trainNetwork - training data format

Hi,

I'm trying to build a concolutional neural network, but I don't get how to prepare training data.

I've got 24 mfcc features as an input and 20 output classes, so as far as I understand I need featureInputLayer and convolution1dLayer.

But when I'm using featureInputLayer I'm getting

"Error using trainNetwork - Invalid network.

(...)

Caused by:

Layer 2: Input data must have one spatial dimension only, one temporal dimension only, or one of each.

Instead, it has 0 spatial dimensions and 0 temporal dimensions."

How to fix this?

my simplified layers:

layers = [ ...

featureInputLayer(numFeatures)

convolution1dLayer(filterSize, numFilters, 'Padding', 'same')

reluLayer

averagePooling1dLayer(2, 'Stride', 2)

fullyConnectedLayer(300)

reluLayer

fullyConnectedLayer(numOutputs)

softmaxLayer

classificationLayer];

My data structure:

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

XTrain -> Nx24 double

Ytrain -> Nx1 categorical

To be honest I've tested all combinations of XTrain and YTrain transpositions, but still the same issue .

I've also tried using imageInputLayer as below, but then I've received another error

"Error using trainNetwork - Number of observations in X and Y disagree"

layers = [ ...

imageInputLayer([numFeatures 1])

convolution2dLayer([filterSize 1], numFilters, 'Padding', 'same')

reluLayer

averagePooling2dLayer([2 1], 'Stride', 2)

fullyConnectedLayer(300)

reluLayer

fullyConnectedLayer(numOutputs)

softmaxLayer

classificationLayer];

Please help me finding the right data structure, because the description provided in help is unclear to me.

Thanks,

Zbigniew

PS. versions of Matlab, I'm working on: R2021b, R2023a

[NAVNEET NAYAN]

what is the size of your feature?

Answer 1

[Matt J]

With your imageInputLayer network, try,

XTrain =reshape(XTrain.', 24,1,1,N);

so that the format of the input is SxSxCxB.