|
wolffd@0
|
1 function net = gtm(dim_latent, nlatent, dim_data, ncentres, rbfunc, ...
|
|
wolffd@0
|
2 prior)
|
|
wolffd@0
|
3 %GTM Create a Generative Topographic Map.
|
|
wolffd@0
|
4 %
|
|
wolffd@0
|
5 % Description
|
|
wolffd@0
|
6 %
|
|
wolffd@0
|
7 % NET = GTM(DIMLATENT, NLATENT, DIMDATA, NCENTRES, RBFUNC), takes the
|
|
wolffd@0
|
8 % dimension of the latent space DIMLATENT, the number of data points
|
|
wolffd@0
|
9 % sampled in the latent space NLATENT, the dimension of the data space
|
|
wolffd@0
|
10 % DIMDATA, the number of centres in the RBF model NCENTRES, the
|
|
wolffd@0
|
11 % activation function for the RBF RBFUNC and returns a data structure
|
|
wolffd@0
|
12 % NET. The parameters in the RBF and GMM sub-models are set by calls to
|
|
wolffd@0
|
13 % the corresponding creation routines RBF and GMM.
|
|
wolffd@0
|
14 %
|
|
wolffd@0
|
15 % The fields in NET are
|
|
wolffd@0
|
16 % type = 'gtm'
|
|
wolffd@0
|
17 % nin = dimension of data space
|
|
wolffd@0
|
18 % dimlatent = dimension of latent space
|
|
wolffd@0
|
19 % rbfnet = RBF network data structure
|
|
wolffd@0
|
20 % gmmnet = GMM data structure
|
|
wolffd@0
|
21 % X = sample of latent points
|
|
wolffd@0
|
22 %
|
|
wolffd@0
|
23 % NET = GTM(DIMLATENT, NLATENT, DIMDATA, NCENTRES, RBFUNC, PRIOR),
|
|
wolffd@0
|
24 % sets a Gaussian zero mean prior on the parameters of the RBF model.
|
|
wolffd@0
|
25 % PRIOR must be a scalar and represents the inverse variance of the
|
|
wolffd@0
|
26 % prior distribution. This gives rise to a weight decay term in the
|
|
wolffd@0
|
27 % error function.
|
|
wolffd@0
|
28 %
|
|
wolffd@0
|
29 % See also
|
|
wolffd@0
|
30 % GTMFWD, GTMPOST, RBF, GMM
|
|
wolffd@0
|
31 %
|
|
wolffd@0
|
32
|
|
wolffd@0
|
33 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
34
|
|
wolffd@0
|
35 net.type = 'gtm';
|
|
wolffd@0
|
36 % Input to functions is data
|
|
wolffd@0
|
37 net.nin = dim_data;
|
|
wolffd@0
|
38 net.dim_latent = dim_latent;
|
|
wolffd@0
|
39
|
|
wolffd@0
|
40 % Default is no regularisation
|
|
wolffd@0
|
41 if nargin == 5
|
|
wolffd@0
|
42 prior = 0.0;
|
|
wolffd@0
|
43 end
|
|
wolffd@0
|
44
|
|
wolffd@0
|
45 % Only allow scalar prior
|
|
wolffd@0
|
46 if isstruct(prior) | size(prior) ~= [1 1]
|
|
wolffd@0
|
47 error('Prior must be a scalar');
|
|
wolffd@0
|
48 end
|
|
wolffd@0
|
49
|
|
wolffd@0
|
50 % Create RBF network
|
|
wolffd@0
|
51 net.rbfnet = rbf(dim_latent, ncentres, dim_data, rbfunc, ...
|
|
wolffd@0
|
52 'linear', prior);
|
|
wolffd@0
|
53
|
|
wolffd@0
|
54 % Mask all but output weights
|
|
wolffd@0
|
55 net.rbfnet.mask = rbfprior(rbfunc, dim_latent, ncentres, dim_data);
|
|
wolffd@0
|
56
|
|
wolffd@0
|
57 % Create field for GMM output model
|
|
wolffd@0
|
58 net.gmmnet = gmm(dim_data, nlatent, 'spherical');
|
|
wolffd@0
|
59
|
|
wolffd@0
|
60 % Create empty latent data sample
|
|
wolffd@0
|
61 net.X = []; |
