Time Delay NN with Early Stopping (Validation)
in [Matlab]
|
Prev: computation time sensitivity to initial value
Next: Is Matlab "libstruct" able to recognize the nested-C structure?
From: Jorge Pasquotto on 23 Apr 2010 15:58 Dear Perez and Norberto, I am having the same problem with Elman networks. I simply cannot get a cross-validation training because all the data set is understood as trainiing set, and no validation interval is taken. (nor test set). Have you found a new solution? Regards "Perez Jose" <fennics(a)gmail.com> wrote in message <gev90g$mni$1(a)fred.mathworks.com>... > Hi!!! > > I don't find the way for use Validation with the FTDNN, and it's important and needed in the task of forecasting. > > Im using R2008a > > This example is taken from the Help of Focused Time-Delay Neural Network (newfftd). You can see there this example explained. > ------------------------------------ > load laser > y = y(1:600)'; > y = con2seq(y); > > ftdnn_net = newfftd(y,y,[1:8],5); > ftdnn_net.trainParam.show = 10; > ftdnn_net.trainParam.epochs = 50; > > p = y(9:end); > t = y(9:end); > Pi=y(1:8); > ftdnn_net = train(ftdnn_net,p,t,Pi); > > yp = sim(ftdnn_net,p,Pi); > yp = cell2mat(yp); > e = yp-cell2mat(t); > rmse = sqrt(mse(e)) > ------------------------------------ > > Well, by default the divideFcn is 'dividerand', so the training should use validation for early stopping. But IT DOESN'T. > You can check it in the parameters of TR if you train like this > > [ftdnn_net, TR] = train(ftdnn_net,p,t,Pi); > > It doesn't divide, even it should. AND it doesn't display any errors. It does if you change it to 'divideblock' > > ftdnn_net.divideFcn ='divideblock'; > > > 1? Solution: Dont make con2seq(y), I mean, train with Matrix and not with cells. Then, it WORKS.. > > But, if you train with matrix..then I don't know how to give the parameter Pi, because it ask me that it needs to have more columns. ??. And it says in the help of 'train' that it's not recommended neither when it's several Time Steps Ahead. > > CONCLUSION Can someone help me to or give me an idea of using training with Validation AND training with cells? > > Should I write my own divide? > > THANKSSSS |