Why do I get a channel mismatch error between predictions and targets when using "trainnet"?

I am trying to train a neural network using the "trainnet" function. My setup involves splitting a dataset into training and test sets, and then training a network with the following relevant code:

% Split data into training and test sets[idxTrain, idxTest] = trainingPartitions(size(myDataTable, 1), [0.85, 0.15]);trainData = myDataTable(idxTrain, :);
% Prepare training datapredictorNames = ["feature1", "feature2", "feature3", "feature4"];XTrain = table2array(trainData(:, predictorNames));TTrain = trainData.targetColumn;
% Define network layerslayers = [    featureInputLayer(numel(predictorNames), Normalization="zscore")    fullyConnectedLayer(16)    reluLayer    fullyConnectedLayer(2)    softmaxLayer];
% Train networkoptions = trainingOptions("lbfgs", ExecutionEnvironment="cpu", Plots="training-progress", Verbose=false);net = trainnet(XTrain, TTrain, layers, "crossentropy", options);
However, I encounter the following error:

Error using trainnet
Number of channels in predictions (2) must match the number of channels in the targets (1).

Why is this happpening?

Answer 1

[MathWorks Support Team]

This issue arises due to a mismatch between the expected format of the target data and what is provided to the "trainnet" function.

• The error indicates that the network's output layer is configured for 2 classes.
• The target data "TTrain" is currently a 431x1 double, and appears to be formatted for only 1 class, but this should be a categorical.

Please find the below workarounds:

1. If you are in R2024a or earlier, convert "TTrain" from type "double" to a categorical array to match the expected input format:

TTrain = categorical(TTrain);

2. Specify feature data directly into "trainnet" as a table, with targets as the last column:

trainData.targetColumn = categorical(trainData.targetColumn);dataTrain = [trainData(:, predictorNames) trainData(:, "targetColumn")];numFeatures = width(dataTrain) - 1;net = trainnet(dataTrain, layers, "crossentropy", options);

3. In R2024b, new cross entropy loss functions were introduced to avoid the need for converting to categoricals.

• "index-crossentropy": for multiclass classification:

net = trainnet(dataTrain, layers, "index-crossentropy", options);

• "binary-crossentropy": if the problem is a binary classification, modify the network to end with a "sigmoidLayer" and use "binary-crossentropy":

layers = [    featureInputLayer(numel(predictorNames), Normalization="zscore")    fullyConnectedLayer(16)    reluLayer    fullyConnectedLayer(1)    sigmoidLayer];net = trainnet(dataTrain, layers, "binary-crossentropy", options);