gmmcGaussianNumEstimate

GMM training and test, w.r.t. varying number of mixtures

Contents

Syntax

Description

[gmmcData, recogRate1, recogRate2, validTrialIndex]=gmmcGaussianNumEstimate(DS, TS, gmmcOpt)

• DS: training set
• TS: test set
• gmmcOpt.arch.gaussianNum: an matrix indicating numbers of Gaussian components, where each column is the number of Gaussian components for each class during a trial
• gmmcOpt.arch.covType: type of covariance matrix, 1: identity times a constant, 2: diagonal, 3: full
• gmmcOpt.train: parameter for training GMM
• gmmcData: GMM parameters
  • gmmcData(i): in which class j has gmmcOpt.arch.gaussianNum(j, i) gaussians
  • gmmcData(i).class(j): gmm of class j at case i
  • gmmcData(i).class(j).gmmPrm(k): gaussian k of class j at case i
  • gmmcData(i).class(j).gmmPrm(k).mu: mean vector
  • gmmcData(i).class(j).gmmPrm(k).sigma: covariance matrix
  • gmmcData(i).class(j).gmmPrm(k).w: weight
• recogRate1: inside-test recognition rate
• recogRate2: outside-test recognition rate
• validTrialIndex: Actually valid index for
• gmmcOpt.arch.gaussianNum. (We need to have this output
• parameters since sometimes we are given a large number of mixtures which cannot be used for GMM training at all.)

Example

[DS, TS]=prData('nonlinearSeparable');
gmmcOpt=gmmcTrain('defaultOpt');
gmmcOpt.arch.gaussianNum=1:10;
gmmcOpt.arch.covType=1;
gmmcOpt.train.maxIteration=50;
plotOpt=1;
[gmmcData, recogRate1, recogRate2, validTrialIndex]=gmmcGaussianNumEstimate(DS, TS, gmmcOpt, plotOpt);
[rr, index]=max(recogRate2);
figure; gmmcPlot(TS, gmmcData(index), '2dPdf');
figure; gmmcPlot(TS, gmmcData(index), 'decBoundary');

DS data count = 250, TS data count = 250
DS class data count = [150 100]
TS class data count = [150 100]
1/10: No. of Gaussian = [1;1] ===> inside RR = 63.2%, outside RR = 62.4%
2/10: No. of Gaussian = [2;2] ===> inside RR = 95.2%, outside RR = 92.8%
3/10: No. of Gaussian = [3;3] ===> inside RR = 96.8%, outside RR = 96%
4/10: No. of Gaussian = [4;4] ===> inside RR = 96.8%, outside RR = 96.4%
5/10: No. of Gaussian = [5;5] ===> inside RR = 97.2%, outside RR = 96%
6/10: No. of Gaussian = [6;6] ===> inside RR = 96.8%, outside RR = 95.2%
7/10: No. of Gaussian = [7;7] ===> inside RR = 96.8%, outside RR = 95.2%
8/10: No. of Gaussian = [8;8] ===> inside RR = 96.8%, outside RR = 94.8%
9/10: No. of Gaussian = [9;9] ===> inside RR = 96.8%, outside RR = 95.2%
10/10: No. of Gaussian = [10;10] ===> inside RR = 98%, outside RR = 94.8%

Published with MATLAB® R2018a