongoDB handling; cleaner best-iteration management.
• Exporter API cleanup
  • Export functions now return just the output path string (except TF SavedModel which also returns the Keras model). Simpler and lighter. See PR #22009 📦
• torch.compile flexibility
  • compile accepts True/False or a mode string "default" | "reduce-overhead" | "max-autotune" across train/val/predict. See PR #21999 ⚙️
• Training robustness
  • Unwrap DDP/compiled models before loss calculation to avoid wrapper-related issues. See PR #22016 🧩
• NMS correctness
  • Fixed early-exit and sorting in TorchNMS to reduce false positives and improve stability. See PR #22014 ✅
• Unified results plotting
  • plot_results auto-detects metrics/losses and works for all tasks; logic centralized in BaseTrainer. See PR #22026 📈
• Segmentation fixes
  • Correct mask resizing/shape handling in validation (e.g., mask_ratio=1) and faster, correct mask plotting with overlap_mask. See PR #22037, PR #22031 🖼️
• Stability & docs
  • Dataset cache auto-recovers on ModuleNotFoundError. See PR #22025
  • Intel GPU discovery now catches all exceptions to prevent environment crashes. See PR #22034
  • Clearer configs and tracker YAMLs; improved docs for compile args; corrected detection boxes column order (track_id position); quickstart removes Seaborn from manual install. PRs #22011, #22028, #22035
  • Construction-PPE dataset docs add a “Business Value” section for real-world ROI. See PR #22029 👷📈

🎯 Purpose & Impact

• Better models, faster: BLX-α crossover + adaptive mutation helps find stronger hyperparameters in fewer iterations, especially for YOLO11 training. 🚀
• More reliable training: Unwrapping before loss and flexible compile modes reduce edge-case failures with DDP and torch.compile.
• Fewer false positives: NMS logic fixes stabilize detections across datasets.
• Cleaner APIs: Exporters returning a path simplifies pipelines, reduces memory, and avoids unexpected large return objects.
• Easier metrics analysis: Unified plotting works consistently across detect/segment/pose/classify; less duplication, clearer visuals.
• Smoother ops: GPU memory cleanup between tuning iterations, robust dataset cache handling, and environment-safe Intel GPU checks reduce friction in diverse setups.
• Clarity for everyone: Improved YAML/docs, corrected detection box order, and streamlined Quickstart make onboarding and maintenance simpler.

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

What's Changed

• Clean up Exporter and remove unnecessary None placeholder by @Laughing-q in #22009
• Allow settable compile mode for torch.compile by @Laughing-q in #21999
• Improve config YAMLs by @glenn-jocher in #22011
• Unwrap DDP model before loss calculation by @Y-T-G in #22016
• Fix TorchNMS.nms() early exit logic by @Y-T-G in #22014
• Revise detection boxes argument documentation by @daniel-mooney in #22028
• Fix mask plotting when number of objects equal to number of images in a batch by @Y-T-G in #22031
• docs: 📝 remove seaborn package from manual installation dependencies list from quickstart docs by @onuralpszr in #22035
• Catch all exceptions during Intel GPU discovery by @Y-T-G in #22034
• Add business value section to Construction-PPE dataset by @UltralyticsAbi in #22029
• Reset labels cache when loading with incompatible numpy version by @Y-T-G in #22025
• Faster torch.Tensor.shape[0] to get length for torch.Tensor by @Laughing-q in #22021
• Clean up and unify plot_results for detect/segment/pose/obb/classify tasks by @Laughing-q in #22026
• Fix mask shape mismatch when validating with mask_ratio=1 by @Laughing-q in #22037
• ultralytics 8.3.198 Improve Tuner with BLX-α gene crossover by @glenn-jocher in #22038

New Contributors

• @daniel-mooney made their first contribution in #22028
• @UltralyticsAbi made their first contribution in #22029

Full Changelog: v8.3.197...v8.3.198