YOLOv5 v4.0 Release

[glenn-jocher]

This PR combines architecture updates, bug fixes and feature additions into a new v4.0 release.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Update and improvements to YOLOv5 models and training methods.

📊 Key Changes

• Updated README with new GPU speed information and v4.0 release details.
• Improved pretrained model checkpoint table with new structure and test results.
• Revised the requirements.txt installation process.
• Enhanced image loading and augmentation methods in datasets.py.
• Modified loss computation to scale object loss by image size and number of detection layers.
• Changed various components, such as activations (from Hardswish to SiLU) and anchor box calculations.

🎯 Purpose & Impact

• Aimed to provide clearer documentation and point to recent developments.
• Model checkpoint adjustments give users the latest performance metrics, helping to compare models and make informed choices.
• Easier and more reliable setup process due to requirements.txt changes enhances user experience.
• Refined data processing and augmentation techniques improve the robustness and accuracy of the model training.
• Adjustments to activation functions and other model components based on the latest research may lead to more efficient and accurate object detection.