|
wolffd@0
|
1 function net = mdninit(net, prior, t, options)
|
|
wolffd@0
|
2 %MDNINIT Initialise the weights in a Mixture Density Network.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 %
|
|
wolffd@0
|
6 % NET = MDNINIT(NET, PRIOR) takes a Mixture Density Network NET and
|
|
wolffd@0
|
7 % sets the weights and biases by sampling from a Gaussian distribution.
|
|
wolffd@0
|
8 % It calls MLPINIT for the MLP component of NET.
|
|
wolffd@0
|
9 %
|
|
wolffd@0
|
10 % NET = MDNINIT(NET, PRIOR, T, OPTIONS) uses the target data T to
|
|
wolffd@0
|
11 % initialise the biases for the output units after initialising the
|
|
wolffd@0
|
12 % other weights as above. It calls GMMINIT, with T and OPTIONS as
|
|
wolffd@0
|
13 % arguments, to obtain a model of the unconditional density of T. The
|
|
wolffd@0
|
14 % biases are then set so that NET will output the values in the
|
|
wolffd@0
|
15 % Gaussian mixture model.
|
|
wolffd@0
|
16 %
|
|
wolffd@0
|
17 % See also
|
|
wolffd@0
|
18 % MDN, MLP, MLPINIT, GMMINIT
|
|
wolffd@0
|
19 %
|
|
wolffd@0
|
20
|
|
wolffd@0
|
21 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
22 % David J Evans (1998)
|
|
wolffd@0
|
23
|
|
wolffd@0
|
24 % Initialise network weights from prior: this gives noise around values
|
|
wolffd@0
|
25 % determined later
|
|
wolffd@0
|
26 net.mlp = mlpinit(net.mlp, prior);
|
|
wolffd@0
|
27
|
|
wolffd@0
|
28 if nargin > 2
|
|
wolffd@0
|
29 % Initialise priors, centres and variances from target data
|
|
wolffd@0
|
30 temp_mix = gmm(net.mdnmixes.dim_target, net.mdnmixes.ncentres, 'spherical');
|
|
wolffd@0
|
31 temp_mix = gmminit(temp_mix, t, options);
|
|
wolffd@0
|
32
|
|
wolffd@0
|
33 ncentres = net.mdnmixes.ncentres;
|
|
wolffd@0
|
34 dim_target = net.mdnmixes.dim_target;
|
|
wolffd@0
|
35
|
|
wolffd@0
|
36 % Now set parameters in MLP to yield the right values.
|
|
wolffd@0
|
37 % This involves setting the biases correctly.
|
|
wolffd@0
|
38
|
|
wolffd@0
|
39 % Priors
|
|
wolffd@0
|
40 net.mlp.b2(1:ncentres) = temp_mix.priors;
|
|
wolffd@0
|
41
|
|
wolffd@0
|
42 % Centres are arranged in mlp such that we have
|
|
wolffd@0
|
43 % u11, u12, u13, ..., u1c, ... , uj1, uj2, uj3, ..., ujc, ..., um1, uM2,
|
|
wolffd@0
|
44 % ..., uMc
|
|
wolffd@0
|
45 % This is achieved by transposing temp_mix.centres before reshaping
|
|
wolffd@0
|
46 end_centres = ncentres*(dim_target+1);
|
|
wolffd@0
|
47 net.mlp.b2(ncentres+1:end_centres) = ...
|
|
wolffd@0
|
48 reshape(temp_mix.centres', 1, ncentres*dim_target);
|
|
wolffd@0
|
49
|
|
wolffd@0
|
50 % Variances
|
|
wolffd@0
|
51 net.mlp.b2((end_centres+1):net.mlp.nout) = ...
|
|
wolffd@0
|
52 log(temp_mix.covars);
|
|
wolffd@0
|
53 end
|
