New Release: Ultralytics v8.3.162

:rocket: Ultralytics v8.3.162 Release Announcement!

We’re excited to announce the release of Ultralytics v8.3.162, packed with improvements that boost reliability, hardware compatibility, and overall usability for the entire YOLO community! This release is all about making your YOLO experience smoother, safer, and more productive—whether you’re developing, training, or exporting models.

:glowing_star: Summary

v8.3.162 introduces standardized model loading for consistent behavior, enhanced device and hardware support (including Intel CPUs and GPUs), better visualization for metrics, safer data augmentations, and various quality-of-life tools for both end users and developers. Dive into the details and see how your workflow just got easier!


:sparkles: New Features & Major Improvements

:white_check_mark: Standardized Model Loading

All torch.load usages are now unified with Ultralytics’ torch_load utility, delivering reliable and predictable model file handling across the codebase.
Thanks @Y-T-G! PR #21260

:white_check_mark: Improved Device Compatibility

Cached text embeddings for YOLOE and YOLO-World now load directly on the specified device (CPU or GPU), preventing those tricky device mismatch errors during training.
Thanks @Laughing-q! PR #21237

:white_check_mark: Intel Hardware Detection

A new utility automatically detects Intel CPUs and GPUs, allowing the tools to recommend OpenVINO exports for users with compatible hardware for optimal inference performance.
Thanks @ambitious-octopus! PR #21242

:white_check_mark: Improved Metrics Plotting

Metrics plots for detection, segmentation, and pose estimation are now clearer, easier to interpret, and visually enhanced for more effective evaluation.
Thanks @Laughing-q! PR #21259

:white_check_mark: Relative Path Support for Grounding Datasets

Open-vocabulary model training now supports relative dataset paths in DATASET_DIR, making it easier to manage and switch between custom datasets.
Thanks @Laughing-q! PR #21256


:hammer_and_wrench: Improvements

  • CopyPaste Augmentation Fix:
    The CopyPaste augmentation process now preserves your original images, preventing unwanted changes during training.
    Thanks @Y-T-G! PR #21262

  • Dependency Version Pinning:
    The ai-edge-litert package is now pinned to versions >=1.2.0,<1.4.0 for more stable TensorFlow SavedModel export workflows.
    Thanks @Laughing-q! PR #21265

  • Optional Typing Stubs:
    We’ve added an optional dependency group for typing stubs, boosting autocompletion and static analysis for developers.
    Thanks @jorenham! PR #21137


:bug: Bug Fixes

  • Open-Vocabulary Evaluation Fixes:
    Improved handling for LVIS in open-vocabulary evaluation.
    Thanks @ImJaewooChoi! PR #21245
  • Assorted Minor Fixes:
    Includes enhancements to dataset handling and various small code improvements for extra stability.

:raising_hands: New Contributors

A big welcome and thank you to @jorenham for their first contribution!
PR #21137


:link: Additional Resources


:triangular_flag: Try It Out!

Upgrade to v8.3.162 today and experience the difference! Your feedback is invaluable—let us know what’s working, how it helps you, and where we can do even better. The progress we make is thanks to the community, and we’re excited to see what you’ll build with the latest Ultralytics release!

Happy training!
— The Ultralytics Team

(As always, you can find installation guides, tutorials, and more on the Ultralytics official documentation site.)