experimental.py Apple MPS device fix

[glenn-jocher]

May resolve #8102

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced model loading for device compatibility in the YOLOv5 ensemble setup.

📊 Key Changes

• Unified model loading process to first load weights on the CPU.
• After loading, the checkpoint is moved to the specified device (e.g., GPU or CPU) and converted to floating point precision (FP32).

🎯 Purpose & Impact

• 🚀 Ensures greater compatibility across different computing devices by first using the CPU for loading weights.
• 🔍 Aids in avoiding potential out-of-memory errors on GPUs during the model loading phase.
• 🌐 Simplifies the model loading pipeline, making it more robust and maintainable for users working with different hardware configurations.

[glenn-jocher]

Updates seem load faster than mapping directly to GPU...

%timeit torch.load('yolov5x.pt', map_location='cpu')['model'].to(torch.device('cuda:0')).float()
%timeit torch.load('yolov5x.pt', map_location=torch.device('cuda:0'))['model'].float()

1 loop, best of 5: 208 ms per loop
1 loop, best of 5: 254 ms per loop