Error using trainNetwork: Predictors must be a N-by-1 cell array of sequences

Hello MATLAB friends,

I’m working on a project involving time series prediction using a Gated Recurrent Unit (GRU) model. The goal is to predict whether a stock will be a ‘winning’ or ‘losing’ stock based on historical data.

However, I’m encountering an error when trying to train the network using the trainNetwork function. The error message is:

Error using trainNetwork

Invalid training data. Predictors must be a N-by-1 cell array of sequences, where N is the number of sequences. All sequences must have the same feature dimension and at least one time step.

Here’s the relevant portion of my code:

% Initialize cell arrays to store training and testing features and labels
training_features = {};
training_labels = {};
testing_features = {};
testing_labels = {};

% Process each CSV file in the combined_stocks table
for f = 1:height(combined_stocks)
    % Load data from CSV file using readtable
    data = readtable(fullfile(directory_path, combined_stocks.Stock(f)), 'VariableNamingRule', 'preserve');
    
    % Selecting the required features
    features = data{:, {'open', 'high', 'low', 'close', 'MA', 'MA_1'}};
    
    % Normalizing the close prices
    normalized_close = (data.close - min(data.close)) / (max(data.close) - min(data.close));
    
    % Calculating logarithmic close prices
    log_close = log(data.close);
    
    % Combine all features into a matrix
    features = [features, normalized_close, log_close];
    
    % Determine the label for the current stock
    if combined_stocks.Maximum_Profit(f) > 0
        labels = ones(height(data), 1); % Winning class
    else
        labels = zeros(height(data), 1); % Losing class
    end
    
    % Convert labels to categorical
    labels = categorical(labels);
    
    % Generate a random permutation of indices
    indices = randperm(size(features, 1));
    
    % Split the data into 70% training and 30% testing
    split_point = round(size(features, 1) * 0.7);
    training_indices = indices(1:split_point);
    testing_indices = indices(split_point+1:end);
    
    % Append the training and testing features and labels to the respective arrays
    training_features = [training_features; {features(training_indices, :)}];
    training_labels = [training_labels; {labels(training_indices)}]; 
    testing_features = [testing_features; {features(testing_indices, :)}];
    testing_labels = [testing_labels; {labels(testing_indices)}]; 
end

% Define the GRU network architecture
layers = [  ...
    sequenceInputLayer(size(training_features{1}, 2))
    gruLayer(100,'OutputMode','last')
    fullyConnectedLayer(2)
    softmaxLayer
    classificationLayer];

% Define the training options
options = trainingOptions('adam', ...
    'MaxEpochs',100, ...
    'MiniBatchSize', 150, ...
    'InitialLearnRate', 0.01, ...
    'GradientThreshold', 1, ...
    'ExecutionEnvironment','auto',...
    'plots','training-progress', ...
    'Verbose',false);

% Train the GRU network
net = trainNetwork(training_features, training_labels, layers, options);

In my code, training_features and training_labels are N-by-1 cell arrays. Each cell in training_features contains a matrix of size M-by-8, where M is the number of time steps and 8 is the number of features. Each cell in training_labels contains a categorical vector of length M, where M is the same as in the corresponding cell of training_features.

Despite this, I’m still getting the error. Any help would be greatly appreciated!

Answer 1

[Krishna]

Hi David,

I understand you're aiming to classify the sequence into a categorical variable of either 0 or 1. For a sequence with input dimensions Mx8, your predicted output should have a dimension of 1, not Mx1. This is because you are using the output mode 'last' in GRU layer. If you want it to output a sequence make it to 'sequence.' This adjustment should address your issue.

Please go through the following documentation to learn more,

https://in.mathworks.com/help/deeplearning/ug/classify-sequence-data-using-lstm-networks.html

Hope this helps.