New Release: Ultralytics v8.3.226

[glenn-jocher]

Ultralytics v8.3.226 — Security-first release, safer parsing, custom Albumentations, and smoother NCNN export

Quick summary: v8.3.226 focuses on security and reliability by replacing unsafe eval() across the codebase, while adding Python API support for custom Albumentations, simplifying NCNN export via the PNNX Python interface, stabilizing multi-GPU training, and improving installer logs and docs. YOLO11 remains the latest stable and recommended Ultralytics model for all use cases.

Highlights

• Security-first: eval() fully replaced with safe parsing using ast.literal_eval() across configs, loaders, image size checks, and Triton utils.
• New: Custom Albumentations via a Python augmentations parameter in model.train(...).
• Simpler deployment: NCNN export now uses the PNNX Python package directly, with clearer logs.
• Stability: DDP final_epoch fix for val=False and safer default Matplotlib settings for analytics.
• UX and docs: Improved requirement checks and logs, link hygiene across docs, and a new one-click KITTI training notebook for YOLO11.

New Features

• Add custom Albumentations in Python with the new augmentations parameter; see the contribution in custom Albumentations in the Python API by @onuralpszr.
• NCNN export pipeline now calls the PNNX Python API directly for a simpler, more reliable workflow; details in Use PNNX Python interface for NCNN conversion by @Y-T-G.

Security & Parsing

• All eval() usage replaced with ast.literal_eval() for deterministic, safe parsing:
  • Config parsing via cfg.smart_value() is now secure.
  • Webcam sources like "0" are safely parsed as integers.
  • Image sizes like "[640, 640]" parse predictably.
  • Triton model metadata parsing is hardened.
• Track the main change in Replace eval() with ast.literal_eval() by @onuralpszr.

Improvements

• Training stability:
  • Fixed undefined final_epoch for DDP when val=False; see Fix undefined final_epoch for DDP when val=False by @Y-T-G.
  • Analytics now imports safe Matplotlib settings by default; covered in Add plt_settings import for Analytics by @onuralpszr.
• Installation & tooling:
  • More informative requirement checks and better behavior with the uv package manager as shown in Improve check_requirements() output with uv by @glenn-jocher.
  • CI link checks are less flaky with a longer retry window via Increase link check retry delay by @glenn-jocher.
• Docs & UX polish:
  • A ready-to-run notebook was added in Add KITTI notebook in docs by @RizwanMunawar.
  • Code and link hygiene improved in Python 3.9 target in Markdown and Python and URL/301 refactor, both by @glenn-jocher.

Bug Fixes

• Resolved rare crash in multi-GPU runs when val=False through a robust final_epoch definition as captured in Fix undefined final_epoch for DDP when val=False by @Y-T-G.
• Prevented CI issues with Matplotlib backend settings via Analytics plt_settings import fix by @onuralpszr.
• Ensured requirement checks render correctly under uv in Improve check_requirements() output with uv by @glenn-jocher.

Try it now

Upgrade to the latest version:

pip install -U ultralytics

Example: use custom Albumentations directly from Python

import albumentations as A
from ultralytics import YOLO

model = YOLO("yolo11n.pt")
custom_transforms = [
    A.Blur(blur_limit=7, p=0.5),
    A.CLAHE(clip_limit=4.0, p=0.5),
]
model.train(data="coco8.yaml", epochs=100, imgsz=640, augmentations=custom_transforms)

Resources

• Review the full list of changes in the release notes at Ultralytics v8.3.226.
• Explore every commit between versions in the full changelog diff from v8.3.225 to v8.3.226.
• Learn how to configure augmentations in the YOLO data augmentation guide.
• Get started with deployment using the NCNN export documentation.
• Train YOLO11 on KITTI in one click via the KITTI dataset guide with Colab quickstart.

What’s Changed (PRs and authors)

• Custom Albumentations in the Python API by @onuralpszr.
• Use PNNX Python interface for NCNN conversion by @Y-T-G.
• Analytics Matplotlib settings fix by @onuralpszr.
• Add KITTI notebook in docs by @RizwanMunawar.
• Increase CI link check retry delay by @glenn-jocher.
• Python 3.9 target in Markdown by @glenn-jocher.
• Fix undefined final_epoch for DDP when val=False by @Y-T-G.
• Python and URL/301 refactor by @glenn-jocher.
• Improve check_requirements() output with uv by @glenn-jocher.
• Replace eval() with ast.literal_eval() for security by @onuralpszr.

Thanks and feedback

Huge thanks to the YOLO community and Ultralytics team for the contributions in v8.3.226. Please upgrade, try the new features, and share your experience or questions in our discussions. Your feedback helps us prioritize and improve future releases!