Trying to make a RNN for 2D signal data classification to 2D classified matrix output. Error using trainNetwork (line 165) Invalid training data. Responses must be a vector of categorical responses, or a cell array of categorical response sequences.

Hello Everyone,

I'm trying to implement a Neural Network classification algorithm for signal data with a shape (800,500) , with 800 being the number of time steps and 500 being the number of observations. I want to train a NN to identify for every timestep if it belongs to class 0, 1, or -1 . So, my responses should have the same shape with the XTrain data, (800,500).

After trying to use Xtrain and YTrain in the form of simple 2 dimensional arrays I understood that this is not possible, so I made changed their form to cell arrays as the bibliography requires.

My XTrain data have the form of a cell array of double sequences and so do the YTrain data. Out of 500 observations I used 70%, which are 350 observations for my Train data.

As the Matlab trainNetwork bibliography suggest:

But, I'm still getting the same error:

The Layers and Options for my RNN so far are the following, please make sugestions :)

Please help

Answer 1

[Srivardhan Gadila]

According to the documentation of the Input Arguments: sequences & responses of the trainNetwork function for the syntax net = trainNetwork(sequences,responses,layers,options) the input data should be of N-by-1 cell array of numeric arrays, where N is the number of observations and each observation must be a c-by-s matrix, where c is the number of features of the sequences and s is the sequence length in case of Vector sequences. Whereas the responses should be N-by-1 cell array of categorical sequences of labels, where N is the number of observations with each observation as a 1-by-s sequence of categorical labels, where s is the sequence length of the corresponding predictor sequence.

The following code may help you:

%% Create network.
inputSize = 800;
numClasses = 3;
numHiddenUnits = 100;

layers = [ ...
    sequenceInputLayer(1,'Name','Sequence Input')    
    lstmLayer(numHiddenUnits,'Name','LSTM Layer')
    fullyConnectedLayer(numClasses,'Name','FC')
    softmaxLayer('Name','Softmax')
    classificationLayer('Name','Classification Layer')];
lgraph = layerGraph(layers);

analyzeNetwork(lgraph)
%% Create Random Training data.
numTrainSamples = 50;
trainData = arrayfun(@(x)rand([1 inputSize]),1:numTrainSamples,'UniformOutput',false)';
trainLabels = arrayfun(@(x)categorical(randi([-1 1], 1,inputSize)),1:numTrainSamples,'UniformOutput',false)';
size(trainData)
size(trainLabels)
%% Train the network.
options = trainingOptions('adam', ...
    'InitialLearnRate',0.005, ...
    'LearnRateSchedule','piecewise',...
    'Verbose',1, ...
    'Plots','training-progress');

net = trainNetwork(trainData,trainLabels,lgraph,options);

[Orestis Marantos]

Hello Srivardhan, thank you so much for your thorough explanations. Could you also recommend me if the way to overpass the high skewness of my problem? I created with simulations more samples.

Specifically, I have 2000 samples each with 800 time steps. In each sample there are some points of interest (signal goes up (on), signal goes down (off)). I would like to detect those points so I thought that MIMO multi-label classification with lstm could solve my problem. So, I want to classify each time step of each sample as 0 if it is not a point of interest, or 1 if it is (changed the -1 also to 1). Eventually I'm interested only in ones so I'm not sure if it is the best method to follow.

Because more than 98% of the timesteps are zeroes the lstm keeps giving my only zeroes after the training. I have implemented a custom weighted binary cross entropy function and I give 0.97 weight to the ones and 0.03 to zeros, but the algorithm keeps giving me only zeros as a result for each prediciton.

Do you have any suggestions on how to solve this highly skewed problem? I read about over-sampling and under-sampling but this won't help in my problem as it is natural to have this very little number of active points (ones).

Thanks a lot in advance