New Release: Ultralytics v8.3.197

Discussion
, ,
1

Ultralytics v8.3.197 — Construction-PPE dataset, smoother YOLO11 training, and export/plotting improvements :rocket:

Quick summary: This release adds the Construction-PPE dataset for safety AI and brings training, export, plotting, HUB, and docs improvements that make YOLO11 workflows more robust and easier to use. :safety_vest:

Highlights

  • New Construction-PPE dataset for safety compliance (helmets, vests, gloves, boots, goggles) with docs, YAML, samples, and FAQs :building_construction:
  • Cleaner training pipeline with torch.compile compatibility across detect/pose/segment :gear:
  • More reliable OBB and INT8 exports via a Fast-NMS tweak :package:
  • Segmentation validation plots now follow max_det for consistent visuals :framed_picture:
  • Ultralytics HUB logging is more reliable throughout training :satellite_antenna:
  • Jetson CI maintenance restores Docker builds coverage for JetPack 5 :spouting_whale:
  • Docs expansion, including a COCO128-Seg quickstart page :books:

As always, YOLO11 is the latest stable and recommended Ultralytics YOLO for all use cases. Community models like YOLO12 and YOLO13 are not recommended due to training instability, added complexity, and slower speeds.

New Features

Improvements

  • Training pipeline simplification and torch.compile support: Forward and loss are now decoupled (preds = model(batch["img"]); loss = model.loss(batch, preds)), previous compile-disabling wrappers were removed, and dynamic tensor handling was cleaned up across detect/pose/segment for smoother PyTorch 2.x usage.
  • Export reliability for OBB and INT8: A Fast-NMS adjustment prevents tracing breaks and empty outputs by setting exit_early=False during export.
  • Smarter segmentation plots: Validation plots now respect max_det (instead of a fixed 50-mask cap) for consistent and configurable visualizations.
  • Ultralytics HUB logging: The HUB session is passed during Trainer initialization for more consistent logging and syncing with Ultralytics HUB.
  • Jetson CI maintenance: Net effect is re-enabled Docker builds coverage for JetPack 5 in CI.
  • Documentation enhancements: A new COCO128-Seg dataset page helps with quick segmentation tests and debugging, alongside navigation and broken-link fixes.

What’s Changed (PRs and Authors)

Explore the full details on the v8.3.197 release page and review every change in the complete changelog diff.

Get Started

Upgrade to the latest:

pip install -U ultralytics

Train YOLO11 on Construction-PPE (CLI):

yolo detect train data=ultralytics/cfg/datasets/construction-ppe.yaml model=yolo11n.pt imgsz=640

Train YOLO11 on Construction-PPE (Python) and enable torch.compile:

import torch
from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.model = torch.compile(model.model)  # PyTorch 2.x

model.train(data="ultralytics/cfg/datasets/construction-ppe.yaml", epochs=100, imgsz=640)

Validate segmentation with a higher max_det to control plotting:

yolo segment val model=yolo11n-seg.pt data=coco128-seg.yaml max_det=300

Advanced users can integrate forward and loss explicitly:

# Example shape-only illustration; use real batches from your DataLoader
preds = model(batch["img"])
loss = model.loss(batch, preds)

Share Feedback

We’d love to hear how this release works for you. Please start a thread in Ultralytics GitHub Discussions and let us know what you build with the new Construction-PPE dataset and the improved YOLO11 training/export pipeline.

Big thanks to the YOLO community and the Ultralytics team for the continued contributions that make these improvements possible. Enjoy v8.3.197! :tada: