ultralytics 8.3.116

v8.3.116 - `ultralytics 8.3.116` RTDETR VarifocalLoss `gamma` and `alpha` parameterization (#20292)

on Python PyPI

🌟 Summary

This release brings enhanced loss function customization for advanced model training, improved export options, clearer documentation, and several usability and performance upgrades across the Ultralytics ecosystem. 🚀

📊 Key Changes

• Customizable Loss Functions: Added gamma and alpha parameters to FocalLoss and VarifocalLoss, allowing users to better handle class imbalance and focus on hard-to-classify examples.
• YOLOE Model Naming Consistency: Updated all YOLOE references and documentation to use the latest YOLO11-based naming (e.g., yoloe-11s-seg.pt), making it easier to select and use the correct models.
• TorchScript Export Enhancement: Added support for exporting models in half-precision (half argument) for TorchScript, improving performance on compatible hardware.
• Improved Non-Max Suppression (NMS): Fixed class filtering logic in NMS to ensure more accurate and reliable detection results when filtering by class.
• Export Flexibility: Introduced explicit control over bounding box output format (xyxy vs. xywh) during export, reducing confusion and making integration with deployment environments smoother.
• Ultralytics Solutions Label Customization: Added options to show/hide labels and confidence scores in Solutions modules, with a unified label formatting method for consistent output.
• Docker Image Update: Upgraded the base Docker image to PyTorch 2.7.0 for better compatibility and performance.
• Documentation Improvements: Refined docstrings, clarified return types, and updated code examples for better developer experience.
• Security Upgrade: Set GitHub workflow permissions to read-only for improved CI/CD security.

🎯 Purpose & Impact

• Greater Training Flexibility: Fine-tune loss function behavior to address challenging datasets, leading to potentially better model accuracy and robustness.
• Easier Model Selection: Consistent YOLOE naming reduces user confusion and errors in both code and documentation.
• Faster, Leaner Exports: Half-precision TorchScript exports enable faster inference and lower memory usage, especially on modern GPUs.
• More Reliable Detections: Improved NMS and class filtering logic ensure users get accurate results when working with specific object classes.
• Seamless Deployment: Explicit bounding box format control and Docker updates make it easier to deploy models in diverse environments.
• Customizable Visual Outputs: Solutions users can now tailor label and confidence display to their needs, enhancing presentation and clarity.
• Better Developer Experience: Improved documentation and code clarity help both new and experienced users work more efficiently.
• Enhanced Security: Workflow permission changes follow best practices, reducing risk in automated processes.

This update is packed with improvements for both model developers and end users, making Ultralytics tools more powerful, flexible, and user-friendly. 💡✨

What's Changed

• Update Tuple return type annotations in docstrings by @RizwanMunawar in #20274
• Fix YOLOE usage in model.py by @RizwanMunawar in #20321
• Fix non_max_suppression when applying classes argument by @RizwanMunawar in #20322
• Add half argument to Torchscript export allowlist by @seungjlee in #20313
• Add show_conf and show_labels for Solutions by @RizwanMunawar in #20282
• FROM pytorch/pytorch:2.7.0-cuda12.6-cudnn9-runtime by @glenn-jocher in #20335
• Remove redundant coordinate conversions for nms=True by @Y-T-G in #20334
• Add workflow-level CI permissions for security by @glenn-jocher in #20336
• ultralytics 8.3.116 RTDETR VarifocalLoss gamma and alpha parameterization by @pow3rpi in #20292

New Contributors

• @seungjlee made their first contribution in #20313
• @pow3rpi made their first contribution in #20292

Full Changelog: v8.3.115...v8.3.116