neuralNetTrain

 neuralNetTrain : trains a neural network to predict different classes.
 Solve a Pattern Recognition Problem with a Neural Network
 Script generated by Neural Pattern Recognition app

 INPUT : 
   X : input data.
   Y : target data.
   numCanidates : number of times to repeat training

 OUTPUT : 
   net : struct with trained neural network.
   fun : the function used for obtaining the scores with the neural network. 

 Script was generated by Neural Pattern Recognition app
 University of Washington, 2016
 This file is part of SuperSegger.

0001 function [net,fun] = neuralNetTrain (X, Y, numCanidates)
0002 % neuralNetTrain : trains a neural network to predict different classes.
0003 % Solve a Pattern Recognition Problem with a Neural Network
0004 % Script generated by Neural Pattern Recognition app
0005 %
0006 % INPUT :
0007 %   X : input data.
0008 %   Y : target data.
0009 %   numCanidates : number of times to repeat training
0010 %
0011 % OUTPUT :
0012 %   net : struct with trained neural network.
0013 %   fun : the function used for obtaining the scores with the neural network.
0014 %
0015 % Script was generated by Neural Pattern Recognition app
0016 % University of Washington, 2016
0017 % This file is part of SuperSegger.
0018 
0019 
0020 fun = @scoreNeuralNet;
0021 
0022 if ~exist('numCanidates','var') || isempty(numCanidates)
0023     numCanidates = 1;
0024 end
0025 
0026 t = [(Y == 0),Y]';
0027 x = X';
0028 
0029 
0030 
0031 % Create a Pattern Recognition Network
0032 % hiddenLayerSizes
0033 net = patternnet([10 15]);
0034 
0035 % Choose a Training Function
0036 % For a list of all training functions type: help nntrain
0037 % 'trainlm' is usually fastest.
0038 % 'trainbr' takes longer but may be better for challenging problems.
0039 % 'trainscg' uses less memory. Suitable in low memory situations.
0040 net.trainFcn = 'trainbr';  % Scaled conjugate gradient backpropagation.
0041 
0042 
0043 % Setup Division of Data for Training, Validation, Testing
0044 net.divideParam.trainRatio = 70/100;
0045 net.divideParam.valRatio = 15/100;
0046 net.divideParam.testRatio = 15/100;
0047 
0048 %Make sure the overfit is real.
0049 net.trainParam.max_fail = net.trainParam.epochs / 10;
0050 
0051 numTotal = numel(t(1,:));
0052 numTrue = numel(find(t(1,:)==0));
0053 numFalse = numel(find(t(1,:)==1));
0054 
0055 errorWeights = {[(t(1,:)==0) * 0 + (t(1,:)==1) * numTotal / numFalse; (t(2,:)==0) * 0 + (t(2,:)==1) * numTotal / numTrue]};
0056 
0057 
0058 
0059 % Train the Network
0060 canidates = {};
0061 for i = 1:numCanidates
0062     [canidates{i}] = train(net,x,t,{},{},errorWeights);
0063 end
0064 
0065 % Test the Network
0066 % y = net(x);
0067 % e = gsubtract(t,y);
0068 % performance = perform(net,t,y);
0069 tind = vec2ind(t);
0070 percentError = [];
0071 for i = 1:numCanidates
0072     net = canidates{i};
0073     y = net(x);
0074     yind = vec2ind(y);
0075     percentError(i) = sum(tind ~= yind)/numel(tind);
0076 end
0077 
0078 [percentErrors, index] = min(percentError);
0079 
0080 net = canidates{index};
0081 disp (['Percent Error from neural network predictions : ',num2str(percentErrors)]);
0082 
0083 end
0084