Troubles in prediction using LSTM

I want to make a sequence-to-sequence regression using LSTM.

The training progress showed the convergence of RMSE and Loss to nearly zero.

However, the prediction is very bad, although I use the training data for test.

Whould you recommand what can I do to modify my program?

I do not know what the problem is. (I tried to change 'numHiddenUnits' and number of 'lstmLayer')

I hope your help.

My program is as follow;

%-------------------------------------------------------------------

% The size of Xtrain is 5x1 cell, and the size of each cell is 300x25000.

% The size of Ytrain is 5x1 cell, and the size of each cell is 3x25000.

numResponses = size(Ytrain{1},1); % 3

featureDimension = size(Xtrain{1},1); % 300

numHiddenUnits = 500;

layers = [ ...

sequenceInputLayer(featureDimension)

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

fullyConnectedLayer(50)

dropoutLayer(0.5)

fullyConnectedLayer(numResponses)

regressionLayer];

maxEpochs = 60;

miniBatchSize = 20;

options = trainingOptions('adam', ...

'MaxEpochs',maxEpochs, ...

'MiniBatchSize',miniBatchSize, ...

'InitialLearnRate',0.01, ...

'GradientThreshold',1, ...

'Shuffle','never', ...

'Plots','training-progress',...

'Verbose',0);

%-----------Training and Test--------------------

net = trainNetwork(Xtrain,Ytrain,layers,options);

h = predict(net,Xtrain,'MiniBatchSize',1);

[nahed zemouri]

i have the same probleme please any one can help me

Answer 1

[Hong Gi Yeom]

I found a problem with my code.

I write the answer for someone who will have the same problem.

To train the neural network, you have to normalize both input and output data.

I did not normalize the output data.

I hope this message can help someone.

[Salma Matoussi]

Hi, could you please explain how to normalize inputs and outputs with matlab for LSTM ? Thanks in advance

[Hong Gi Yeom]

If you want to normalize the 'sig' data, you can used following code. The 'n_sig' is the normalized data m_sig = mean(sig); v_sig = std(sig); n_sig = (sig-m_sig)/v_sig;

[nahed zemouri]

if you want normalise your data you can use this function

function dataout = scaledata(datain,minval,maxval)

%

% Program to scale the values of a matrix from a user specified minimum to a user specified maximum

%

% Usage:

% outputData = scaleData(inputData,minVal,maxVal);

%

% Example:

% a = [1 2 3 4 5];

% a_out = scaledata(a,0,1);

%

% Output obtained:

% 0 0.1111 0.2222 0.3333 0.4444

% 0.5556 0.6667 0.7778 0.8889 1.0000

dataout = datain - min(datain(:));

dataout = (dataout/range(dataout(:)))*(maxval-minval);

dataout = dataout + minval;