RuntimeError: mat2 must be a matrix, got 1-D tensor

1 
new_RBM.py in v_to_h(self, v)
     37 
     38         p_h = F.sigmoid(
---> 39             F.linear((v), (w).squeeze(), bias=h_bias)
     40         ).cuda()
     41 
RuntimeError: mat2 must be a matrix, got 1-D tensor

trouble with linear input tensor. when I remove .squeeze() code then got error like this.

     37 
     38         p_h = F.sigmoid(
---> 39             F.linear((v), (w), bias=h_bias)
     40         ).cuda()
     41 

RuntimeError: The expanded size of the tensor (1) must match the existing size (13000000) at non-singleton dimension 1.  Target sizes: [1, 1].  Tensor sizes: [13000000]

same part, other error. I don’t know how solve this problem.

2

Can you share a minimal reproducible example of this error? (Or at least some of the source code?)

More info on the size of v and w would help solve this issue as it’s mostly likely due to the input Tensor not be compatible with the F.linear operation.

3

Im not sure does it helpful here is it.

    def v_to_h(self, v):
        v = (v.clone().detach()).reshape(-1, 13000000)
        h_bias = (self.h_bias.clone())        
        w = (self.W.clone())

        size_msg = '''
        {} {} {}
        '''.format(torch.flatten(v).size(), torch.flatten(w).size(), h_bias.size())
        print(size_msg)
        
        p_h = F.sigmoid(
            F.linear(torch.flatten(v), torch.flatten(w), bias=h_bias)
        ).cuda()

        sample_h = self.sample_from_p(p_h)
        return p_h, sample_h
4

Yeah, so the issue is that you’re using torch.flatten() on v and w, these two tensors need to be matrices and not vectors, the sizes are defined in the documentation here

5

hmm, okay.
I wondering why does not output size can’t be 13000000. Because torch.flatten(v) and bias size is 13000000, and original weight size 1, 13000000. Thats reason why I changed the size.

Set it up all output feature size got error.

6

So the issue is the size of the bias Tensor,

v.shape #returns torch.Size([1300])
weight.shape #returns torch.Size([1, 1300])
bias.shape #returns torch.Size([1300])
out = F.linear(v.unsqueeze(0), weight) #linear without bias
out.shape #returns torch.Size([1, 1])

The bias needs to be same shape as out (so just a scalar). This is the issue

7

really thank you… i got a other issue that overflow… i will try it to figure out. thanks a lot.