ultralytics 8.3.130

v8.3.130 - `ultralytics 8.1.130` Faster `model.fuse()` operations (#20466)

on Python PyPI

🌟 Summary

This release boosts model initialization speed by optimizing how model layers are fused, adds clearer training metric access for users, and strengthens ONNX export testing and workflow security. 🚀

📊 Key Changes

• Faster Model Initialization: Improved the model.fuse() process by performing layer fusion on the CPU before moving data to the GPU, making model startup more efficient.
• Better Training Monitoring: Added a new example in the documentation showing how to use the on_model_save callback to access and print key training metrics after each checkpoint.
• Enhanced ONNX Export Testing: Expanded ONNX export tests to cover more scenarios, improving reliability and catching more potential issues.
• Workflow Security & Maintenance: Updated GitHub workflow permissions for safer automated formatting and labeling, and added missing license headers for better compliance.

🎯 Purpose & Impact

• Faster Model Startups: Users will notice quicker model initialization, especially on systems with limited GPU memory, leading to a smoother experience when loading or deploying models. ⚡
• Easier Training Insights: The improved callback documentation helps users—both beginners and experts—track training progress and results more transparently, making it easier to debug and optimize models.
• More Reliable Model Exports: Broader ONNX export testing ensures exported models work as expected across different setups, reducing surprises in deployment.
• Improved Security & Compliance: Tighter workflow permissions and proper licensing keep the codebase safer and more professional for all contributors.

Overall, this update makes Ultralytics models faster to use, easier to monitor during training, and more robust for deployment and development.

What's Changed

• Add on_model_save callback with Python example to Callbacks docs by @RizwanMunawar in #20531
• Potential fix for format.yml missing permissions by @glenn-jocher in #20556
• Add ONNX CUDA CI by @glenn-jocher in #20448
• ultralytics 8.1.130 Faster model.fuse() operations by @dianyo in #20466

New Contributors

• @dianyo made their first contribution in #20466

Full Changelog: v8.3.129...v8.3.130