yolo.py --profile updates

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved model profiling in YOLOv5.

📊 Key Changes

• Removed the model.train() line that initially sets the model to training mode.
• During profiling, the model is now set to evaluation mode and fused with model.eval().fuse() call.
• Changed the profiling image batch size from 16 to 8 when CUDA is available.

🎯 Purpose & Impact

• Ensures the model is in the correct state (evaluation mode) for profiling, potentially leading to more accurate performance measurements.
• Model fusion during evaluation optimizes the model for inference, which should result in a faster and more efficient profiling process.
• Reduced batch size may lower memory requirements during profiling, making it more accessible on devices with limited GPU memory and potentially speeding up the profiling process.

🚀 These changes could offer developers a better understanding of the model's performance characteristics and enable smoother experiences when optimizing and deploying YOLOv5 models.