Why moving model and tensors to GPU?

6

Hi I have a similar question here, if I have already created the Tensor in CUDA, do I still need to explicitly copy tensor?

which means should I do it like create tensor and copy?

torch.tensor(batchD, dtype=torch.float).cuda(device=self.device, non_blocking=True)

or directly create tensor in CUDA?

torch.tensor(batchD.state, device=self.device, dtype=torch.float)
1 Like