Search code examples
numpytorch

Torch7: Addition layer with Tensors of inconsistent size like numpy

I would like to add two tensors with one different dimmension.

For example:

x = torch.ones(4,5)
y = torch.ones(4,3,5)

In numpy I'm cappable of doing this with:

import numpy as np
x = np.ones((4,5))
y = np.ones((4,3,5))
y + x[:,None,:]

#Prints out
array([[[ 2.,  2.,  2.,  2.,  2.],
    [ 2.,  2.,  2.,  2.,  2.],
    [ 2.,  2.,  2.,  2.,  2.]],

   [[ 2.,  2.,  2.,  2.,  2.],
    [ 2.,  2.,  2.,  2.,  2.],
    [ 2.,  2.,  2.,  2.,  2.]],

   [[ 2.,  2.,  2.,  2.,  2.],
    [ 2.,  2.,  2.,  2.,  2.],
    [ 2.,  2.,  2.,  2.,  2.]],

   [[ 2.,  2.,  2.,  2.,  2.],
    [ 2.,  2.,  2.,  2.,  2.],
    [ 2.,  2.,  2.,  2.,  2.]]])

It has a shape of (4,3,5)

Is there any way to reproduce this on a nn.CMulTable()? When I view x tensor like this x:view(4,1,5) it give me an error inconsistent tensor size.

m = nn.CAddTable()
m:forward({y, x:view(4,1,5)})

Any ideas?

Solution

  • Use expandAs to obtain a 4x3x5 tensor:

    m:forward({y, x:view(4,1,5):expandAs(y)})