LSTM error with number of X and Y observations

I am using lstm regression network to denoise speech. The predictor input consists of 9 consecutive noisy STFT vectors. The target is corresponding clean STFT vector. The length of each vector is 129.

Here 's the network I defined:

layers = [

sequenceInputLayer([129 9 1],"Name","sequence")

flattenLayer("Name","flatten")

lstmLayer(128,"Name","lstm")

fullyConnectedLayer(129,"Name","fc_1")

reluLayer("Name","relu")

fullyConnectedLayer(129,"Name","fc_2")

regressionLayer("Name","regressionoutput")];

I trained the network with X and Y of sizes:

size(X):

129 9 1 254829

size(Y):

129 254829

I got the error "Invalid training data. X and Y must have the same number of observations". I think that maybe the network I defined is wrong. I am new with lstm network to do sequence-to-sequence regression. What should I do with my network or training data?

Thanks for your help!

Answer 1

[jibrahim]

Hi Daniel,

Please refer to the help of trainNetwork for info on correct input sizes for sequences.

The inputs must be cell arrays. Here is an example with two observations:

opts = trainingOptions('adam');
trainNetwork({randn(129,9,1) , randn(129,9,1)}, {randn(129,1),randn(129,1)},layers,opts) 

[Daniel Guo]

It works! Thank you a lot :)