ultralytics 8.3.59

v8.3.59 - `ultralytics 8.3.59` Add ability to load any `torchvision` model as module (#18564)

on Python PyPI

🌟 Summary

The latest release, v8.3.59, introduces the ability to load any torchvision model as a backbone, along with several quality-of-life updates, including enhanced Docker support, dataset path refinements, and usability improvements in documentation and tools. 🚀

📊 Key Changes

• 🔥 Custom TorchVision Backbone Support: Allows integration of any torchvision model (e.g., EfficientNet, MobileNet, ResNet) as YOLO backbones! Includes options for pretrained weights and layer customization.
• 🖼️ Expanded Segmentation Mask Support: Added compatibility for .jpg masks alongside existing .png support.
• 🐛 Validation Enhancements for INT8 Calibration: New checks ensure calibration datasets meet batch size requirements, providing more robust error handling.
• 🛠️ Improved Docker Environment: Simplified JupyterLab installations and introduced retry mechanisms for Docker image pushes.
• 🔧 Updated Dataset Paths: Refined YAML dataset path structures for better organization and reduced misconfigurations.
• ⚙️ Enhanced Multi-Processing Documentation: Help added for common Windows-related training errors (e.g., RuntimeError) with clear solutions.
• 📊 New Benchmarks: Extended NVIDIA DeepStream and Coral TPU performance benchmarks for development on Jetson devices and Raspberry Pi (including Pi 5).

🎯 Purpose & Impact

• Flexibility & Power with TorchVision Backbones:
  • Users can now integrate models like ConvNext and MobileNet directly into YOLO pipelines.
  • Pretrained weights accelerate training for both object detection and classification tasks. 🔄
• Streamlined Segmentation Workflows:
  • Compatibility with .jpg masks eliminates a frequent need for manual file conversions, saving time. 🕒
• INT8 Improvements:
  • The validation on calibration size prevents breakdowns in compression workflows, ensuring higher-quality deployment setups.
• Smoother Docker & DevOps:
  • Better Docker resilience and JupyterLab setup reduce installation friction for developers. 🐳
• Training Guidance on Windows:
  • Clear troubleshooting advice mitigates pitfalls for users launching scripts in Windows environments for seamless training experiences.
• Enhanced Benchmark Documentation:
  • Developers can now better choose the hardware and YOLO model precision (e.g., FP32, FP16, or INT8) for NVIDIA Jetson or Edge TPU use cases. 📈

This release offers powerful new capabilities for model customization and smoother workflows, making it a significant upgrade for developers working with YOLO and associated tools. 🎉

What's Changed

• Add instructions to enable W&B logging by @Y-T-G in #18546
• Add warning about Windows multi-processing error when launching training by @Y-T-G in #18547
• Ultralytics Refactor https://ultralytics.com/actions by @glenn-jocher in #18555
• Use uv for Dockerfile-jupyter by @glenn-jocher in #18567
• Add retries to Docker pushes by @glenn-jocher in #18565
• Update Benchmarks for NVIDIA DeepStream running on NVIDIA Jetson by @lakshanthad in #18603
• Add benchmarks for Pi 4B/5 by @Skillnoob in #18580
• Update package-seg.yaml by @RizwanMunawar in #18594
• Verify dataset >= batch size on INT8 export calibration by @Y-T-G in #18611
• Fix incorrect docstring for bbox_iou function by @visionNoob in #18579
• Include .jpg in mask converter by @Y-T-G in #18576
• ultralytics 8.3.59 Add ability to load any torchvision model as module by @Y-T-G in #18564

New Contributors

• @visionNoob made their first contribution in #18579

Full Changelog: v8.3.58...v8.3.59