|
wolffd@0
|
1 function [mixparams, y, z, a] = mdnfwd(net, x)
|
|
wolffd@0
|
2 %MDNFWD Forward propagation through Mixture Density Network.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 % MIXPARAMS = MDNFWD(NET, X) takes a mixture density network data
|
|
wolffd@0
|
6 % structure NET and a matrix X of input vectors, and forward propagates
|
|
wolffd@0
|
7 % the inputs through the network to generate a structure MIXPARAMS
|
|
wolffd@0
|
8 % which contains the parameters of several mixture models. Each row
|
|
wolffd@0
|
9 % of X represents one input vector and the corresponding row of the
|
|
wolffd@0
|
10 % matrices in MIXPARAMS represents the parameters of a mixture model
|
|
wolffd@0
|
11 % for the conditional probability of target vectors given the input
|
|
wolffd@0
|
12 % vector. This is not represented as an array of GMM structures to
|
|
wolffd@0
|
13 % improve the efficiency of MDN training.
|
|
wolffd@0
|
14 %
|
|
wolffd@0
|
15 % The fields in MIXPARAMS are
|
|
wolffd@0
|
16 % type = 'mdnmixes'
|
|
wolffd@0
|
17 % ncentres = number of mixture components
|
|
wolffd@0
|
18 % dimtarget = dimension of target space
|
|
wolffd@0
|
19 % mixcoeffs = mixing coefficients
|
|
wolffd@0
|
20 % centres = means of Gaussians: stored as one row per pattern
|
|
wolffd@0
|
21 % covars = covariances of Gaussians
|
|
wolffd@0
|
22 % nparams = number of parameters
|
|
wolffd@0
|
23 %
|
|
wolffd@0
|
24 % [MIXPARAMS, Y, Z] = MDNFWD(NET, X) also generates a matrix Y of the
|
|
wolffd@0
|
25 % outputs of the MLP and a matrix Z of the hidden unit activations
|
|
wolffd@0
|
26 % where each row corresponds to one pattern.
|
|
wolffd@0
|
27 %
|
|
wolffd@0
|
28 % [MIXPARAMS, Y, Z, A] = MLPFWD(NET, X) also returns a matrix A giving
|
|
wolffd@0
|
29 % the summed inputs to each output unit, where each row corresponds to
|
|
wolffd@0
|
30 % one pattern.
|
|
wolffd@0
|
31 %
|
|
wolffd@0
|
32 % See also
|
|
wolffd@0
|
33 % MDN, MDN2GMM, MDNERR, MDNGRAD, MLPFWD
|
|
wolffd@0
|
34 %
|
|
wolffd@0
|
35
|
|
wolffd@0
|
36 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
37 % David J Evans (1998)
|
|
wolffd@0
|
38
|
|
wolffd@0
|
39 % Check arguments for consistency
|
|
wolffd@0
|
40 errstring = consist(net, 'mdn', x);
|
|
wolffd@0
|
41 if ~isempty(errstring)
|
|
wolffd@0
|
42 error(errstring);
|
|
wolffd@0
|
43 end
|
|
wolffd@0
|
44
|
|
wolffd@0
|
45 % Extract mlp and mixture model descriptors
|
|
wolffd@0
|
46 mlpnet = net.mlp;
|
|
wolffd@0
|
47 mixes = net.mdnmixes;
|
|
wolffd@0
|
48
|
|
wolffd@0
|
49 ncentres = mixes.ncentres; % Number of components in mixture model
|
|
wolffd@0
|
50 dim_target = mixes.dim_target; % Dimension of targets
|
|
wolffd@0
|
51 nparams = mixes.nparams; % Number of parameters in mixture model
|
|
wolffd@0
|
52
|
|
wolffd@0
|
53 % Propagate forwards through MLP
|
|
wolffd@0
|
54 [y, z, a] = mlpfwd(mlpnet, x);
|
|
wolffd@0
|
55
|
|
wolffd@0
|
56 % Compute the postion for each parameter in the whole
|
|
wolffd@0
|
57 % matrix. Used to define the mixparams structure
|
|
wolffd@0
|
58 mixcoeff = [1:1:ncentres];
|
|
wolffd@0
|
59 centres = [ncentres+1:1:(ncentres*(1+dim_target))];
|
|
wolffd@0
|
60 variances = [(ncentres*(1+dim_target)+1):1:nparams];
|
|
wolffd@0
|
61
|
|
wolffd@0
|
62 % Convert output values into mixture model parameters
|
|
wolffd@0
|
63
|
|
wolffd@0
|
64 % Use softmax to calculate priors
|
|
wolffd@0
|
65 % Prevent overflow and underflow: use same bounds as glmfwd
|
|
wolffd@0
|
66 % Ensure that sum(exp(y), 2) does not overflow
|
|
wolffd@0
|
67 maxcut = log(realmax) - log(ncentres);
|
|
wolffd@0
|
68 % Ensure that exp(y) > 0
|
|
wolffd@0
|
69 mincut = log(realmin);
|
|
wolffd@0
|
70 temp = min(y(:,1:ncentres), maxcut);
|
|
wolffd@0
|
71 temp = max(temp, mincut);
|
|
wolffd@0
|
72 temp = exp(temp);
|
|
wolffd@0
|
73 mixpriors = temp./(sum(temp, 2)*ones(1,ncentres));
|
|
wolffd@0
|
74
|
|
wolffd@0
|
75 % Centres are just copies of network outputs
|
|
wolffd@0
|
76 mixcentres = y(:,(ncentres+1):ncentres*(1+dim_target));
|
|
wolffd@0
|
77
|
|
wolffd@0
|
78 % Variances are exp of network outputs
|
|
wolffd@0
|
79 mixwidths = exp(y(:,(ncentres*(1+dim_target)+1):nparams));
|
|
wolffd@0
|
80
|
|
wolffd@0
|
81 % Now build up all the mixture model weight vectors
|
|
wolffd@0
|
82 ndata = size(x, 1);
|
|
wolffd@0
|
83
|
|
wolffd@0
|
84 % Return parameters
|
|
wolffd@0
|
85 mixparams.type = mixes.type;
|
|
wolffd@0
|
86 mixparams.ncentres = mixes.ncentres;
|
|
wolffd@0
|
87 mixparams.dim_target = mixes.dim_target;
|
|
wolffd@0
|
88 mixparams.nparams = mixes.nparams;
|
|
wolffd@0
|
89
|
|
wolffd@0
|
90 mixparams.mixcoeffs = mixpriors;
|
|
wolffd@0
|
91 mixparams.centres = mixcentres;
|
|
wolffd@0
|
92 mixparams.covars = mixwidths;
|
|
wolffd@0
|
93
|
