torch.nn.functional.Conv3d groups problem

1

Hi everyone!

Since I use custom weights in my network I often rely on torch.nn.functional.conv3d rather than torch.nn.Conv3d.
Today I wanted to start to experiment with different group sizes, however there seems to be a bug in the current pytorch 0.2 version?
Following code snippet for example doesn’t work for me:

import torch
myX = torch.autograd.Variable(torch.randn(1,2,5,5,5))
myW = torch.autograd.Variable(torch.randn(4, 2, 3, 3, 3))
y = torch.nn.functional.conv3d(myX, myW, groups=2)

The resulting error:

in conv3d

return f(input, weight, bias)
RuntimeError: size mismatch, m1: [2 x 54], m2: [27 x 27] at /opt/conda/conda-bld/pytorch_1502008109146/work/torch/lib/TH/generic/THTensorMath.c:1293

If I use

c3d = torch.nn.Conv3d(2,4,3,groups=2)
y = c3d(myX)

everything works fine. Nevertheless I would like to be able to use the nn.functional.conv3d interface.
Does anyone face the same problem? Is this a bug for which I should open an issue on github?

Greetings!

2

this definitely seems like a bug. if you would open an issue at https://github.com/pytorch/pytorch/issues i’ll take care of it right away.

3

Hi,
what is the difference between conv3d and conv2d?

Can you give me some suggestions?

Thank you so much.

4

conv3d is for convolution in 3D space (depth x height x width) like videos, 2D is for images

5

I looked into this more closely. This is actually because you gave myW of incorrect size.

If groups=2, then myW has to be of shape (4, 1, 3, 3, 3), i.e. (outputChannels, inputChannels // groups, kT, kH, kW)