ultralytics 8.3.162

v8.3.162 - `ultralytics 8.3.162` Replace `torch.load` calls with patched `torch_load` method that defaults to `weights_only=False` (#21260)

on Python PyPI

🌟 Summary

This release (v8.3.162) brings improved reliability and consistency to model loading, enhanced hardware compatibility, and several quality-of-life updates for both developers and users. 🚀🛠️

📊 Key Changes

• Standardized Model Loading:
  All direct uses of torch.load are replaced with Ultralytics' torch_load utility, ensuring consistent and robust model file handling throughout the codebase.
• Improved Device Compatibility:
  Loading of cached text embeddings in YOLOE and YOLO-World now explicitly supports both CPU and GPU, preventing device mismatch errors.
• Intel Hardware Detection:
  Added a new utility to detect Intel CPUs and GPUs, allowing Ultralytics tools to recommend OpenVINO exports for optimal performance on Intel hardware.
• Enhanced Metrics Plotting:
  Metric plots for detection, segmentation, and pose tasks now include clearer labeling and improved processing, making evaluation results easier to interpret.
• Relative Path Support for Grounding Datasets:
  Open-vocabulary model training now supports relative dataset paths, simplifying custom dataset management.
• CopyPaste Augmentation Fix:
  The CopyPaste augmentation no longer modifies original images in-place, preserving data integrity during training.
• Dependency Version Pinning:
  The ai-edge-litert package is now pinned to versions >=1.2.0,<1.4.0 to ensure stable TensorFlow SavedModel exports.
• Optional Typing Stubs:
  Introduced an optional dependency group for typing stubs, improving code completion and static analysis for developers.
• Assorted Bug Fixes:
  Includes fixes for open-vocabulary evaluation, dataset handling, and minor code improvements.

🎯 Purpose & Impact

• Reliability & Consistency:
  By standardizing model loading with torch_load, users and developers benefit from fewer bugs and more predictable behavior when working with PyTorch models.
• Better Hardware Support:
  Users with Intel hardware now receive tailored export recommendations, and device-aware embedding loading prevents frustrating training errors.
• Improved Usability:
  Clearer metric plots, support for relative dataset paths, and safer augmentations make the platform easier and safer to use for both new and advanced users.
• Developer Experience:
  Optional typing stubs and codebase improvements enhance code quality, autocompletion, and maintainability.
• Export Stability:
  Pinning dependencies reduces the risk of export failures, ensuring smoother model conversion workflows.

Overall, v8.3.162 delivers a more robust, user-friendly, and developer-friendly experience across the Ultralytics ecosystem! 🎉

What's Changed

• Optional dependency group for typing stubs by @jorenham in #21137
• YOLOE: specify device for text embedding cache loading by @Laughing-q in #21237
• Fix is_lvis check for open-vocabulary models evaluation by @ImJaewooChoi in #21245
• Fix metrics plotting for Pose and Segment tasks by @Laughing-q in #21259
• Support intel XPU discovery by @ambitious-octopus in #21242
• Add DATASET_DIR relative path compatibility for grounding datasets by @Laughing-q in #21256
• Avoid modifying original image during CopyPaste augmentation by @Y-T-G in #21262
• Pin ai-edge-litert>=1.2.0,<1.4.0 by @Laughing-q in #21265
• ultralytics 8.3.162 Replace torch.load calls with patched torch_load method that defaults to weights_only=False by @Y-T-G in #21260

New Contributors

• @jorenham made their first contribution in #21137

Full Changelog: v8.3.161...v8.3.162