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