Why am I getting an Input data size error using "arrayDatastore" with "trainNetwork" (Deep Learning Toolbox)?

I am using the Deep Learning Toolbox in R2023a and wanted to train a very simple network with one 2-dimensional feature input and one regression output. The training data is provided in form of a combined "arrayDatastore" which has two columns. The first column has 2D-row vectors in each row, for the input features.

 

%% generate data
inData = randn(1000,3);
xTrain = arrayDatastore(inData(:,1:2)); 
yTrain = arrayDatastore(inData(:,3));
dataTrain = combine(xTrain,yTrain);

%% define network
layers = [
    featureInputLayer(2,"Name","featureinput")
    fullyConnectedLayer(20,"Name","fc_1")
    reluLayer("Name","relu1")
    fullyConnectedLayer(20,"Name","fc_2")
    reluLayer("Name","relu2")
    fullyConnectedLayer(20,"Name","fc_3")
    reluLayer("Name","relu3")
     fullyConnectedLayer(1,"Name","fc_4")
    regressionLayer("Name","regressionoutput")];
lgraph = layerGraph(layers);

%% train network
options = trainingOptions('adam');
options.MaxEpochs = 1;
netTest = trainNetwork(dataTrain,layers,options);

However, I received the error message:

 

Error using trainNetwork
Input datastore returned more than one observation per row for network input 1.

Error in trainNetwork_ds_error (line 26)
netTest = trainNetwork(dataTrain,layers,options);

How can I create a compatible datastore from my simple input?

Answer 1

[MathWorks Support Team]

Inside the documentation for using datastores in Deep Learning, it is written that the format of predictors for feature input data should be a column vector c-by-1, if c is the number of input features:

https://www.mathworks.com/help/releases/R2023a/deeplearning/ug/datastores-for-deep-learning.html

To workaround this, you can transpose the input data using the (') operator and change the iteration dimension in the "arrayDatastore". For example:

xTrain = arrayDatastore(inData(:,1:2)', "IterationDimension", 2);