Loss and IoU speed improvements

[glenn-jocher]

Last index slicing in build_targes() 3x faster (3us to 1us) in ec5ccde

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Refactor of bounding box operations and loss calculations in YOLOv5 for enhanced clarity and efficiency.

📊 Key Changes

• 🔧 Modified bbox_iou function to handle box coordinates more consistently and cleanly.
• 💡 Changed how targets are built by simplifying tensor concatenation.
• 🧹 Cleaned up usage of tensor operations by applying squeeze() and more expressive slicing.
• 🔍 Refined computation of IoU by directly using box_area and avoiding unnecessary transposes.
• 🛠 Relocated box_iou from metrics.py to general.py to consolidate IoU related operations.

🎯 Purpose & Impact

• 💎 Improved Code Quality: The simpler, more legible code will be easier to maintain and less prone to bugs.
• 🚀 Increased Efficiency: The changes potentially make the code execute faster and use less memory by optimizing tensor operations.
• 🛁 Cleaner Interfaces: Users will experience cleaner internal operations, leading to potentially fewer errors and improved performance when training or using YOLOv5 models.