How can I convert a data set of doubles into a cell arrays?

I'm starting using Matlab for machine learning and I have a data set of doubles with structure:

6x70117 double

And need to convert it into a:

6x70117 cell

to be able to train the DNN network because I'm receiving this error:

Invalid training data. For sequence-to-one networks, training data must be cell arrays

I've been trying with some loops but always distort the structure of the data set. Is there any other functional way to do this?

Thank you in advance

[Stephen23]

Where M is your numeric matrix:

C = num2cell(M)

[Jan]

Are you sure, that you need a 6x70117 cell? The error message might mean a scalar cell containing the matrix also.

[Manuel Alejandro]

I'm not sure, still looking for the solution. It didn't work with num2cell function, neither making a loop to read the size and data of the original data set and put it into a cell array, bellow it's the code I'm using up to the step of convert the data type.

%Data splitting into training and validation
Training = 0.8
Validation = 0.1

load DataSet %Data structure with 87648x6

[Datalenght , Variables] = size (DataSet);
%Data set preparation
TrainingSteps = floor(Training*Datalenght); 
ValidationSteps = floor(Validation*Datalenght); 
TestSteps = Datalenght-(TrainingSteps+ValidationSteps);
indexTraining = 1:TrainingSteps; 
indexTraining = indexTraining';
indexValidation = (TrainingSteps+1:TrainingSteps+ValidationSteps);
indexValidation = indexValidation';
indexTest = TrainingSteps+ValidationSteps+1:Datalenght;
indexTest = indexTest';
TrainingData = DataSet(indexTraining,:);
ValidationData=DataSet(indexValidation,:);
TestData = DataSet(indexTest,:);
%Data set standarization
mu = mean(TrainingData);
sig = std(TrainingData);
TrainingStandardizedData = (TrainingData) - (mu) / (sig);
ValidationStandardizedData = (ValidationData) - (mu) / (sig);
TestStandardizedData = (TestData) - (mu) / (sig);
%Predictors and responses definition
XTraining = TrainingStandardizeddata(1:end-1,:)';
YTraining = TrainingStandardizeddata(2:end,:)';
XrTraining = cell(size(XTraining));
%Loop for converting predictors into cell array
for i=1:size(XTraining)
    XrTraining{i,1} = XTraining(:,i);
end

[Jan]

@Manuel Alejandro: Although this line is working sometimes:

for i=1:size(XTraining)    

it is a souce of bugs or at least confusions. Remember that size() replies a vector. The colon operator : uses the first element of a vector only and this might work by accident. Prefer to specify the dimension explicitly:

for i=1:size(XTraining, 1)  % Modern Matlab versions: height(XTraining)    

The posted code does not clarify, what you need as input. Which command causes the "Invalid training data" error and what does the documentation explain about the required inputs?

By the way, always post the complete error messages in the forum, not just a part of it. This helps to understand what the problem is.

[Manuel Alejandro]

Dear, thank you for your reply

Originally both, "XTraining" and "YTraining" are doubles with 6x70117, but since for:

[net info] = trainNetwork(XTraining,YTraining,layers,options);

the predictors must be a cell array, when I run it I receive this error:

Invalid training data. For sequence-to-one networks, training data must be cell arrays

So, when I convert XTraining from "double" to "cell", either by using "num2cell" or a loop, I receive then this error:

Error using trainNetwork
Invalid training data. Predictors and responses must have the same number of observations.

Error in HybridDNN5 (line 171)
[net info] = trainNetwork(XTraining,YTraining,layers,options);

The idea of the code is to train a DNN for a regression task with 5 predictors and 1 response.

[Amanjit Dulai]

The reason for using cell arrays is if you are training on multiple time series. For example, if you were trying to train a model to predict voltage from current for a machine, and you have data recorded from multiple different machines, you would use one cell for each time series from each machine. But it sounds like you only have one time series, in which case you should be able to train without a cell array.

Also, the error you received about "sequence-to-one" suggests your network might not be configured for the right problem. "Sequence-to-one" problems are where the input is a sequence but the output is not (for example, classifying an entire sequence is a sequence-to-one problem).

The code below shows how to train an LSTM for a simple sequence to sequence problem on a single time series:

[X, T] = maglev_dataset;

X = cell2mat(X);

T = cell2mat(T);

trainingFraction = 0.8;

validationFraction = 0.1;

numTimeSteps = size(X,2);

numTrain = floor(numTimeSteps*trainingFraction);

numValidation = floor(numTimeSteps*validationFraction);

numTest = numTimeSteps - numTrain - numValidation;

XTrain = X(:, 1:numTrain);

TTrain = T(:, 1:numTrain);

XValidation = X(:, (numTrain + 1):(numTrain + numValidation));

TValidation = T(:, (numTrain + 1):(numTrain + numValidation));

XTest = X(:, (numTrain + numValidation + 1):end);

TTest = T(:, (numTrain + numValidation + 1):end);

layers = [

sequenceInputLayer(1)

lstmLayer(20)

lstmLayer(20)

lstmLayer(20)

fullyConnectedLayer(1)

regressionLayer

];

options = trainingOptions('adam', ...

'MaxEpochs', 1000, ...

'ValidationData', {XValidation, TValidation}, ...

'Plots', 'training-progress');

net = trainNetwork(XTrain, TTrain, layers, options);

YTest = predict(net, XTest);

rmse = sqrt(mean((TTest - YTest).^2));