Invalid training data. The output size (4) of the last layer does not match the number of classes (5).How to match the size for neural network?

clear all;

load emg_E1.mat;

for i = 1:24

namestr = ['s' num2str(i) '=','emg_E1(4441:504440,i)',';'];

eval(namestr);

end

rowDist=ones(1,160);

rowDist=3125*rowDist;

for i=1:24

namestr = ['c' num2str(i) '=','mat2cell','(','s' num2str(i) ',' 'rowDist',')',';'];

eval(namestr);

end

for i=1:24

for j=1:160

namestr = ['m' num2str(j) '=','c' num2str(i),'{',num2str(j),'}',';'...

'm' num2str(j) '=','m' num2str(j),'''',';'...

'c' num2str(i) '{',num2str(j),'}' '=','m' num2str(j) ';'];

eval(namestr);

end

end

Signals=[c1;c2;c3;c4;c5;c6;c7;c8;c9;c10;c11;c12;c13;c14;c15;c16;c17;c18;c19;c20;c21;c22;c23;c24;];

Labels=zeros(3840,1);

Labels=categorical(Labels);

Labels(1:960)='1';Labels(961:1920)='2';Labels(1921:2880)='3';Labels(2881:3840)='4';

XTrain=Signals;YTrain=Labels;

%=====定义 LSTM 网络架构=====%

inputSize = 1;

numHiddenUnits = 100;

numClasses = 4;

layers = [ ...

sequenceInputLayer(inputSize)

bilstmLayer(numHiddenUnits,'OutputMode','last')

fullyConnectedLayer(numClasses)

softmaxLayer

classificationLayer]

maxEpochs = 100;

miniBatchSize = 3125;

options = trainingOptions('adam', ...

'MaxEpochs',10, ...

'MiniBatchSize', 150, ...

'InitialLearnRate', 0.001, ...

'SequenceLength', 625, ...

'GradientThreshold', 1, ...

'ExecutionEnvironment',"auto",...

'plots','training-progress', ...

'Verbose',false);

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

Answer 1

[Soumya]

Hi @龙春 谢,

The ‘categorical’ data type in MATLAB is used to represent data that can take on a fixed number of discrete categories. Initially, ‘Labels’ is created as a numeric vector of ‘zeros’ with length 3840. When it is immediately converted to a categorical array, a single category ('0') is created based on the initial numeric values. Later, when the entries in ‘Labels’ are reassigned with the string values ('1', '2', '3', and '4'), they get added as new categories. As a result, the ‘Labels’ array ends up having five categories.

This mismatch causes the error when training the neural network, which is configured for only four output classes. The presence of an unintended fifth category ('0') causes the output layer to be incompatible with the label data.

To resolve this issue, assign the labels before converting the array to categorical, as shown below:

Labels = zeros(3840,1); 
Labels(1:960) = 1;  
Labels(961:1920) = 2; 
Labels(1921:2880) = 3; 
Labels(2881:3840) = 4; 
Labels = categorical(Labels); 

With this approach, the number of categories in the categorical array matches the number of defined output classes.

Please refer to the following documentation to know more about the ‘categorical’ function:

https://www.mathworks.com/help/matlab/ref/categorical.html

I hope this helps.