Formatting data for Deep Learning toolbox and the trainNetwork function

Hello,

I am struggling a lot with MATLAB trying to train a neural network to recognise gait events in kinematic walking data (sensors on patient's legs that give me their positions in 3D). My dataset (for now) is a 10661 by 96 matrix (10661 points with 96 features). The features are 3 positions (because it is in 3D) * 4 sensors * 2 for both feet * 4 different experiment conditions (where the patient walked with different speeds and inclination), which is equal to 96 features. My labels are inside a 96by1 vector with 1 for heel strike, -1 for toe off, 0 else (which are the only gait events I want to train to recognise for now). My code from that point:

% making a categorical array out of my labels array

labels_cat = categorical(labels,[-1 0 1],{'toe off' 'else' 'heel strike'});

numClasses = 3; % Number of classes in your problem

%% defining the neural net

inputSize = 3*4*2*4; % 96 features

hiddenSize = 100;

numClasses = 3; % the number of classes (heel strike, toe off, and other)

layers = [

imageInputLayer([1 1 inputSize], 'Name', 'input')

fullyConnectedLayer(50, 'Name', 'fc1')

reluLayer('Name', 'relu1')

fullyConnectedLayer(numClasses, 'Name', 'fc2')

softmaxLayer('Name', 'softmax')

classificationLayer('Name', 'classoutput')

];

%% training

% Specify training options

options = trainingOptions('adam', ...

'MaxEpochs', 20, ...

'MiniBatchSize', 32, ...

'InitialLearnRate', 0.001, ...

'Plots', 'training-progress');

net = trainNetwork(data, labels_cat, layers, options);

I have tried every imaginable way to format the data (especially the labels) and I always get an error. Either a supposed size mismatch between X and Y (when both are exactly 10661x1 and 10661x96), or something else. I might be doing everything wrong, I never touched MATLAB for ML before and I'm not an expert on pytorch / tensorflow either.

Thanks :)

Answer 1

[Sivylla Paraskevopoulou]

Hi Alexei,

You have a feature matrix but your network expects images as input data. The first layer of your network is an imageInputLayer.

You can classify your data with machine learning or deep learning, and I think you will find these two examples helpful:

• Human Activity Recognition Using Signal Feature Extraction and Machine Learning
• Sequence-to-Sequence Classification Using Deep Learning

If you choose deep learning, see in the above example how a very simple LSTM network is created for classifying sensor data:

layers = [ ...

sequenceInputLayer(numFeatures)

lstmLayer(numHiddenUnits,'OutputMode','sequence')

fullyConnectedLayer(numClasses)

softmaxLayer

classificationLayer];

[Alexei Ermochkine]

Thank you for your help! This works for me