|
wolffd@0
|
1 function [prob,a] = mdnprob(mixparams, t)
|
|
wolffd@0
|
2 %MDNPROB Computes the data probability likelihood for an MDN mixture structure.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 % PROB = MDNPROB(MIXPARAMS, T) computes the probability P(T) of each
|
|
wolffd@0
|
6 % data vector in T under the Gaussian mixture model represented by the
|
|
wolffd@0
|
7 % corresponding entries in MIXPARAMS. Each row of T represents a single
|
|
wolffd@0
|
8 % vector.
|
|
wolffd@0
|
9 %
|
|
wolffd@0
|
10 % [PROB, A] = MDNPROB(MIXPARAMS, T) also computes the activations A
|
|
wolffd@0
|
11 % (i.e. the probability P(T|J) of the data conditioned on each
|
|
wolffd@0
|
12 % component density) for a Gaussian mixture model.
|
|
wolffd@0
|
13 %
|
|
wolffd@0
|
14 % See also
|
|
wolffd@0
|
15 % MDNERR, MDNPOST
|
|
wolffd@0
|
16 %
|
|
wolffd@0
|
17
|
|
wolffd@0
|
18 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
19 % David J Evans (1998)
|
|
wolffd@0
|
20
|
|
wolffd@0
|
21 % Check arguments for consistency
|
|
wolffd@0
|
22 errstring = consist(mixparams, 'mdnmixes');
|
|
wolffd@0
|
23 if ~isempty(errstring)
|
|
wolffd@0
|
24 error(errstring);
|
|
wolffd@0
|
25 end
|
|
wolffd@0
|
26
|
|
wolffd@0
|
27 ntarget = size(t, 1);
|
|
wolffd@0
|
28 if ntarget ~= size(mixparams.centres, 1)
|
|
wolffd@0
|
29 error('Number of targets does not match number of mixtures')
|
|
wolffd@0
|
30 end
|
|
wolffd@0
|
31 if size(t, 2) ~= mixparams.dim_target
|
|
wolffd@0
|
32 error('Target dimension does not match mixture dimension')
|
|
wolffd@0
|
33 end
|
|
wolffd@0
|
34
|
|
wolffd@0
|
35 dim_target = mixparams.dim_target;
|
|
wolffd@0
|
36 ntarget = size(t, 1);
|
|
wolffd@0
|
37
|
|
wolffd@0
|
38 % Calculate squared norm matrix, of dimension (ndata, ncentres)
|
|
wolffd@0
|
39 % vector (ntarget * ncentres)
|
|
wolffd@0
|
40 dist2 = mdndist2(mixparams, t);
|
|
wolffd@0
|
41
|
|
wolffd@0
|
42 % Calculate variance factors
|
|
wolffd@0
|
43 variance = 2.*mixparams.covars;
|
|
wolffd@0
|
44
|
|
wolffd@0
|
45 % Compute the normalisation term
|
|
wolffd@0
|
46 normal = ((2.*pi).*mixparams.covars).^(dim_target./2);
|
|
wolffd@0
|
47
|
|
wolffd@0
|
48 % Now compute the activations
|
|
wolffd@0
|
49 a = exp(-(dist2./variance))./normal;
|
|
wolffd@0
|
50
|
|
wolffd@0
|
51 % Accumulate negative log likelihood of targets
|
|
wolffd@0
|
52 prob = mixparams.mixcoeffs.*a;
|
