Add support for different normalization layers

[vardanagarwal]

Adding normalization layers like LazyBatchNorm, GroupNorm, InstanceNorm, LayerNorm, etc. to the optimizer group training strategy.

#7375

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced normalization layers handling during optimization in the YOLOv5 training script.

📊 Key Changes

• Broadened normalization layers recognition: The code now identifies a range of normalization layers, not just BatchNorm2d, ensuring proper optimizer configuration.

🎯 Purpose & Impact

• Improved flexibility: This change allows YOLOv5 to work with more types of normalization layers, which can be beneficial for experimenting with novel model architectures.
• Optimization accuracy: By correctly identifying which model weights should not have weight decay applied, the training process is more robust, potentially improving model performance.

This update is a technical improvement that can have positive effects on training custom models with different normalization layers, helping maintain YOLOv5's adaptability. 💡🛠

[glenn-jocher]

@vardanagarwal PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