New Release: Ultralytics v8.3.247

:rocket: Ultralytics v8.3.247 is out! :test_tube::bar_chart:

TL;DR: Ray Tune + Weights & Biases (W&B) hyperparameter sweeps now get unique, trial-specific run names, making tuning runs dramatically easier to track (no more “everything is train”) :white_check_mark:


:glowing_star: New Features (Highlights)

:label: Ray Tune trials now log uniquely to W&B

Ray Tune trials logged to W&B no longer all show up as train—each trial gets a clear, unique name like my_run_00000.

This is handled in PR #23084 by @raimbekovm:

  • Preserves the original base run name before it’s removed from training args
  • Appends a Ray Tune trial_id suffix for uniqueness
  • Naming format: {base_name}_{trial_suffix} :white_check_mark:

Why it matters: your W&B dashboard now mirrors your local Ray Tune trial structure, making sweeps much easier to interpret and compare :control_knobs:


:satellite: Improvements

:chart_decreasing: Cleaner, faster training uploads to Ultralytics Platform

Training upload & logging quality improvements in PR #23105 by @glenn-jocher:

  • Deduplicates plots by type (avoids repeated uploads / noisy results)
  • Downsamps large curve payloads (e.g., PR curves) to reduce upload size and speed up logging
  • Cleaner console output with a clearer Platform: prefix + a direct “view results” URL
  • Fewer unnecessary GPU-monitoring warnings (notably on macOS / when CUDA isn’t available) :fire_extinguisher:

:puzzle_piece: Bug Fixes & Stability

:framed_picture: Validation plotting stability fixes

Two small-but-important fixes to prevent rare validation plotting issues:

  • Empty prediction guard to prevent crashes in DetectionValidator.plot_predictions via PR #23112 by @raimbekovm
  • Removed a redundant bbox slice during bbox format conversion via PR #23093 by @raimbekovm

:package: Docs Updates (Deployment + Clarity)

:wrench: Export docs: output tensor formats (ONNX/TensorRT)

A new FAQ-style section clarifies typical output tensor structures for detect/seg/pose exports—super helpful if you’re writing custom post-processing.

See PR #23057 by @raimbekovm.

:broom: General documentation & code clarity cleanups

Docstring + clarity improvements (including Boxes.xywh examples and clearer tensor dimension naming) in PR #23107 by @raimbekovm.


:white_check_mark: How to Update

pip install -U ultralytics

Then run your usual training/tuning workflows—especially Ray Tune + W&B sweeps—to see the improved run naming in action :test_tube:


:magnifying_glass_tilted_right: Changelog + Release Links

You can review the full diff in the GitHub v8.3.246…v8.3.247 comparison, and details in the v8.3.247 release page.


:speech_balloon: Feedback Welcome!

Give v8.3.247 a spin and let us know how it goes—especially if you’re running Ray Tune sweeps or exporting to ONNX/TensorRT. If anything looks off, please share logs + minimal repro steps so we can jump on it quickly :raising_hands: