Best way to convolve on different channels with single kernel - revisited

autograd
1

Hi,
I tried to utilize the following topic:

what I tried is:

kernel = np.ones((1, 3, 1))
padding_value = math.floor(kernel.shape[1]/2.)
torch_kernel = torch.Tensor(kernel).permute(0, 2, 1)
data = [[[1., 1.], [2., 2.], [3., 3.], [4., 4.], [5., 5.], [6., 6.], [7., 7.]], [[1., 1.], [2., 2.], [3., 3.], [4., 4.], [5., 5.], [6., 6.], [7., 7.]]]
torch_data = torch.Tensor(data).permute(0, 2, 1)
data_reshaped = torch_data.reshape(-1, 1, torch_data.shape[2])
conv_out = (F.conv1d(data_reshaped, torch_kernel, padding=padding_value, bias=None) / torch_kernel.shape[2]).view(torch_data.shape[0], torch_data.shape[2], torch_data.shape[1])

conv_out is the following:

tensor([[[1.0000, 2.0000],
         [3.0000, 4.0000],
         [5.0000, 6.0000],
         [4.3333, 1.0000],
         [2.0000, 3.0000],
         [4.0000, 5.0000],
         [6.0000, 4.3333]],

        [[1.0000, 2.0000],
         [3.0000, 4.0000],
         [5.0000, 6.0000],
         [4.3333, 1.0000],
         [2.0000, 3.0000],
         [4.0000, 5.0000],
         [6.0000, 4.3333]]])

What I would like to see instead is:

tensor([[[1.0000, 1.0000],
         [2.0000, 2.0000],
         [3.0000, 3.0000],
         [4.0000, 4.0000],
         [5.0000, 5.0000],
         [6.0000, 6.0000],
         [4.3333, 4.3333]],

        [[1.0000, 1.0000],
         [2.0000, 2.0000],
         [3.0000, 3.0000],
         [4.0000, 4.0000],
         [5.0000, 5.0000],
         [6.0000, 6.0000],
         [4.3333, 4.3333]]])

I am basically trying to do a convolution for a multiple series with fixed kernel for each series. If I unpack one dimension into batches, then I need multi-channel kernel, which doesn’t do the same thing. Could you please help me achieve the desired form of this operation? Maybe this approach my choice of approach is entirely wrong?

2

You could permute the result tensor as seen here:

conv_out = (F.conv1d(data_reshaped, torch_kernel, padding=padding_value, bias=None) / torch_kernel.shape[2])
ret = conv_out.view(torch_data.shape).permute(0, 2, 1)
print(ret)
> tensor([[[1.0000, 1.0000],
           [2.0000, 2.0000],
           [3.0000, 3.0000],
           [4.0000, 4.0000],
           [5.0000, 5.0000],
           [6.0000, 6.0000],
           [4.3333, 4.3333]],

          [[1.0000, 1.0000],
           [2.0000, 2.0000],
           [3.0000, 3.0000],
           [4.0000, 4.0000],
           [5.0000, 5.0000],
           [6.0000, 6.0000],
           [4.3333, 4.3333]]])

and you could also use a grouped conv approach:

out = F.conv1d(torch_data, torch_kernel.expand(2, -1, -1), padding=padding_value, bias=None, groups=2) / torch_kernel.shape[2]
out = out.permute(0, 2, 1)
print(out)
1 Like