Tied "conv1d" and "conv_transpose1d" not geting the same result as the input

quantization
1

Here is a snippet of my test code：

    data = torch.tensor([[[1.], [2.], [3.], [4.]]])  # N=1 * L=4 * C_in=1
    data = data.permute(0, 2, 1)
    conv = nn.Conv1d(1, 1, 2, bias=None)
    print(conv.weight)
    output = conv(data)
    print("output : {}".format(output))
    dweight = conv.weight.transpose(0, 1).flip(2, )   # dweight = conv.weight.transpose(0, 1)
    print("dweight : {}".format(dweight))

    dconv = F.conv_transpose1d(input=output, weight=dweight, bias=None)
    print(dconv)   # != data

The result I suppose it would be is: input -> conv1d(input) -> **conv_transpose1d(conv1d(input)) should be equal to input.

But they are not equal, whether I flip the temporal axis or not, but they are supposed to be the same, right?

I’m really confused and frustrated here, could anyone figure it out?

2

Please don’t tag specific people, as this might discourage others to answer.

3

I think I misunderstand the “tied weight” concept.

I wrote the conv_transposed1d in doubly block circulant matrix form and I find that one don’t need to flip the temporal axis actually.

Suppose the conv1d’s matrix is and the corresponding conv_transpose1d’s matrix is .

The square matrix apprently is not always identity matrix. So the result need not to be identical to the input.