zero-mAP fix return .detach() to EMA

[glenn-jocher]

Resolves ultralytics/hub#82

Signed-off-by: Glenn Jocher [email protected]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced stability in model updates during training for the YOLOv5 project.

📊 Key Changes

• Adjusted the Exponential Moving Average (EMA) update function by detaching the model state_dict tensors during the EMA update process.

🎯 Purpose & Impact

• Purpose: To ensure that the gradients are not calculated or tracked during the update of EMA, improving computational efficiency and reducing potential errors.
• Impact: Users should experience more stable training as this change will prevent inadvertent influence on gradient calculations from the EMA update step. This contributes to more reliable convergence and performance of the YOLOv5 models. 🚀