Ultralytics YOLO

Read full text
Authors: Glenn Jocher, Jing Qiu, Ayush Chaurasia
Cited by: 0

Abstract

🌟 Summary (single-line synopsis) Ray Tune + Weights & Biases (W&B) tuning runs now get unique, trial-specific names in Ultralytics 8.3.247, making hyperparameter sweeps much easier to track 🧪📊 📊 Key Changes (Priority / Current PR #23084) Ray Tune trials logged to W&B no longer all show up as train 🏷️ Saves the original run name before it gets removed from training args Uses Ray Tune's trial_id to append a unique suffix Naming format: {base_name}_{trial_suffix} (example: my_run_00000) ✅ Platform/HUB training upload improvements (PR #23105) 🛰️ Deduplicates plots by type (avoids repeated uploads/noisy results pages) Downsamples large curve payloads (e.g., PR curves) before upload to reduce size and speed up logging 📉 Cleaner console logging with a clearer "Platform:" prefix and a direct "view results" URL 🔗 Fewer unnecessary GPU-monitoring warnings (especially on macOS / when CUDA isn't available) 🧯 Validation plotting stability fixes (PR #23112, #23093) 🖼️ Prevents crashes when predictions are empty (if not preds: return) Removes a redundant bbox slice during bbox format conversion (small cleanup, safer assumptions) Export docs: output tensor formats (PR #23057) 📦 Adds a clear FAQ section explaining typical output tensor structures for detect/seg/pose exports (ONNX/TensorRT), with pointers to examples Docs & code clarity cleanups (PR #23107) 🧹 Fixes docstring examples (e.g., Boxes.xywh) and clarifies tensor dimension naming in transformer code 🎯 Purpose & Impact Much clearer hyperparameter tuning in W&B 🎛️ You can now immediately distinguish Ray Tune trials in W&B, matching the local run directory structure—no more "everything is named train" confusion. More reliable & efficient experiment logging 🚀 Plot deduplication + curve downsampling reduces upload noise and payload size, improving the experience when viewing runs in Ultralytics HUB/Platform. Fewer training/validation edge-case failures 🧩 The empty-prediction guard prevents rare but frustrating validation plotting crashes. Easier deployment integrations 🔧 Better export tensor documentation helps anyone writing custom ONNX/TensorRT post-processing pipelines. What's Changed Fix plots logging by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23105 Add empty preds guard in DetectionValidator.plot_predictions by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23112 Remove redundant bbox slice in plot_predictions by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23093 Add documentation for output tensor format by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23057 Fix documentation and improve code clarity by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23107 ultralytics 8.3.247 Improve Ray Tune trial names logged to W&B by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23084 Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.246...v8.3.247

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
3 January 2026