5

I´m trying out PyTorch's DCGAN Tutorial, using a Mac with an M1 chip.

For some reason, when loading images with the Dataloader, the resulting samples are corrupted when using: device = torch.device("mps"). But when using the device = torch.device("cpu") I get the correct result as shown below:

the result is correct

I checked that MPS is correctly configured in my environment. I'm using PyTorch's nightly build, which supports MPS acceleration on my MacOs version.

Is this a known issue?

Michael M.
  • 10,486
  • 9
  • 18
  • 34
Pablo Arjonilla
  • 51
  • 1
  • I don't know the answer to this and can't test it myself without an M1, but mps support is relatively new to pytorch and I recommend dropping a question in the mps category of the pytorch forum so someone can look closer at this: https://discuss.pytorch.org/c/metal-performance-shader/ – dbish Aug 07 '22 at 18:41

0 Answers0