Wavelet 2D Scattering transform of an input image

Question

I am trying to do a 2D scattering transform on an input image. When I run the following code I get this error: "The filters are not compatible for multiplication."? Can anyone help please? Thanx!

import torch
from kymatio import Scattering2D
import numpy as np
import PIL
from PIL import Image

FILENAME = "add a png file path"
image = PIL.Image.open(FILENAME).convert("L")

a = np.array(image).astype(np.float32)
x = torch.from_numpy(a)
imageSize=x.shape

scattering = Scattering2D(J=2, shape=imageSize, L=8)
Sx = scattering.forward(x)

print(Sx.size()) 

Answer 1

Seems to be working OK for a small 1KB .png with equal width and height (square, not rectangle):

import torch
from kymatio import Scattering2D
import numpy as np
import PIL
from PIL import Image

FILENAME = "/path/to/dir/small_size_1_KB.png"
image = PIL.Image.open(FILENAME).convert("L")

a = np.array(image).astype(np.float64)
x = torch.from_numpy(a)
imageSize = x.shape

scattering = Scattering2D(J=2, shape=imageSize, L=8)

Sx = scattering.forward(x)

print(Sx.size())

Output

torch.Size([81, 19, 19])

The error that you're getting though is in this method ( backend_torch.py ), should have to do something with tensor sizes:

def cdgmm(A, B, inplace=False):
    """
        Complex pointwise multiplication between (batched) tensor A and tensor B.

        Parameters
        ----------
        A : tensor
            input tensor with size (B, C, M, N, 2)
        B : tensor
            B is a complex tensor of size (M, N, 2)
        inplace : boolean, optional
            if set to True, all the operations are performed inplace

        Returns
        -------
        C : tensor
            output tensor of size (B, C, M, N, 2) such that:
            C[b, c, m, n, :] = A[b, c, m, n, :] * B[m, n, :]
    """
    A, B = A.contiguous(), B.contiguous()
    if A.size()[-3:] != B.size():
        raise RuntimeError('The filters are not compatible for multiplication!')

    if not iscomplex(A) or not iscomplex(B):
        raise TypeError('The input, filter and output should be complex')

    if B.ndimension() != 3:
        raise RuntimeError('The filters must be simply a complex array!')

    if type(A) is not type(B):
        raise RuntimeError('A and B should be same type!')


    C = A.new(A.size())

    A_r = A[..., 0].contiguous().view(-1, A.size(-2)*A.size(-3))
    A_i = A[..., 1].contiguous().view(-1, A.size(-2)*A.size(-3))

    B_r = B[...,0].contiguous().view(B.size(-2)*B.size(-3)).unsqueeze(0).expand_as(A_i)
    B_i = B[..., 1].contiguous().view(B.size(-2)*B.size(-3)).unsqueeze(0).expand_as(A_r)

    C[..., 0].view(-1, C.size(-2)*C.size(-3))[:] = A_r * B_r - A_i * B_i
    C[..., 1].view(-1, C.size(-2)*C.size(-3))[:] = A_r * B_i + A_i * B_r

    return C if not inplace else A.copy_(C)

Source

https://github.com/edouardoyallon/pyscatwave/blob/master/scatwave/utils.py