Error using trainNetwork. Output size does not match the response size.

I have two sets of data, pressuredata and velocitydata, both are 100x100x10 doubles.

I am attempting to use trainNetwork to predict the velocity based on the pressure. I am met with the following error when I run the code:

Error using trainNetwork
Invalid training data. The output size ([1 1 100000]) of the last layer does not
match the response size ([100 100 10]).

Error in Airfoil_in_cross_flow_data (line 116)
net = trainNetwork(inputData, outputData, layers, options);

What changes to I need to make to the layer settings to get this to function? Current code is below:

%% Define and prepare the input and output data
inputData = pressuredata; % Set input = pressure data
outputData = velocitydata; % Set output = velocity field

%% Determine input size
inputSize = numel(inputData); % 100 x 100 x 10
outputSize = numel(outputData); % 100 x 100 x 10

%% Set layer variables
layers = [
    imageInputLayer([100 100 10])
    fullyConnectedLayer(100) % Adjust the number of neurons in the fully connected layer as needed
    reluLayer % Activation function
    fullyConnectedLayer(outputSize)
    regressionLayer % Regression layer for continuous output
];

%% Define training options
options = trainingOptions('adam', ...
    'MaxEpochs', 50, ...
    'MiniBatchSize', 32, ...
    'Verbose', true);

%% Train network
net = trainNetwork(inputData, outputData, layers, options);

[Matt J]

Why are you flattening the data with ,

inputData(:), outputData(:)

[Zachary]

Apologies - that was a carry over from another example I was playing around with. Removing that sees a new error, where there is a difference between the output and response size.

Answer 1

[Ranjeet]

Hi Zachary,

I understand that you want to train a neural network with input and output of size [100, 100, 10]. The input has been given as imageInputLayer but there is a fullyConnectedLayer in the output before regression layer. FullyConnectedLayer does not provide output in the expected dimension but it specifies the output as a one-dimensional vector.

There needs to be reshaping of the data before feeding into regression layer. Following is the modified network and a custom nnet layer to resize the output from fully connected layer has been given:

Modified network:

pressuredata = rand(100, 100, 10);
velocitydata = rand(100, 100, 10);

%% Define and prepare the input and output data
inputData = pressuredata; % Set input = pressure data
outputData = velocitydata; % Set output = velocity field

%% Determine input size
inputSize = size(inputData); % 100 x 100 x 10
outputSize = size(outputData); % 100 x 100 x 10

%% Set layer variables
layers = [
    imageInputLayer([100 100 10])
    fullyConnectedLayer(100) % Adjust the number of neurons in the fully connected layer as needed
    reluLayer % Activation function
    fullyConnectedLayer(numel(outputData))
    reshapeLayer("reshape layer", outputSize)
    regressionLayer % Regression layer for continuous output
];
%% analyze network
analyzeNetwork(layers);

%% Define training options
options = trainingOptions('adam', ...
    'MaxEpochs', 50, ...
    'MiniBatchSize', 32, ...
    'Verbose', true);

%% Train network
net = trainNetwork(inputData, outputData, layers, options);

reshapeLayer class:

classdef reshapeLayer < nnet.layer.Layer
    properties
        outputSize;
    end
    properties (Learnable)
    end
    methods
        function layer = reshapeLayer(name, outputSize)
            layer.Name = name;
            layer.outputSize = outputSize;
        end
        function [Z] = predict(layer, X)
            Z = reshape(X,layer.outputSize(1), layer.outputSize(2), layer.outputSize(3),[]);
        end
    end
    
end

You may refer the following example as well:

Reshaping fully connected output