Can anyone help me in reshaping a fully connected layer output to a image?

I am trying to make a deep layer network in which I want to connect my pretrained convolutional layer at the intermediate step? I am unable to write a custom layer for it. I am sharing the code for the custom layer, please le me know where am I going wrong in it.

classdef fullytandem < nnet.layer.Layer

properties

% (Optional) Layer properties

% Layer properties go here

end

properties (Learnable)

% (Optional) Layer learnable parameters

% Layer learnable parameters go here

W;

b;

Z;

end

methods

function layer = fullytandem(input_shape, output_shape, im_layer, name)

% (Optional) Create a myLayer

% This function must have the same name as the layer

% Layer constructor function goes here

layer.Name = name;

layer.Description = 'fully self';

layer.b = rand([im_layer,1]);

layer.W = rand([im_layer,input_shape]);

end

function Z = predict(layer, X)

load('net_28s11.mat')

% Forward input data through the layer at prediction time and

% output the result

%

% Inputs:

% layer - Layer to forward propagate through

% X - Input data

% Output:

% Z - Output of layer forward function

% Layer forward function for prediction goes here

disp('W')

size(layer.W)

disp('X')

size(X)

weights = layer.W;

bias = layer.b;

Y = fullyconnect(X,weights,bias,'DataFormat','SSCB');

% Z = layer.W*X + layer.b';

% disp(size(Z))

% outputSize = X.OutputSize;

disp('Y')

size(Y)

Z = reshape(Y, [4,7,1]);

disp('Z')

size(Z)

% Z = predict(net,dlarray(Z));

end

%

end

end

Answer 1

[Srivardhan Gadila]

Instead of writing the code for fullyconnected layer you can make use of the existing fullyConnectedLayer & write the custom layer code only for the reshape operation as follows:

layers = [ ...
    imageInputLayer([100 1 1],'Name','input','Normalization','none')
    fullyConnectedLayer(100, 'Name','fc')
    reshapeLayer("reshape")
    convolution2dLayer(5,20,'Name','conv')];

dlnet = dlnetwork(layerGraph(layers));
analyzeNetwork(dlnet)

Custom layer code for reshape operation:

classdef reshapeLayer < nnet.layer.Layer

    properties

    end

    properties (Learnable)

    end

    methods
        function layer = reshapeLayer(name)
            layer.Name = name;
        end

        function [Z] = predict(layer, X)
            Z = reshape(X,10,10,1,[]);
        end
    end
    
end

[Aggraj Gupta]

Thanks alot for the answer, Srivardhan. I am still rusky on how to connect this reshape layer to the pretrained network? Say, I have a network saved in the .mat file. We can use this network as predict(net,XTest). How to add this pretrained network layers after the reshape layer?

Could you please guide me on the same?

Thanks again!

[Berk Baris Celik]

This code is not working.

[Gil Oliveira]

It should work if you replace the first line with

classdef reshapeLayer < nnet.layer.Layer & nnet.layer.Formattable

as recommended by the analyzeNetwork() function when you run Srivardhan Gadila's code.

[Atallah Baydoun]

Hey @Srivardhan Gadila

I have implemented this reshape function and I am trying to reshape from 512 to 8 x 8 x 8.

I have changed the reshape settings to Z = reshape(X,8,8,8);

For some reason, the output in the network is 8.

Any idea why this is happening?

Attached the analyzeNetwork

[Atallah Baydoun]

Just found the answer.

You have to add a line to the function as follows:

function [Z] = predict(layer, X)

Z = reshape(X, 8, 8, 8, []);

Z = dlarray(Z,'SSSCB');