Ultralytics v8.3.129 Release: Smarter Augmentation, Smoother Exports, and Enhanced Docs!
Summary
We’re excited to announce Ultralytics v8.3.129, a release focused on intelligent data augmentation, stronger export and inference reliability, and clearer documentation. These upgrades make training and deploying your models with Ultralytics smoother and more intuitive than ever before!
Key Changes
New Features
-
Automatic Mosaic Augmentation Selection
Now, Ultralytics will automatically determine the best image selection method for mosaic data augmentation depending on your dataset’s cache settings. This takes the guesswork out of configuration and ensures consistent results!
ultralytics 8.3.129 full dataset buffer with cache=“ram” by @Y-T-G -
TensorBoard Setup Made Easy
Introducing clear command-line instructions for enabling or disabling TensorBoard logging, so you can manage experiment tracking with ease.
Add instructions to enable TensorBoard by @Y-T-G
Improvements
-
TensorRT Export Enhancement
The default maximum input size (max_shape) for exported TensorRT models has been increased, making it easier to run inference on bigger images without hitting errors.
TensorRT: Bump default dynamic max_shape up to 1280 by @Laughing-q -
Robust DLA Core Handling
Upgrades in how DLA core settings are read from model metadata ensure more reliable loading of TensorRT engines under diverse configurations.
Scope getting dla from metadata by @Laughing-q -
Better File Validation
File checking functions are now more flexible, supporting bothPathobjects and string types for file paths and extensions.
Update check_yaml to accept Path and str by @kaanrkaraman -
Benchmarking Docs Update
Our documentation now recommends more representative datasets for benchmarking models, so you’ll get more accurate and relevant performance metrics.
Fix benchmark note in docs by @lakshanthad
Purpose & Impact
- Simpler Training Configurations: Eliminates manual tweaks, especially with different dataset caching strategies.
- Resilient Model Export & Inference: Exported models cope better with large and variable input sizes and metadata.
- Improved User Experience: Clearer setup guides and more flexible tools ease the workflow for beginners and experts alike.
- Accurate Performance Insights: Representative benchmarking leads to more meaningful insights across various hardware.
Community Shout-Outs
We’re welcoming @kaanrkaraman, who made their first contribution with an improvement to YAML file validation—thank you for helping make the Ultralytics ecosystem better!
See all contributors in the v8.3.129 full changelog.
Try It Out and Share Your Feedback
Update to Ultralytics v8.3.129 and experience these improvements for yourself! We’d love to hear your thoughts, suggestions, or bug reports—your feedback directly shapes future releases.
On behalf of the entire YOLO community and Ultralytics team, thank you for your continued contributions and support!
Happy training, exporting, and deploying!