C-LSTM Input of 4D to predict 12X1 values

I am trying to use a C-LSTM to predict 12 values for each time step with regression.

I'm happy to provide as much information as I can. However, to me, it seems clear that I'm providing the correct inputs for the C-LSTM architecture...

Please help! This is important medical research!

Thank you all.

I tried it with two other formats, one sort of worked (RMSE = .54)

and then I tried it with a cell array around each variable such that the training and validation are just one cell array with a time series of 528*n and kinematics is 12*n and somehow that worked

layers = [ ...

sequenceInputLayer(inputSize,'Name','input')

sequenceFoldingLayer('Name','fold')

convolution2dLayer(filterSize,numFilters,'Name','conv')

batchNormalizationLayer('Name','bn')

reluLayer('Name','relu')

sequenceUnfoldingLayer('Name','unfold')

flattenLayer('Name','flatten')

lstmLayer(numHiddenUnits,'OutputMode','sequence','Name','lstm')

reluLayer('Name','relu2')

fullyConnectedLayer(numFeatures*2, 'Name','fc1')

reluLayer('Name','relu3')

fullyConnectedLayer(numFeatures*2, 'Name','fc2')

reluLayer('Name','relu4')

fullyConnectedLayer(numFeatures*2, 'Name','fc3')

fullyConnectedLayer(numClasses, 'Name','fcl')

regressionLayer('Name','regression')];

lgraph = layerGraph(layers);

lgraph = connectLayers(lgraph,'fold/miniBatchSize','unfold/miniBatchSize');

Here are my layers

inputsize is 528 1 1

filtersize is 64 1

numFilters is 50

numclasses is 12

Answer 1

[Srivardhan Gadila]

Refer to the documentation of the Input Arguments: sequences & responses of the trainNetwork function for the syntax

net = trainNetwork(sequences,responses,layers,options) to know the format of the training data.

You can execute the following code to understand the data format:

inputSize = [528 1 1];
filterSize = [64 1];
numFilters = 50;
numClasses = 12;
numHiddenUnits = 200;
numResponses = numClasses;
numFeatures = 10;

layers = [ ...
    sequenceInputLayer(inputSize,'Name','input')
    
    sequenceFoldingLayer('Name','fold')
    
    convolution2dLayer(filterSize,numFilters,'Name','conv')
    batchNormalizationLayer('Name','bn')
    reluLayer('Name','relu')
    
    sequenceUnfoldingLayer('Name','unfold')
    flattenLayer('Name','flatten')
    
    lstmLayer(numHiddenUnits,'OutputMode','sequence','Name','lstm')
    reluLayer('Name','relu2')
    fullyConnectedLayer(numFeatures*2,  'Name','fc1')
    reluLayer('Name','relu3')
    fullyConnectedLayer(numFeatures*2,  'Name','fc2')
    reluLayer('Name','relu4')
    fullyConnectedLayer(numFeatures*2,  'Name','fc3')
    fullyConnectedLayer(numClasses, 'Name','fcl')
    
    regressionLayer('Name','regression')];
lgraph = layerGraph(layers);
lgraph = connectLayers(lgraph,'fold/miniBatchSize','unfold/miniBatchSize');


analyzeNetwork(lgraph)
%%
numTrainSamples = 50;
trainData = arrayfun(@(x)rand([inputSize(:)' 1]),1:numTrainSamples,'UniformOutput',false)';
trainLabels = arrayfun(@(x)rand(numResponses,1),1:numTrainSamples,'UniformOutput',false)';
size(trainData)
size(trainLabels)
%%
options = trainingOptions('adam', ...
    'InitialLearnRate',0.005, ...
    'LearnRateSchedule','piecewise',...
    'Verbose',1, ...
    'Plots','training-progress');
net = trainNetwork(trainData,trainLabels,lgraph,options);