wolffd@0: function net = mdninit(net, prior, t, options)
wolffd@0: %MDNINIT Initialise the weights in a Mixture Density Network.
wolffd@0: %
wolffd@0: %	Description
wolffd@0: %
wolffd@0: %	NET = MDNINIT(NET, PRIOR) takes a Mixture Density Network NET and
wolffd@0: %	sets the weights and biases by sampling from a Gaussian distribution.
wolffd@0: %	It calls MLPINIT for the MLP component of NET.
wolffd@0: %
wolffd@0: %	NET = MDNINIT(NET, PRIOR, T, OPTIONS) uses the target data T to
wolffd@0: %	initialise the biases for the output units after initialising the
wolffd@0: %	other weights as above.  It calls GMMINIT, with T and OPTIONS as
wolffd@0: %	arguments, to obtain a model of the unconditional density of T.  The
wolffd@0: %	biases are then set so that NET will output the values in the
wolffd@0: %	Gaussian  mixture model.
wolffd@0: %
wolffd@0: %	See also
wolffd@0: %	MDN, MLP, MLPINIT, GMMINIT
wolffd@0: %
wolffd@0: 
wolffd@0: %	Copyright (c) Ian T Nabney (1996-2001)
wolffd@0: %	David J Evans (1998)
wolffd@0: 
wolffd@0: % Initialise network weights from prior: this gives noise around values
wolffd@0: % determined later
wolffd@0: net.mlp = mlpinit(net.mlp, prior);
wolffd@0: 
wolffd@0: if nargin > 2
wolffd@0:   % Initialise priors, centres and variances from target data
wolffd@0:   temp_mix = gmm(net.mdnmixes.dim_target, net.mdnmixes.ncentres, 'spherical');
wolffd@0:   temp_mix = gmminit(temp_mix, t, options);
wolffd@0:   
wolffd@0:   ncentres = net.mdnmixes.ncentres;
wolffd@0:   dim_target = net.mdnmixes.dim_target;
wolffd@0: 
wolffd@0:   % Now set parameters in MLP to yield the right values.
wolffd@0:   % This involves setting the biases correctly.
wolffd@0:   
wolffd@0:   % Priors
wolffd@0:   net.mlp.b2(1:ncentres) = temp_mix.priors;
wolffd@0:   
wolffd@0:   % Centres are arranged in mlp such that we have
wolffd@0:   % u11, u12, u13, ..., u1c, ... , uj1, uj2, uj3, ..., ujc, ..., um1, uM2, 
wolffd@0:   % ..., uMc
wolffd@0:   % This is achieved by transposing temp_mix.centres before reshaping
wolffd@0:   end_centres = ncentres*(dim_target+1);
wolffd@0:   net.mlp.b2(ncentres+1:end_centres) = ...
wolffd@0:     reshape(temp_mix.centres', 1, ncentres*dim_target);
wolffd@0:   
wolffd@0:   % Variances
wolffd@0:   net.mlp.b2((end_centres+1):net.mlp.nout) = ...
wolffd@0:     log(temp_mix.covars);
wolffd@0: end