Ema patch fix #279

[NanoCode012]

Do not save keys in ["module", "process_group", "reducer"] for ema attributes for fix #279

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved multi-GPU support and cleaner code for YOLOv5.

📊 Key Changes

• 🖥️ Introduced multi-processing and distributed training enhancements to better support multi-GPU setups.
• 🛠️ Made modifications to allow devices to be set more flexibly when initializing training, including automatic device selection if not specified.
• ✂️ Removed redundant print statements to declutter console output during training.
• 👾 Refactored Create_dataloader function to support distributed training with the split flag and rank-aware samplers.
• 🧽 Cleaned up code within the training loop to ensure proper device assignment of the train.py script.

🎯 Purpose & Impact

• 📈 The changes pave the way for more efficient and scalable training on multiple GPUs, leading to faster model training and experimentation times.
• 🔍 The modifications and refactoring improve code maintainability, making it easier for other developers to work with and contribute to the codebase.
• ✨ The cleaner console output and code structure enhance the user experience by making it easier to monitor and understand training progress.

[NanoCode012]

Trained on 1,2,4 GPU using python train.py --weights yolov5s.pt --epochs 4 --img 320 --device 0.. and used python test.py --weights weights/last.pt to test.

[NanoCode012]

Sorry. I made a mistake with merging my branches. Will reopen in new branch.