New Release: Ultralytics v8.3.246

Discussion
, ,
1

:rocket: Ultralytics v8.3.246 is out!

TL;DR: Ultralytics v8.3.246 improves end-of-training reporting by uploading rich, interactive plot data + class names so results are easier to explore and interpret :chart_increasing::label::sparkles:

You can review the release details directly in the Ultralytics v8.3.246 release page, and browse the complete diff in the v8.3.245 → v8.3.246 changelog comparison.

:sparkles: New Features (Highlights)

:bar_chart: Rich plot-data upload on train completion (priority)

At on_train_end(), Ultralytics now collects plots from both the trainer and validator and uploads them with the final "training_complete" event—making dashboards and reports far more informative :puzzle_piece::outbox_tray:

Includes data for:

  • Confusion matrix
  • PR curves
  • Metric–confidence curves

Implemented in PR #23104 by @glenn-jocher.

:label: Class names included in the final payload

The "training_complete" message now includes classNames, so plots/results can be labeled correctly (no more guessing label mappings) :white_check_mark:

Added in PR #23104 by @glenn-jocher.

:desktop_computer: Structured plot metadata for interactive visualizations

Plot callbacks (on_plot) now provide raw arrays + plot type, enabling interactive rendering:

  • Confusion matrix → raw matrix counts
  • PR curve → x/y arrays + per-class AP
  • Metric-confidence → x/y arrays

Shipped in PR #23104 by @glenn-jocher.

:hammer_and_wrench: Improvements

:toolbox: Better diagnostics when installs fail

check_requirements() now surfaces more detail (including captured command output), making dependency issues easier to debug—especially in CI/containers :warning:

From PR #23097 by @Y-T-G.

:desktop_computer::wood: NVML init failures are now logged

If GPU telemetry can’t initialize, Ultralytics prints a warning instead of failing silently.

From PR #23096 by @glenn-jocher.

:package: Export dependency updates & safeguards

  • Bump onnxslim minimum to >=0.1.82 to reduce export tooling issues via PR #23101 by @onuralpszr
  • ExecuTorch export adds a guard ruamel.yaml<0.19.0 to prevent version-related failures via PR #23102 by @onuralpszr

:spouting_whale: Dockerfile example updates

Documentation/usage improvements in PR #23103 by @glenn-jocher.

:lady_beetle: Bug Fixes / Correctness

:white_check_mark: Try it now

Upgrade with:

pip install -U ultralytics

Then run a quick check:

yolo predict model=yolo11n.pt source=bus.jpg

:speech_balloon: Feedback welcome

If you notice anything off with plot uploads, exports, or environment checks, please share details (logs + OS/Python/PyTorch + command used). Community reports help us iterate quickly and keep Ultralytics YOLO tooling solid for everyone :folded_hands: