Using pytorch from source. cat(): argument 'tensors' (position 1) must be tuple of Variables, not Variable

1

I had to compile pytorch from src(some bugs have been fixed in master branch), but when I launched my code, which works well with 0.2 version(current latest release), I get an error:

dec_h = torch.cat(last_hc[0], 1) 

RuntimeError: cat(): argument 'tensors' (position 1) must be tuple of Variables, not Variable

What should I use? What differ “src cat” from current version?
ty

1 Like
2

Supposing last_hc[0] is a variable, just use (last_hc[0],) instead as suggested by the error message.

3

Maybe I don’t understand something. Let me construct some simple example to talk about something more specific :slight_smile:

import torch
A = torch.autograd.Variable(torch.FloatTensor([[[1, 2, 3], [1, 2, 3]], [[4, 5, 6], [4, 5, 6]], [[7, 8, 9], [7,8, 9]]]))
torch.cat((A,), 1)

Output:

Variable containing:
(0 ,.,.) = 
  1  2  3
  1  2  3

(1 ,.,.) = 
  4  5  6
  4  5  6

(2 ,.,.) = 
  7  8  9
  7  8  9
[torch.FloatTensor of size 3x2x3]

which is not what I want which is like this:

1     2     3     4     5     6     7     8     9
    1     2     3     4     5     6     7     8     9
[torch.FloatTensor of size 2x9]
4

You want to give it a collection of tensors, rather than an existing tensor.

You can either:

A = torch.autograd.Variable(torch.FloatTensor([[1, 2, 3], [1, 2, 3]]))
B = torch.autograd.Variable(torch.FloatTensor([[4, 5, 6], [4, 5, 6]]))
C = torch.autograd.Variable(torch.FloatTensor([[7, 8, 9], [7, 8, 9]]))
torch.cat((A, B, C), 1)

or just use the same A, get the (roughly) same result by

A.transpose(0, 1).view(2, 9)
1 Like
5

Now I see, thank you!