wolffd@0
|
1 %DEMGMM3 Demonstrate density modelling with a Gaussian mixture model.
|
wolffd@0
|
2 %
|
wolffd@0
|
3 % Description
|
wolffd@0
|
4 % The problem consists of modelling data generated by a mixture of
|
wolffd@0
|
5 % three Gaussians in 2 dimensions with a mixture model using diagonal
|
wolffd@0
|
6 % covariance matrices. The priors are 0.3, 0.5 and 0.2; the centres
|
wolffd@0
|
7 % are (2, 3.5), (0, 0) and (0,2); the covariances are all axis aligned
|
wolffd@0
|
8 % (0.16, 0.64), (0.25, 1) and the identity matrix. The first figure
|
wolffd@0
|
9 % contains a scatter plot of the data.
|
wolffd@0
|
10 %
|
wolffd@0
|
11 % A Gaussian mixture model with three components is trained using EM.
|
wolffd@0
|
12 % The parameter vector is printed before training and after training.
|
wolffd@0
|
13 % The user should press any key to continue at these points. The
|
wolffd@0
|
14 % parameter vector consists of priors (the column), and centres (given
|
wolffd@0
|
15 % as (x, y) pairs as the next two columns). The diagonal entries of
|
wolffd@0
|
16 % the covariance matrices are printed separately.
|
wolffd@0
|
17 %
|
wolffd@0
|
18 % The second figure is a 3 dimensional view of the density function,
|
wolffd@0
|
19 % while the third shows the axes of the 1-standard deviation circles
|
wolffd@0
|
20 % for the three components of the mixture model.
|
wolffd@0
|
21 %
|
wolffd@0
|
22 % See also
|
wolffd@0
|
23 % GMM, GMMINIT, GMMEM, GMMPROB, GMMUNPAK
|
wolffd@0
|
24 %
|
wolffd@0
|
25
|
wolffd@0
|
26 % Copyright (c) Ian T Nabney (1996-2001)
|
wolffd@0
|
27
|
wolffd@0
|
28 % Generate the data
|
wolffd@0
|
29 ndata = 500;
|
wolffd@0
|
30
|
wolffd@0
|
31 % Fix the seeds for reproducible results
|
wolffd@0
|
32 randn('state', 42);
|
wolffd@0
|
33 rand('state', 42);
|
wolffd@0
|
34 data = randn(ndata, 2);
|
wolffd@0
|
35 prior = [0.3 0.5 0.2];
|
wolffd@0
|
36 % Mixture model swaps clusters 1 and 3
|
wolffd@0
|
37 datap = [0.2 0.5 0.3];
|
wolffd@0
|
38 datac = [0 2; 0 0; 2 3.5];
|
wolffd@0
|
39 datacov = [1 1;1 0.25; 0.4*0.4 0.8*0.8];
|
wolffd@0
|
40 data1 = data(1:prior(1)*ndata,:);
|
wolffd@0
|
41 data2 = data(prior(1)*ndata+1:(prior(2)+prior(1))*ndata, :);
|
wolffd@0
|
42 data3 = data((prior(1)+prior(2))*ndata +1:ndata, :);
|
wolffd@0
|
43
|
wolffd@0
|
44 % First cluster has axis aligned variance and centre (2, 3.5)
|
wolffd@0
|
45 data1(:, 1) = data1(:, 1)*0.4 + 2.0;
|
wolffd@0
|
46 data1(:, 2) = data1(:, 2)*0.8 + 3.5;
|
wolffd@0
|
47
|
wolffd@0
|
48 % Second cluster has axis aligned variance and centre (0, 0)
|
wolffd@0
|
49 data2(:,2) = data2(:, 2)*0.5;
|
wolffd@0
|
50
|
wolffd@0
|
51 % Third cluster is at (0,2) with identity matrix for covariance
|
wolffd@0
|
52 data3 = data3 + repmat([0 2], prior(3)*ndata, 1);
|
wolffd@0
|
53
|
wolffd@0
|
54 % Put the dataset together again
|
wolffd@0
|
55 data = [data1; data2; data3];
|
wolffd@0
|
56
|
wolffd@0
|
57 clc
|
wolffd@0
|
58 disp('This demonstration illustrates the use of a Gaussian mixture model')
|
wolffd@0
|
59 disp('with diagonal covariance matrices to approximate the unconditional')
|
wolffd@0
|
60 disp('probability density of data in a two-dimensional space.')
