Algorithm breaks when using cuda

[papalotis]

The ESN Algo breaks (throws an exception) when it is run with a GPU. More specifically the function finalize in /echotorch/nn/RRCell.py.

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
~/bachelor/test.py in <module>
     68
     69 # Finalize training
---> 70 esn.finalize()
     71
     72 # Train MSE

~/libs/EchoTorch/echotorch/nn/ESN.py in finalize(self)
    187         """
    188         # Finalize output training
--> 189         self.output.finalize()
    190
    191         # Not in training mode anymore

~/libs/EchoTorch/echotorch/nn/RRCell.py in finalize(self, train)
    178
    179                 # Algo
--> 180                 ridge_xTx = self.xTx + self.ridge_param * torch.eye(self.input_dim + self.with_bias, dtype=self.dtype)
    181                 inv_xTx = ridge_xTx.inverse()
    182                 self.w_out.data = torch.mm(inv_xTx, self.xTy).data

RuntimeError: expected type torch.cuda.FloatTensor but got torch.FloatTensor

This error was created by running the Mackey Glass example The workaround that I have for now is

 # Algo
algo_eye = torch.eye(self.input_dim + self.with_bias, dtype=self.dtype)

#convert to cuda if necessary
if self.xTx.is_cuda:
    algo_eye = algo_eye.cuda()

ridge_xTx = self.xTx + self.ridge_param * algo_eye

Definitely not the best solution but works for now

[MrMonotreme]

Yes, I had a similar issue. Could a commit be scheduled to remedy this problem? I attempted using a global change to the gpu device, but was unsuccessful in changing the device of a resident tensor in the EchoTorch library. I believe @papalotis' solution will work best for my ends.

[nschaetti]

Solved in 6159b46ed280df738b10e0aa3d290dc4314b883f (22/01/2021), dev branch No tests with NARMA-10 with CUDA