|
wolffd@0
|
1 function net = mdn(nin, nhidden, ncentres, dim_target, mix_type, ...
|
|
wolffd@0
|
2 prior, beta)
|
|
wolffd@0
|
3 %MDN Creates a Mixture Density Network with specified architecture.
|
|
wolffd@0
|
4 %
|
|
wolffd@0
|
5 % Description
|
|
wolffd@0
|
6 % NET = MDN(NIN, NHIDDEN, NCENTRES, DIMTARGET) takes the number of
|
|
wolffd@0
|
7 % inputs, hidden units for a 2-layer feed-forward network and the
|
|
wolffd@0
|
8 % number of centres and target dimension for the mixture model whose
|
|
wolffd@0
|
9 % parameters are set from the outputs of the neural network. The fifth
|
|
wolffd@0
|
10 % argument MIXTYPE is used to define the type of mixture model.
|
|
wolffd@0
|
11 % (Currently there is only one type supported: a mixture of Gaussians
|
|
wolffd@0
|
12 % with a single covariance parameter for each component.) For this
|
|
wolffd@0
|
13 % model, the mixture coefficients are computed from a group of softmax
|
|
wolffd@0
|
14 % outputs, the centres are equal to a group of linear outputs, and the
|
|
wolffd@0
|
15 % variances are obtained by applying the exponential function to a
|
|
wolffd@0
|
16 % third group of outputs.
|
|
wolffd@0
|
17 %
|
|
wolffd@0
|
18 % The network is initialised by a call to MLP, and the arguments PRIOR,
|
|
wolffd@0
|
19 % and BETA have the same role as for that function. Weight
|
|
wolffd@0
|
20 % initialisation uses the Matlab function RANDN and so the seed for
|
|
wolffd@0
|
21 % the random weight initialization can be set using RANDN('STATE', S)
|
|
wolffd@0
|
22 % where S is the seed value. A specialised data structure (rather than
|
|
wolffd@0
|
23 % GMM) is used for the mixture model outputs to improve the efficiency
|
|
wolffd@0
|
24 % of error and gradient calculations in network training. The fields
|
|
wolffd@0
|
25 % are described in MDNFWD where they are set up.
|
|
wolffd@0
|
26 %
|
|
wolffd@0
|
27 % The fields in NET are
|
|
wolffd@0
|
28 %
|
|
wolffd@0
|
29 % type = 'mdn'
|
|
wolffd@0
|
30 % nin = number of input variables
|
|
wolffd@0
|
31 % nout = dimension of target space (not number of network outputs)
|
|
wolffd@0
|
32 % nwts = total number of weights and biases
|
|
wolffd@0
|
33 % mdnmixes = data structure for mixture model output
|
|
wolffd@0
|
34 % mlp = data structure for MLP network
|
|
wolffd@0
|
35 %
|
|
wolffd@0
|
36 % See also
|
|
wolffd@0
|
37 % MDNFWD, MDNERR, MDN2GMM, MDNGRAD, MDNPAK, MDNUNPAK, MLP
|
|
wolffd@0
|
38 %
|
|
wolffd@0
|
39
|
|
wolffd@0
|
40 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
41 % David J Evans (1998)
|
|
wolffd@0
|
42
|
|
wolffd@0
|
43 % Currently ignore type argument: reserved for future use
|
|
wolffd@0
|
44 net.type = 'mdn';
|
|
wolffd@0
|
45
|
|
wolffd@0
|
46 % Set up the mixture model part of the structure
|
|
wolffd@0
|
47 % For efficiency we use a specialised data structure in place of GMM
|
|
wolffd@0
|
48 mdnmixes.type = 'mdnmixes';
|
|
wolffd@0
|
49 mdnmixes.ncentres = ncentres;
|
|
wolffd@0
|
50 mdnmixes.dim_target = dim_target;
|
|
wolffd@0
|
51
|
|
wolffd@0
|
52 % This calculation depends on spherical variances
|
|
wolffd@0
|
53 mdnmixes.nparams = ncentres + ncentres*dim_target + ncentres;
|
|
wolffd@0
|
54
|
|
wolffd@0
|
55 % Make the weights in the mdnmixes structure null
|
|
wolffd@0
|
56 mdnmixes.mixcoeffs = [];
|
|
wolffd@0
|
57 mdnmixes.centres = [];
|
|
wolffd@0
|
58 mdnmixes.covars = [];
|
|
wolffd@0
|
59
|
|
wolffd@0
|
60 % Number of output nodes = number of parameters in mixture model
|
|
wolffd@0
|
61 nout = mdnmixes.nparams;
|
|
wolffd@0
|
62
|
|
wolffd@0
|
63 % Set up the MLP part of the network
|
|
wolffd@0
|
64 if (nargin == 5)
|
|
wolffd@0
|
65 mlpnet = mlp(nin, nhidden, nout, 'linear');
|
|
wolffd@0
|
66 elseif (nargin == 6)
|
|
wolffd@0
|
67 mlpnet = mlp(nin, nhidden, nout, 'linear', prior);
|
|
wolffd@0
|
68 elseif (nargin == 7)
|
|
wolffd@0
|
69 mlpnet = mlp(nin, nhidden, nout, 'linear', prior, beta);
|
|
wolffd@0
|
70 end
|
|
wolffd@0
|
71
|
|
wolffd@0
|
72 % Create descriptor
|
|
wolffd@0
|
73 net.mdnmixes = mdnmixes;
|
|
wolffd@0
|
74 net.mlp = mlpnet;
|
|
wolffd@0
|
75 net.nin = nin;
|
|
wolffd@0
|
76 net.nout = dim_target;
|
|
wolffd@0
|
77 net.nwts = mlpnet.nwts;
|
