% Classifier design based on GMM

vsdOpt=vsdOptSet;

% ====== Load DS.mat
fprintf('Loading DS.mat...\n');
load DS.mat
load bestInputIndex.mat
bestInputIndex=1:size(DS.input,1);		% This is for HMM
DS2=DS;								% DS2 is DS after input selection and normalization
DS2.input=DS2.input(bestInputIndex, :);				% Use the selected features
DS2.inputName=DS2.inputName(bestInputIndex);			% Update the input names based on the selected features
if vsdOpt.useInputNormalize
	[DS2.input, mu, sigma]=inputNormalize(DS2.input);	% Input normalization
end
priors=classDataCount(DS2);					% Take data count of each class as the class probability
% ====== Down sample the training data
%fprintf('Down sampling DS2...\n');
%DS2.input=DS2.input(:, 1:vsdOpt.downSampleRate:end);
%DS2.output=DS2.output(:, 1:vsdOpt.downSampleRate:end);

% ====== GMM classifier
cvData=cvDataGen(DS2, 2);	% 2-fold cross-validation
DS=cvData(1).DS;
TS=cvData(1).TS;
count1=classDataCount(DS); count2=classDataCount(TS);
vecOfGaussianNum=1:min([count1, count2]);
vecOfGaussianNum=[2, 4, 8, 16, 32, 64, 128, 256, 512, 1024];
vecOfGaussianNum=2:2:30;
gmmTrainPrm=gmmTrainPrmSet; gmmTrainPrm.plotOpt=1;
covType=1;
figure;
[gmmData, recogRate1, recogRate2]=gmmcTrainEvalWrtGaussianNum(DS, TS, vecOfGaussianNum, covType, gmmTrainPrm);
print -dpng gmmcTrainEvalWrtGaussianNum
[maxRr, index]=max(recogRate2);
line(vecOfGaussianNum(index), maxRr*100, 'marker', 'o', 'color', 'r');
fprintf('Saving gmmData...\n');
save gmmData gmmData index recogRate1 recogRate2