Precision-Recall Curve Feature Addition

[glenn-jocher]

This PR adds precision-recall curve plotting to the first and last training epochs, as well as any direct calls to test.py. Curve is saved as precision-recall_curve.png.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancements to test and plotting functions in yolov5.

📊 Key Changes

• Changed save_dir and save_txt parameter defaults and types in test.py.
• Removed merge option from Non-Max Suppression (NMS) and associated parameters.
• Added plots parameter to control the generation of precision-recall curves and batch image saving.
• Updated ap_per_class function in utils/general.py to optionally plot precision-recall curves.
• Cleaned up and removed redundant sections of code such as the removal of existing output images in train.py.

🎯 Purpose & Impact

• 🔄 The refactoring enhances code clarity by using more descriptive defaults for directory paths and clarifying the purpose of save_txt parameter.
• 🧹 Removal of the merge option simplifies the NMS process and reflects updated best practices or model optimizations.
• 📈 The plots parameter allows users to control the output of visual data more granitely, which can be useful for large-scale training or when visual output is not needed, potentially saving resources.
• 📊 Precision-recall curve plotting provides a useful visualization of model performance, directly from the ap_per_class function.
• 🧽 Overall streamlining of the codebase can lead to easier maintenance and better performance, benefiting users with clearer options and potentially faster execution.

[glenn-jocher]

Example precision-recall curve for YOLOv5x on COCO, evaluated at mAP@0.5.

[glenn-jocher]

Addition complete, colab tests on branch complete, CI passing, merging.