Error using trainNetwork . Number of observations in X and Y disagree.

Good Evening,

I'm trying to train a neural network but "Error using trainNetwork . Number of observations in X and Y disagree." error came out. I am not able to understand where is the problem.

XTrain is a 607x39x367x1 double and loadYTrain is 607x1 double (before training i changed to categorical also.).

can you please help me to resolve this?. its urgent.

Thank you all in advance.

Here is my code:

layers = [

imageInputLayer([39 367 1])

convolution2dLayer(3, 32, 'Padding', 'same')

reluLayer

convolution2dLayer(3, 64, 'Padding', 'same')

reluLayer

maxPooling2dLayer(2, 'Stride', 2)

dropoutLayer(0.2)

fullyConnectedLayer(128)

reluLayer

dropoutLayer(0.2)

fullyConnectedLayer(64)

reluLayer

fullyConnectedLayer(15)

softmaxLayer

classificationLayer];

% Set training options

options = trainingOptions('adam', ...

'MaxEpochs', 20, ...

'MiniBatchSize', 16, ...

'InitialLearnRate', 0.0001, ...

'ValidationData', {test_data, test_labels_categorical}, ...

'Plots', 'training-progress');

train_labels_categorical = categorical(train_labels);

% Train the model

trained_model = trainNetwork(train_data, train_labels_categorical, layers, options);

Answer 1

[Katja Mogalle]

Hello,

The Deep Learning Toolbox in MATLAB expects in-memory image data to be presented as a h-by-w-by-c-by-N numeric array, where h, w, and c are the height, width, and number of channels of the images, respectively, and N is the number of images.

See https://www.mathworks.com/help/deeplearning/ref/trainnetwork.html?s_tid=doc_ta#mw_199f8be9-53f8-4bac-acdf-2846778f5904

If I understand your data correctly, your array is currently of size N-by-h-by-w-by-c. You can permute the training and test data to bring them in the correct format as follows:

train_data = permute(train_data,[2,3,4,1]);

I hope this helps.

Katja

[Gowri Prasood]

Thankyou for the help. but still same error. and my input data is mfcc features extracted from speech signals. and y_train is its labels.

[Katja Mogalle]

OK, let's have a closer look. This is the script I generated based on your description. After permuting the training and validation data, the code below runs without errors. Can you confirm that you have the same sizes of training and validation data? Thanks

% Generate some random training data
numObsTrain = 607;
train_data = rand(numObsTrain,39,367,1);
% Observations must be the 4th dimension, not the first
train_data = permute(train_data,[2,3,4,1]);
size(train_data) % this should return 39 367 1 607
train_labels = randi(15,numObsTrain,1);
train_labels_categorical = categorical(train_labels);

% Generate some random validation data
numObsTest = 100;
test_data = rand(numObsTest,39,367,1);
% Observations must be the 4th dimension, not the first
test_data = permute(test_data,[2,3,4,1]);
size(test_data) % this should return 39 367 1 100
test_labels = randi(15,numObsTest,1);
test_labels_categorical = categorical(test_labels);

% Set up the layers of the network
layers = [
    imageInputLayer([39 367 1])
    convolution2dLayer(3, 32, 'Padding', 'same')
    reluLayer
    convolution2dLayer(3, 64, 'Padding', 'same')
    reluLayer
    maxPooling2dLayer(2, 'Stride', 2)
    dropoutLayer(0.2)
    fullyConnectedLayer(128)
    reluLayer
    dropoutLayer(0.2)
    fullyConnectedLayer(64)
    reluLayer
    fullyConnectedLayer(15)
    softmaxLayer
    classificationLayer];

% Set training options
options = trainingOptions('adam', ...
    'MaxEpochs', 20, ...
    'MiniBatchSize', 16, ...
    'InitialLearnRate', 0.0001, ...
    'ValidationData', {test_data, test_labels_categorical}, ...
    'Plots', 'training-progress');

% Train the model
trained_model = trainNetwork(train_data, train_labels_categorical, layers, options);

[Gowri Prasood]

numObsTest = 151

[Gowri Prasood]

mfcc size is a csv file size = 758x14313 double. spliting data by using cvpartition , x_train=607x14313 and x_test = 151x14313. label is also a csv file size of 758x1. 15 unique labels.

[Katja Mogalle]

You mentioned initially, that "XTrain is a 607x39x367x1". Now you say that "x_train=607x14313". So I assume the data was somehow flattened to save it to file, but I don't know how exactly.

You'd need to reshape the data ack into a 4-D array which the convolutional neural network can interpret as height-by-width-by-channels-by-numObservations. But to to this, you need to figure out how the data was flattened. I suspect you'd need one of the following commands to unflatten the data:

x_train = reshape(x_train,[607,39,367,1])

or

x_train = reshape(x_train,[607,367,39,1])

Then, don't forget to permute the observation dimension into the fourth dimension as I showed earlier.

As you are using a 2D convolutional neural network, you have to decide which dimensions of your data represent the two "spatial" dimensions and which ones represents the "channel/feature" dimension. I suspect your data only has one dimension that could be interpreted as "space" (or time). So another option you could have a look at, are 1D CNNs or even recurrent networks. Here is an example showing speech emotion recognition based on gtcc coefficients and which uses a recurrent neural network: https://www.mathworks.com/help/audio/ug/speech-emotion-recognition.html

[Gowri Prasood]

thankyou for the help. now i am able to train the network but. validation accuracy is very low and validation loss is very high.