Custom deep learning layers in MATLAB

I am working on a image classification problem with CNN. I was trying to build my own customized convolution layer for a deep learning architecture in matlab. The following class was written for the same.

classdef CustomConvolutionLayer < nnet.layer.Layer ... % & nnet.layer.Acceleratable properties FilterSize NumFilters

end properties (Learnable) % Layer learnable parameters Weights

end

methods

function layer = CustomConvolutionLayer(filterSize)

% Set layer name.

layer.Name = "SC";

% Set layer description.

layer.Description = "Sparse Convolution";

% Initialize layer weights.

layer.FilterSize = filterSize;

layer.NumFilters = 1;

layer.Weights = randn(filterSize);

end

function Z = predict(layer, X)

Z = conv2(X,layer.Weights,'same');

end

function [dLdX,dLdW] = backward(layer,X,~,dLdZ,~)

dLdW = zeros(size(layer.Weights),'like',layer.Weights);

tW=conv2(X,dLdZ,'same');

fvalue = mean2(tW(:));

dLdW(:,:)=fvalue;

dLdX = conv2(dLdZ, rot90(layer.Weights, 2), 'same');

end

end

end

when I am trying to make use of the custom layer with the following deep learning architecture

inputSize = [256 256 1]; numClasses = 25;

layers = [ imageInputLayer(inputSize)

sparseConvolutionLayer03([5 5]) batchNormalizationLayer reluLayer fullyConnectedLayer(numClasses) softmaxLayer classificationLayer];

options = trainingOptions('sgdm', ... 'MaxEpochs',1, ... 'ValidationData',imdsValidation, ... 'ValidationFrequency',7, ... 'Verbose',true, ... 'Plots','training-progress');

net = trainNetwork(imdsTrain,layers,options); disp('Training Over...')

I’m getting the following error

Error in funCNNclassifier01 (line 34) net = trainNetwork(imdsTrain,layers,options); Caused by: Error using conv2 N-D arrays are not supported.

My input size is 256×256 only, but if I check its size, it was shown that 256×256×1×128, which is why this error is thrown. I am very unsure why / how the dimension is changing from 2D to 4D. Your assistance would be more appreciated.

Answer 1

[Venu]

Hi @Kaja Mohideen

When you use the default settings for the imageInputLayer, it automatically assumes that the input data represents images with multiple channels, hence interpreting it as a 4D array (height x width x channels x batch size). Even if your input data is grayscale (single-channel), it will still interpret it as a 4D array by default.

To avoid this, you need to explicitly specify 'Normalization' as 'none' in the imageInputLayer to inform the network that the input data is a 2D array.

imageInputLayer(inputSize, 'Normalization', 'none');

When you define your training data, ensure that the images are loaded as grayscale and reshaped to the desired input size

imdsTrain = imageDatastore('path_to_training_images', 'IncludeSubfolders', true, 'LabelSource', 'foldernames', 'ReadFcn', @(x)reshape(imread(x),[256,256,1]));

Hope this helps