Error using adapt in Neural Networks

I am trying to use Haar Wavelet Decomposition and feed it into a Neural Network(NARX) in order to predict the coefficients. I have tried the same code with simple inputs and it works just fine. However, when I enter the coefficients from the 'Haar' wavelet, it gives an error saying "Performance PERF is not finite". Below is the code and the corresponding error.

Code

=====

clc CSV_PATH=''; CSV_FILE = 'test2.csv'; ROWS = 931;

%Reading Input Data

[numData,textData,row] = xlsread(strcat(CSV_PATH,CSV_FILE));

for i=3:ROWS

        data(i-2) = (row(i,1))   ;

end

CV = cell2mat(data);

%Wavelet Decomposition

[C,L] = wavedec(CV,1,'haar');

%Neural Networks

inputSeries = tonndata((1:465),true,false);

targetSeries = tonndata(C(1:465),true,false);

% Create a Nonlinear Autoregressive Network with External Input

inputDelays = 1:2;

feedbackDelays = 1:2;

hiddenLayerSize = [4 4 4];

net = narxnet(inputDelays,feedbackDelays,hiddenLayerSize);

[Xs,Xi,Ai,Ts] = preparets(net,inputSeries,{},targetSeries);

[net,a,e,pf] = adapt(net,Xs,Ts,Xi,Ai);

Error

=====

??? Error using ==> mse at 238

Performance PERF is not finite.

Error in ==> adaptwb>adapt_network at 94

      gE(net.outputConnect,:) = feval(net.performFcn,'dperf_de',net,T(:,ts),

      ...

Error in ==> adaptwb at 39

    [out1,out2,out3] = adapt_network(in1,in2,in3,in4);

Error in ==> network.adapt at 121

[net,Ac,tr] = feval(adaptFcn,net,Pd,T,Ai);

Error in ==> test_haar at 36

[net,a,e,pf] = adapt(net,Xs,Ts,Xi,Ai);

Can someone help me to identify the problem?

Answer 1

[Greg Heath]

One hidden layer is sufficient

What were your simple inputs ?

Why not test with a MATLAB data set so that we can compare answers ?

What happens when you use train instead of adapt ?

 help nndatasets

 help narxnet

 close all, clear all, clc

     [X,T]                 = simplenarx_dataset;
     net                   = narxnet(1:2,1:2,10);
     [ Xs , Xi ,Ai ,Ts ]   = preparets(net, X,{},T);
     ts                    = cell2mat(Ts);
     MSE00s                = mean(var(ts',1))  %0.0992 Reference MSE

     rng('default')
     [net tr Ys Es Xf Af ] = train( net, Xs, Ts, Xi, Ai);
     NMSEs                 = mse(cell2mat(Es))/MSE00s   % 1.4394e-07

    rng('default')
    [net Ys Es Xf Af tr ] = adapt( net, Xs, Ts, Xi, Ai);
     NMSEs                 = mse(cell2mat(Es))/MSE00s   % 1.5749e-07

Hope this helps.

Thank you for formally accepting my answer

Greg

PS The default H = 10 can be reduced

[Koustubh Gaikwad]

Thanks for the reply Greg.

What were your simple inputs ?

--> I tried it with simple inputs like (1:500) and it worked. However, when I tried (1:1000) it somehow gave me the same error. I also tried with a more realistic dataset and it happened to work for that as well. However, only for certain inputs I get this error.

Why not test with a MATLAB data set so that we can compare answers ?

-->As I mentioned, it works on certain inputs

What happens when you use train instead of adapt ?

-->train works instead of adapt.

Also, I managed to run it for the same set of inputs by first dividing the entire dataset by 10 and then after prediction I multiple by 10. It may not be a good way to do this.

[Greg Heath]

I don't understand.

 input  = 1:500
 target = ?

Again, use a MATLAB timeseries data set

 help nndatsets