How to use my trained LSTM model to test the remaining unused normal signals ?

I want to use my trained LSTM model to test the remaining unused normal signals to see the performance. How can I change the code to test my signals (normals without labels) in my code? I have changed the ratio now but is it another way? besides when I want to start read the unused files for instance from i=3600 when I change the following code (change i to start to read from 3600 until 5600 for 2000 files) I don't get the proper result for classification how can I read from specific number of files then?

clear all
close all

TrainRatio=0.1;
ValidationRatio=0.8;

folder='/Users/pooyan/Desktop/cmms/class/normal/';  % change this path to your normal data folder
audio_files=dir(fullfile(folder,'*.ogg'));
nfileNum=length(audio_files);
nfileNum=30
normal=[];
for i = 1:nfileNum
    normal_name = [folder audio_files(i).name]; 
    normal(i,:) = audioread(normal_name);   
end 
normal=normal';
nLabels = repelem(categorical("normal"),nfileNum,1);

folder='/Users/pooyan/Desktop/cmms/class/anomaly/'; % change this path to your anomaly data folder
audio_files=dir(fullfile(folder,'*.ogg'));
afileNum=length(audio_files);
afileNum=30

anomaly=[];
for i = 1:afileNum 
    anomaly_name = [folder audio_files(i).name]; 
    anomaly(i,:) = audioread(anomaly_name);   
end 
anomaly=anomaly';
aLabels = repelem(categorical("anomaly"),afileNum,1);

% randomize the inputs if necessary
% normal=normal(:,randperm(nfileNum, nfileNum));
% anomaly=anomaly(:,randperm(afileNum, afileNum));

nTrainNum = round(nfileNum*TrainRatio);
aTrainNum = round(afileNum*TrainRatio);
nValidationNum = round(nfileNum*ValidationRatio);
aValidationNum = round(afileNum*ValidationRatio);

audioTrain = [normal(:,1:nTrainNum),anomaly(:,1:aTrainNum)];
labelsTrain = [nLabels(1:nTrainNum);aLabels(1:aTrainNum)];

audioValidation = [normal(:,nTrainNum+1:nTrainNum+nValidationNum),anomaly(:,aTrainNum+1:aTrainNum+aValidationNum)];
labelsValidation = [nLabels(nTrainNum+1:nTrainNum+nValidationNum);aLabels(aTrainNum+1:aTrainNum+aValidationNum)];

audioTest = [normal(:,nTrainNum+nValidationNum+1:end),anomaly(:,aTrainNum+aValidationNum+1:end)];
labelsTest = [nLabels(nTrainNum+nValidationNum+1:end); aLabels(aTrainNum+aValidationNum+1:end)];

fs=44100;

%  Create an audioFeatureExtractor object 
%to extract the centroid and slope of the mel spectrum over time.
aFE = audioFeatureExtractor("SampleRate",fs, ...    %Fs
    "SpectralDescriptorInput","melSpectrum", ...
    "spectralCentroid",true, ...
    "spectralSlope",true);

featuresTrain = extract(aFE,audioTrain);
[numHopsPerSequence,numFeatures,numSignals] = size(featuresTrain);
numHopsPerSequence;
numFeatures;
numSignals;
%treat the extracted features as sequences and use a
%sequenceInputLayer as the first layer of your deep learning model. 

featuresTrain = permute(featuresTrain,[2,1,3]); %permute switching dimensions in array
featuresTrain = squeeze(num2cell(featuresTrain,[1,2]));%remove dimensions

numSignals = numel(featuresTrain); %number of signals of normal and anomalies

[numFeatures,numHopsPerSequence] = size(featuresTrain{1});


%Extract the validation features.

featuresValidation = extract(aFE,audioValidation);
featuresValidation = permute(featuresValidation,[2,1,3]);
featuresValidation = squeeze(num2cell(featuresValidation,[1,2]));



%Define the network architecture.

layers = [ ...
    sequenceInputLayer(numFeatures)
    lstmLayer(50,"OutputMode","last")
    fullyConnectedLayer(numel(unique(labelsTrain))) %%labelTrain=audio
    softmaxLayer
    classificationLayer];

%To define the training options

options = trainingOptions("adam", ...
    "Shuffle","every-epoch", ...
    "ValidationData",{featuresValidation,labelsValidation}, ... %%labelValidatin=audioValidation
    "Plots","training-progress", ...
    "Verbose",false);

%To train the network

net = trainNetwork(featuresTrain,labelsTrain,layers,options);


%Test the network %10 preccent 


%classify(net,permute(extract(aFE,audioTest),[2 257 35]))

TestFeature=extract(aFE, audioTest);
for i=1:size(TestFeature, 3)

TestFeatureIn = TestFeature(:,:,i)';
classify(net,TestFeatureIn)
predict(i) = classify(net,TestFeatureIn);

end

plotconfusion(labelsTest,predict')

Answer 1

[Anshika Chaurasia]

Hi Pooyan,

It is my understanding that you have trained your LSTM sucessfully and you are facing problem while testing.

You want to test the trained LSTM on signals i.e. normals without labels.

Let's suppose you have total 5600 signals and out of which you will test on 2000 signals (i = 3600 to 5600) and train/validate on 3600 signals (i = 1 to 3600). All the signals i.e. normals are stored in variable normal.

I am suggesting following code based on the information provided by you:

nTrainNum = round(3600*TrainRatio);
aTrainNum = round(3600*TrainRatio);

nValidationNum = round(3600*ValidationRatio);
aValidationNum = round(3600*ValidationRatio);

audioTrain = [normal(:,1:nTrainNum),anomaly(:,1:aTrainNum)];
labelsTrain = [nLabels(1:nTrainNum);aLabels(1:aTrainNum)];
audioValidation = [normal(:,nTrainNum+1:nTrainNum+nValidationNum),anomaly(:,aTrainNum+1:aTrainNum+aValidationNum)];
labelsValidation = [nLabels(nTrainNum+1:nTrainNum+nValidationNum);aLabels(aTrainNum+1:aTrainNum+aValidationNum)];
audioTest = [normal(:,3601:5600),anomaly(:,3601:5600)];
labelsTest = [nLabels(3601:5600); aLabels(3601:5600)];

.....
.....

% Extract test feature
TestFeature=extract(aFE, audioTest);
TestFeature = permute(TestFeature,[2,1,3]);
TestFeature = squeeze(num2cell(TestFeature,[1,2]));

predTest = classify(net,TestFeature);

You can also refer to following documents:

https://www.mathworks.com/help/deeplearning/ug/classify-sequence-data-using-lstm-networks.html#ClassifySequenceDataUsingLSTMNetworksExample-5

Classify Gender Using LSTM Networks - MATLAB & Simulink (mathworks.com)

Hope it helps!