Multi variable prediction LSTM

Hi,

I am trying to train a network on a dataset. I would like to predict the coefficients of a spline that represent a contact force between to impacting objects and the total contact time. The spline coefficient vector has variable length depending on the case but as of now I get predictions of the same length ( set to the maximum lenght of the coefficients vector). This is the code I've written

inputTable = table;  
predictorNames = {'v_i', 'E_tip', 'rho_tip', 'v_tip', 'Y_tip', ...
                  'Radius', 'E_plate', 'rho_plate', 'v_plate', 'Y_plate', 'Insulator'};
predictors = inputTable{:, predictorNames}';

maxLen = max(cellfun(@numel, inputTable.Sp)); 
responseCoeffs = zeros(maxLen, length(inputTable.Sp)); 
scalarResponse = inputTable.ContactTime; 

for i = 1:length(inputTable.Sp)
    coeffs = inputTable.Sp{i}; 
    responseCoeffs(1:numel(coeffs), i) = coeffs; 
end

predictors = normalize(predictors, 2);
responseCoeffs = normalize(responseCoeffs, 2);
scalarResponse = normalize(scalarResponse);

numFeatures = size(predictors, 1); 
numSplineCoeffs = maxLen; 

% LSTM 
layers = [
    sequenceInputLayer(numFeatures)  
    lstmLayer(50, 'OutputMode', 'sequence')  
    fullyConnectedLayer(100)  
    reluLayer  
    fullyConnectedLayer(numSplineCoeffs + 1) 
    regressionLayer  
];

options = trainingOptions('adam', ...
    'MaxEpochs', 1000, ...
    'MiniBatchSize', 32, ...
    'Shuffle', 'every-epoch', ...
    'Plots', 'training-progress', ...
    'Verbose', true);

fullResponse = [responseCoeffs; scalarResponse'];

net = trainNetwork(predictors, fullResponse, layers, options);

This is my first time trying to train a network so any advice even on where to find more information on how to select the corret type of network would be much appreciated.

Thanks in advance!

Answer 1

[praguna manvi]

Hi @Riccardo,

Choosing the right architecture depends on the complexity of the problem. Here are some resources that discuss which approach to take based on the type of problem:

• Deep Learning Tips and Tricks: https://www.mathworks.com/help/deeplearning/ug/deep-learning-tips-and-tricks.html
• Choosing a Network Architecture: https://www.mathworks.com/campaigns/offers/next/all-about-choosing-a-network-architecture.html

Since you have only 1800 samples, it would be easier to fit them with a lighter network, as more parameters require more data. Also, since R2023b, it is recommended to use "trainnet" instead. Refer to the following link for more examples:

• Train Network: https://www.mathworks.com/help/deeplearning/ref/trainnet.html