How to implement cross validation with back propogation network

Sir, How to implement cross validation methods such as k fold and leave one out with back propogation network... i have tried with SVm works good.. but dont know how to merge k fold with bpn... .. thanks

[Subha]

i've tried this code...

    load dataset4_bp_kn_fs
TrainingSet=data;
GroupTrain=target;
Indices = crossvalind('Kfold',GroupTrain , 10);
for i=1:10
     test = (Indices == i); train = ~test;
net = newff(TrainingSet(train,:),GroupTrain(train,:),20,{},'trainscg');
[net,TR] = traingd(net,TrainingSet(train,:),TrainingSet(test,:))
 a = sim(net,TrainingSet(train,:));
     b=sim(net,TrainingSet(test,:));
end

where, data is 16 x 54 and target is 1x54 i'm getting error as, ??? Index exceeds matrix dimensions. and

??? Error using ==> network.subsref at 83 Reference to non-existent field 'lr'.

Error in ==> traingd at 141 lr = net.trainParam.lr; ..

i've made few trials too like setting the target as 3x54 matrix but dono how to proceed with this... really in a confused state..

Answer 1

[Tom Lane]

I am not a nnet expert, but I am under the impression that your inputs should have one column per observation (rather than one row as in the Statistics Toolbox). If that is the case you may need to use "train" and "test" to index into columns rather than rows. Also, I believe traingd wants training set target values as its third input, not X data for the test set.

[Subha]

Sir with your piece of advice i've done few modification.. like (test,:) as (:,test), now its working good , Accuracy seem to be low, have to try some thing to improve it... but really Matlab is like an ocean.. i've to learn lots more... .. Thank you Sir..

Answer 2

[laplace laplace]

how did you apply the crossvalind command to column vectors??

[laplace laplace]

generaly if your data have a dimension how do you apply the crossvalind command?