|
wolffd@0
|
61 disp('We begin by generating the data from a mixture of three Gaussians')
|
wolffd@0
|
62 disp('with axis aligned covariance structure and plotting it.')
|
wolffd@0
|
63 disp(' ')
|
wolffd@0
|
64 disp('The first cluster has centre (0, 2).')
|
wolffd@0
|
65 disp('The second cluster has centre (0, 0).')
|
wolffd@0
|
66 disp('The third cluster has centre (2, 3.5).')
|
wolffd@0
|
67 disp(' ')
|
wolffd@0
|
68 disp('Press any key to continue')
|
wolffd@0
|
69 pause
|
wolffd@0
|
70
|
wolffd@0
|
71 fh1 = figure;
|
wolffd@0
|
72 plot(data(:, 1), data(:, 2), 'o')
|
wolffd@0
|
73 set(gca, 'Box', 'on')
|
wolffd@0
|
74
|
wolffd@0
|
75 % Set up mixture model
|
wolffd@0
|
76 ncentres = 3;
|
wolffd@0
|
77 input_dim = 2;
|
wolffd@0
|
78 mix = gmm(input_dim, ncentres, 'diag');
|
wolffd@0
|
79
|
wolffd@0
|
80 options = foptions;
|
wolffd@0
|
81 options(14) = 5; % Just use 5 iterations of k-means in initialisation
|
wolffd@0
|
82 % Initialise the model parameters from the data
|
wolffd@0
|
83 mix = gmminit(mix, data, options);
|
wolffd@0
|
84
|
wolffd@0
|
85 % Print out model
|
wolffd@0
|
86 disp('The mixture model has three components and diagonal covariance')
|
wolffd@0
|
87 disp('matrices. The model parameters after initialisation using the')
|
wolffd@0
|
88 disp('k-means algorithm are as follows')
|
wolffd@0
|
89 disp(' Priors Centres')
|
wolffd@0
|
90 disp([mix.priors' mix.centres])
|
wolffd@0
|
91 disp('Covariance diagonals are')
|
wolffd@0
|
92 disp(mix.covars)
|
wolffd@0
|
93 disp('Press any key to continue.')
|
wolffd@0
|
94 pause
|
wolffd@0
|
95
|
wolffd@0
|
96 % Set up vector of options for EM trainer
|
wolffd@0
|
97 options = zeros(1, 18);
|
wolffd@0
|
98 options(1) = 1; % Prints out error values.
|
wolffd@0
|
99 options(14) = 20; % Number of iterations.
|
wolffd@0
|
100
|
wolffd@0
|
101 disp('We now train the model using the EM algorithm for 20 iterations.')
|
wolffd@0
|
102 disp(' ')
|
wolffd@0
|
103 disp('Press any key to continue.')
|
wolffd@0
|
104 pause
|
wolffd@0
|
105
|
wolffd@0
|
106 [mix, options, errlog] = gmmem(mix, data, options);
|
wolffd@0
|
107
|
wolffd@0
|
108 % Print out model
|
wolffd@0
|
109 disp(' ')
|
wolffd@0
|
110 disp('The trained model has priors and centres:')
|
wolffd@0
|
111 disp(' Priors Centres')
|
wolffd@0
|
112 disp([mix.priors' mix.centres])
|
wolffd@0
|
113 disp('The data generator has priors and centres')
|
wolffd@0
|
114 disp(' Priors Centres')
|
wolffd@0
|
115 disp([datap' datac])
|
wolffd@0
|
116 disp('Model covariance diagonals are')
|
wolffd@0
|
117 disp(mix.covars)
|
wolffd@0
|
118 disp('Data generator covariance diagonals are')
|
wolffd@0
|
119 disp(datacov)
|
wolffd@0
|
120 disp('Note the close correspondence between these parameters and those')
|
wolffd@0
|
121 disp('of the distribution used to generate the data.')
|
wolffd@0
|
122 disp(' ')
|
wolffd@0
|
123 disp('Press any key to continue.')
