How to replace convolutional layer in CNN with Fourier convolutional layer?

I try to add a new layer (Fourier convolutional) to CNN or embed it inside the convolutional layer, I already, tried to use class function to do that but every time, I get errors. Could you guide me to solve this problem?

classdef FFTConvLayer < nnet.layer.Layer
    properties
        Kernel
    end

    methods
        function layer = FFTConvLayer(kernelSize, numFilters)
            % Constructor to initialize the layer properties
            layer.Kernel = double(randn([kernelSize, numFilters])); % Initialize kernel weights
        end
        
        function Z = predict(layer, img)
            % img = double(img); % Convert input image to double
            fft_img = fftshift(fft2(img));
            real_part_img = real(fft_img);
            imag_part_img = imag(fft_img);
            imgOut1 = [real_part_img(:); imag_part_img(:)];   

            fft_layer = fftshift(fft2(layer.Kernel));
            real_part_layer = real(fft_layer);
            imag_part_layer = imag(fft_layer);
            imgOut2 = [real_part_layer(:); imag_part_layer(:)];   
            
            Z = ifft2(reshape(imgOut1 .* imgOut2, size(img, 1), size(img, 2)));
        end
    end
end

Answer 1

[Shivansh]

Hi Abdulrahman,

I understand that you want to implement a Fourier convolutional layer for CNN. It is difficult to find the error without the context of the problem and information about the input and output.

Make sure that you are using the class object correctly in your neural network. A sample code snippet would look like this:

kernelSize = [3, 3]; % Example kernel size
numFilters = 10; % Example number of filters
layer = FFTConvLayer(kernelSize, numFilters);

The other possible problem can be the size of image and input and output configuration for the layers. You should evaluate the configurations of the model again before setting the input and the output configurations of the layer.

You can find more information about customised layer in the neural networks here: https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layers.html.

Hope it helps!