Runtime error when you try to create a tensor using arg "devcie ='cuda' in ContinuousDiceStateActionWightValueLearning

[xiaobaijh]

In classs ContinuousDiceStateActionWightValueLearning ,line 400 we use state = torch.FloatTensor(state, device=self.device) to create a tensor,but if self.device is 'cuda'，it will throw error
'RuntimeError: legacy constructor expects device type: cpu but device type: cuda was passed'，it should be state = torch.tensor(state, device=self.device).So as the other tensors in this class.

[malone-c]

+1 on this issue. This is preventing me from fitting weight functions on GPU using either the DICE or MWL methods.

The torch.FloatTensor function is an outdated part of the legacy API. See here for example: https://discuss.pytorch.org/t/cannot-construct-tensor-directly-on-gpu-in-torch-1-10-1/153751/2.