torch.empty() for speed improvements

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Optimization of tensor initialization in YOLOv5 by using torch.empty instead of torch.zeros.

📊 Key Changes

• Replaced torch.zeros with torch.empty for initializing tensors in various files including models/common.py, models/yolo.py, utils/autobatch.py, utils/loggers/__init__.py, and utils/torch_utils.py.
• This change affects the model warmup process, forward pass, autobatch functionality, TensorBoard logging, and model information functions.

🎯 Purpose & Impact

• 🚀 Purpose: To optimize memory allocation during tensor initialization. torch.empty allocates memory without initializing it, which can be slightly faster than initializing to zero with torch.zeros.
• ⏱ Potential Impact: May result in a minor performance improvement during model initialization and inference, as less time is spent on initializing tensors.
• 👩‍💻 Users: Could benefit from reduced latency in model operations, especially when repeatedly initializing large numbers of tensors.