deep learning layer with different output dimension than the input

I want to create a layer where it inputs 3D data with dimesnion labels 'CBT' and outputs reshaped data with dimesion 'SCBT'.

I tried using the 'ProjectAndReshapeLayer' given by Mathworks but it says outputs must have the same dimensions as inputs.

I tried using a stripdim() command inside the forward function definiton but with no success.

Answer 1

[Kartik]

Hi,

You can create a custom layer in MATLAB to achieve this. Here's an example implementation:

classdef ReshapeLayer < nnet.layer.Layer
    
    properties
        InputDimLabels = {'CBT'} % input dimension labels
        OutputDimLabels = {'SCBT'} % output dimension labels
    end
    
    methods
        function layer = ReshapeLayer(name)
            layer.Name = name;
            layer.Description = "Reshape Layer";
        end
        
        function Z = predict(layer, X)
            % Reshape input data to output shape
            Z = reshape(X, [], size(X,3));
        end
        
        function [dLdX] = backward(layer, X, Z, dLdZ, ~)
            % Reshape gradients back to input shape
            dLdX = reshape(dLdZ, size(X));
        end
        
        function outputSize = forwardPropagateSize(layer, inputSize)
            % Output size calculation based on input size and output dimension labels
            outputSize = [prod(inputSize(1:end-1)), length(layer.OutputDimLabels)];
        end
        
        function inputSize = backwardPropagateSize(layer, outputSize)
            % Input size calculation based on output size and input dimension labels
            inputSize = [outputSize(1:end-1), length(layer.InputDimLabels)];
        end
        
        function outputNames = getOutputArguments(layer)
            % Output argument names based on output dimension labels
            outputNames = layer.OutputDimLabels;
        end
        
        function inputNames = getInputArguments(layer)
            % Input argument names based on input dimension labels
            inputNames = layer.InputDimLabels;
        end
        
        function outputSize = getOutputSize(layer, ~)
            % Output size based on output dimension labels
            outputSize = [NaN, length(layer.OutputDimLabels)];
        end
        
        function inputSize = getInputSize(layer)
            % Input size based on input dimension labels
            inputSize = [NaN, length(layer.InputDimLabels)];
        end
        
        function tf = isValidInputSize(layer, inputSize)
            % Check if input size is compatible with input dimension labels
            tf = isequal(inputSize(end), length(layer.InputDimLabels));
        end
        
        function tf = isValidOutputSize(layer, outputSize)
            % Check if output size is compatible with output dimension labels
            tf = isequal(outputSize(end), length(layer.OutputDimLabels));
        end
    end
end

You can use this layer in your neural network by creating an instance of the layer and adding it to the network:

reshapeLayer = ReshapeLayer('reshape_layer');
layers = [
    imageInputLayer([32 32 3], 'Name', 'input', 'Normalization', 'none', 'DataAugmentation', 'none', 'DimensionLabels', {'Height', 'Width', 'Channels'})
    reshapeLayer
    fullyConnectedLayer(10, 'Name', 'fc')
    softmaxLayer('Name', 'softmax')
    classificationLayer('Name', 'output')
    ];

lgraph = layerGraph(layers);