'NoneType' object has no attribute 't'

[schonkopf]

While running the following codes:
net = NMF(data.shape, rank=self.basis_num, W=torch.Tensor(self.W), trainable_W=False).cuda()
net.fit(data.cuda(), verbose=True, max_iter=200, tol=1e-18, beta=1)

I received the error message, any idea?
/content/pytorchNMF/torchnmf/nmf.py in fit(self, V, beta, tol, max_iter, verbose, alpha, l1_ratio)
281
282 with torch.no_grad():
--> 283 WH = self.reconstruct(H, W)
284 loss_init = previous_loss = beta_div(WH, V, beta).mul(2).sqrt().item()
285

/content/pytorchNMF/torchnmf/nmf.py in reconstruct(H, W)
486 @staticmethod
487 def reconstruct(H, W):
--> 488 return F.linear(H, W)
489
490

/usr/local/lib/python3.7/dist-packages/torch/nn/functional.py in linear(input, weight, bias)
1690 ret = torch.addmm(bias, input, weight.t())
1691 else:
-> 1692 output = input.matmul(weight.t())
1693 if bias is not None:
1694 output += bias

AttributeError: 'NoneType' object has no attribute 't'

[yoyolicoris]

@schonkopf
If the Vshape argumet (the first argument) is specified, it will overrided other arguments like W and H.
To use custom template tensor, in your case, you have to manually set W and H.

Example:

net = NMF(W=torch.Tensor(self.W), trainable_W=False, H=(data.shape[0], self.basis_num)).cuda()

Seems I didn't write the document clearly, I will add some assertions in the module to let user know what happened.

[yoyolicoris]

Hi @schonkopf,

In the newest commit e4563b2 I add an warning note in class docstring. Also, I changed the behavior of trainable_* arguments to only take effect when template tensor is given, it's also written in docstring.

While running the following codes:
net = NMF(data.shape, rank=self.basis_num, W=torch.Tensor(self.W), trainable_W=False).cuda()
net.fit(data.cuda(), verbose=True, max_iter=200, tol=1e-18, beta=1)

This way will work without error now, though W would be replaced by random weights.