Neural Network Data division

Hi all,

I am trying to build up a NN model in Matlab 2008a.

I am dividing the dataset using the following command

net.divideParam.trainRatio = 0.6;

net.divideParam.valRatio = 0.2;

net.divideParam.testRatio = 0.2;

Now after training, I want to have those datasets (_i.e._ target and network output only) that were used for training, testing and validation in three different files so that I can manually do the regression analysis. Can anyone please suggest the suitable command to so this.

Thanks for your kind help.

[Greg Heath]

What do you mean by "regression analysis" ?

What quantities do you want that are not available via the output of TRAIN ?

Greg

[Arijit]

Hi Greg,

By regression analysis, I want to mean that after performing TRAIN, I can see the plots of target vs output for TRAIN, TEST and VALIDATION datasets. I guess those values (target and output) are getting stored at "tr". Now I want to have those values extracted from "tr" to three different files and plot that graph manually. I want to do it just because definitely there will be some values for "train dataset" which will be out of the user defined performance limit. To delete or identify those, I want to have this information.

Thanks for your information and help. Looking forward for your valuable comment.

Arijit

[Arijit]

Hi Greg,

I got the answer of my previous question. If my output is "Y", then we can get the train output from "Y(tr.trainInd)".

Thanks

Arijit

Answer 1

[Luca Cavazzana]

after calling train you get as second input argument the training record tr. The attribute testMask{1} contains a matrix (the same size of the input matrix) whose value are 1 if the element is in the training set, NaN otherwise. You can find indices this way:

[net, tr] = train(net, input, targets);
testInd = find(~isnan(tr.testMask{1}(1,:)));

inputTest = input(:,testInd);
targetTest = target(:,testInd);

In the same way, trainMask{1} and valMask{1} gives you the training and validation mask.

edit: instead of the whole testInd=find(...) thing, I forgot there's already the attribute tr.testInd

[Arijit]

Thanks Luca.

I did using tr.trainInd and got the desired output.

Arijit