Cannnot training due to Expected CUDA but got CPU instead

[thinhnggia]

Yesterday before the update of B4 band B5, I can still train the model from scratch, however, after the update, I see that it seems that the drop_connect function in the utils is not set to CUDA as I think, so it would conflict with the input data which is loaded to GPU already

[lukemelas]

Yes, this makes sense. I'll fix it tomorrow.

[changwuxie]

@thinhnggia
you can advise the fuction 'drop_connect' in utils.py
change: random_tensor += torch.rand([batch_size, 1, 1, 1], dtype=inputs.dtype)
to : random_tensor += torch.rand([batch_size, 1, 1, 1], dtype=inputs.dtype,device=inputs.device)
Then it works!

[lukemelas]

Thanks, @changwuxie ! Just fixed it.