|
wolffd@0
|
124 pause
|
wolffd@0
|
125
|
wolffd@0
|
126 clc
|
wolffd@0
|
127 disp('We now plot the density given by the mixture model as a surface plot.')
|
wolffd@0
|
128 disp(' ')
|
wolffd@0
|
129 disp('Press any key to continue.')
|
wolffd@0
|
130 pause
|
wolffd@0
|
131
|
wolffd@0
|
132 % Plot the result
|
wolffd@0
|
133 x = -4.0:0.2:5.0;
|
wolffd@0
|
134 y = -4.0:0.2:5.0;
|
wolffd@0
|
135 [X, Y] = meshgrid(x,y);
|
wolffd@0
|
136 X = X(:);
|
wolffd@0
|
137 Y = Y(:);
|
wolffd@0
|
138 grid = [X Y];
|
wolffd@0
|
139 Z = gmmprob(mix, grid);
|
wolffd@0
|
140 Z = reshape(Z, length(x), length(y));
|
wolffd@0
|
141 c = mesh(x, y, Z);
|
wolffd@0
|
142 hold on
|
wolffd@0
|
143 title('Surface plot of probability density')
|
wolffd@0
|
144 hold off
|
wolffd@0
|
145 drawnow
|
wolffd@0
|
146
|
wolffd@0
|
147 clc
|
wolffd@0
|
148 disp('The final plot shows the centres and widths, given by one standard')
|
wolffd@0
|
149 disp('deviation, of the three components of the mixture model. The axes')
|
wolffd@0
|
150 disp('of the ellipses of constant density are shown.')
|
wolffd@0
|
151 disp(' ')
|
wolffd@0
|
152 disp('Press any key to continue.')
|
wolffd@0
|
153 pause
|
wolffd@0
|
154
|
wolffd@0
|
155 % Try to calculate a sensible position for the second figure, below the first
|
wolffd@0
|
156 fig1_pos = get(fh1, 'Position');
|
wolffd@0
|
157 fig2_pos = fig1_pos;
|
wolffd@0
|
158 fig2_pos(2) = fig2_pos(2) - fig1_pos(4);
|
wolffd@0
|
159 fh2 = figure('Position', fig2_pos);
|
wolffd@0
|
160
|
wolffd@0
|
161 h = plot(data(:, 1), data(:, 2), 'bo');
|
wolffd@0
|
162 hold on
|
wolffd@0
|
163 axis('equal');
|
wolffd@0
|
164 title('Plot of data and covariances')
|
wolffd@0
|
165 for i = 1:ncentres
|
wolffd@0
|
166 v = [1 0];
|
wolffd@0
|
167 for j = 1:2
|
wolffd@0
|
168 start=mix.centres(i,:)-sqrt(mix.covars(i,:).*v);
|
wolffd@0
|
169 endpt=mix.centres(i,:)+sqrt(mix.covars(i,:).*v);
|
wolffd@0
|
170 linex = [start(1) endpt(1)];
|
wolffd@0
|
171 liney = [start(2) endpt(2)];
|
wolffd@0
|
172 line(linex, liney, 'Color', 'k', 'LineWidth', 3)
|
wolffd@0
|
173 v = [0 1];
|
wolffd@0
|
174 end
|
wolffd@0
|
175 % Plot ellipses of one standard deviation
|
wolffd@0
|
176 theta = 0:0.02:2*pi;
|
wolffd@0
|
177 x = sqrt(mix.covars(i,1))*cos(theta) + mix.centres(i,1);
|
wolffd@0
|
178 y = sqrt(mix.covars(i,2))*sin(theta) + mix.centres(i,2);
|
wolffd@0
|
179 plot(x, y, 'r-');
|
wolffd@0
|
180 end
|
wolffd@0
|
181 hold off
|
wolffd@0
|
182
|
wolffd@0
|
183 disp('Note how the data cluster positions and widths are captured by')
|
wolffd@0
|
184 disp('the mixture model.')
|
wolffd@0
|
185 disp(' ')
|
wolffd@0
|
186 disp('Press any key to end.')
|
wolffd@0
|
187 pause
|
wolffd@0
|
188
|
wolffd@0
|
189 close(fh1);
|
wolffd@0
|
190 close(fh2);
|
wolffd@0
|
191 clear all;
|
wolffd@0
|
192
|