MATLAB Answers

How to save a neural network to test on a new dataset?

13 views (last 30 days)
Joana on 16 Dec 2019
Commented: Dheeraj Singh on 23 Dec 2019
I am using the following code to train and test NN for 2-class classification. I need to save the trained network to test on a diffreent data set. I tried the save net command, but it just saved the results and not the trained model.
Can nayone please help to get that.?
load iris.mat; % Matlab also provides this dataset (load fisheriris.mat)
% Call features & labels
feat=f; label=l;
% Programmer: Jingwei Too
function NN=jNN(feat,label,kfold,Hiddens,Maxepochs)
% Layer
if length(Hiddens)==1
h1=Hiddens(1); net=patternnet(h1);
elseif length(Hiddens)==2
h1=Hiddens(1); h2=Hiddens(2); net=patternnet([h1 h2]);
elseif length(Hiddens)==3
h1=Hiddens(1); h2=Hiddens(2); h3=Hiddens(3);
net=patternnet([h1 h2 h3]);
% rng('default');
% Divide data into k-folds
% Pre
pred2=[]; ytest2=[]; Afold=zeros(kfold,1);
% Neural network start
for i=1:kfold
% Call index of training & testing sets; testIdx=fold.test(i);
% Call training & testing features and labels
xtrain=feat(trainIdx,:); ytrain=label(trainIdx);
xtest=feat(testIdx,:); ytest=label(testIdx);
% Set Maximum epochs
net.trainParam.epochs= Maxepochs;
% to prevent early stopping
net.trainParam.max_fail = 500;
net.trainParam.min_grad = 0.000000000000001;
net.plotFcns = {'plotperform','plottrainstate','ploterrhist', ...
'plotconfusion', 'plotroc'};
% Training model
% Perform testing
% Confusion matrix
% Get accuracy for each fold
% Store temporary result for each fold
pred2=[pred2(1:end,:),pred]; ytest2=[ytest2(1:end);ytest];
% Overall confusion matrix
save net
[~,confmat]=confusion(dummyvar(ytest2)',pred2); confmat=transpose(confmat);
% Average accuracy over k-folds
% Store results
NN.fold=Afold; NN.acc=acc; NN.con=confmat;
fprintf('\n Classification Accuracy (NN): %g %%',acc);
% figure, plotperform(tr)
%figure, plottrainstate(tr)
% figure, ploterrhist(e)
% figure, plotconfusion(ytest2,pred)
% figure, plotroc(Labels,y)


Sign in to comment.

Accepted Answer

Dheeraj Singh
Dheeraj Singh on 20 Dec 2019
save will generally save all the variables present in the workspace at point of time.
So, instead of saving the model inside the function, you can return the model
function [net,NN]=jNN(feat,label,kfold,Hiddens,Maxepochs)
and then use the save command
save net
Also, if you trying to use the iris dataset in MATLAB use iris.dat
load iris.dat
feat=iris(:,1:4); label=iris(:,5);
I used the following code and it is working for me:
feat=iris(:,1:4); label=iris(:,5);
% Programmer: Jingwei Too
[net,NN]=jNN(feat,label,5,[10 10 10],10000)
save net


Joana on 20 Dec 2019
Hi Dheeraj,
Thanks for your answer.
I have one more question regarding data preparation for the Input. I have EEG data for 2-classes, recorded at 1200hz, with 32 EEG channels. i have extracted each class for 1 second for 100 trials. So the data is in the format of number of channels x sampling frequency x trials = 32x1200x200.
I tried the above program by converting the data to 2D as: 38,400x200.
SO input layers neurons are 38,400. and 3 hidden layers with neurons [10 10 10], one output neuron.
Is it the right way to do it.? or should i try something else.? (This gives acceptable results though but i am confused if inout neurons can be this much)
I tried using cell array function but it gives an error that 'Data distribution doesn't have equal number of time steps'.
I'll highly appreciate if you can guide me on this one. :)

Sign in to comment.

More Answers (0)

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by