wolffd@0: function [mixparams, y, z, a] = mdnfwd(net, x)
wolffd@0: %MDNFWD	Forward propagation through Mixture Density Network.
wolffd@0: %
wolffd@0: %	Description
wolffd@0: %	 MIXPARAMS = MDNFWD(NET, X) takes a mixture density network data
wolffd@0: %	structure NET and a matrix X of input vectors, and forward propagates
wolffd@0: %	the inputs through the network to generate a structure MIXPARAMS
wolffd@0: %	which contains the parameters of several mixture models.   Each row
wolffd@0: %	of X represents one input vector and the corresponding row of the
wolffd@0: %	matrices in MIXPARAMS  represents the parameters of a mixture model
wolffd@0: %	for the conditional probability of target vectors given the input
wolffd@0: %	vector.  This is not represented as an array of GMM structures to
wolffd@0: %	improve the efficiency of MDN training.
wolffd@0: %
wolffd@0: %	The fields in MIXPARAMS are
wolffd@0: %	  type = 'mdnmixes'
wolffd@0: %	  ncentres = number of mixture components
wolffd@0: %	  dimtarget = dimension of target space
wolffd@0: %	  mixcoeffs = mixing coefficients
wolffd@0: %	  centres = means of Gaussians: stored as one row per pattern
wolffd@0: %	  covars = covariances of Gaussians
wolffd@0: %	  nparams = number of parameters
wolffd@0: %
wolffd@0: %	[MIXPARAMS, Y, Z] = MDNFWD(NET, X) also generates a matrix Y of the
wolffd@0: %	outputs of the MLP and a matrix Z of the hidden unit activations
wolffd@0: %	where each row corresponds to one pattern.
wolffd@0: %
wolffd@0: %	[MIXPARAMS, Y, Z, A] = MLPFWD(NET, X) also returns a matrix A  giving
wolffd@0: %	the summed inputs to each output unit, where each row  corresponds to
wolffd@0: %	one pattern.
wolffd@0: %
wolffd@0: %	See also
wolffd@0: %	MDN, MDN2GMM, MDNERR, MDNGRAD, MLPFWD
wolffd@0: %
wolffd@0: 
wolffd@0: %	Copyright (c) Ian T Nabney (1996-2001)
wolffd@0: %	David J Evans (1998)
wolffd@0: 
wolffd@0: % Check arguments for consistency
wolffd@0: errstring = consist(net, 'mdn', x);
wolffd@0: if ~isempty(errstring)
wolffd@0:   error(errstring);
wolffd@0: end
wolffd@0: 
wolffd@0: % Extract mlp and mixture model descriptors
wolffd@0: mlpnet = net.mlp;
wolffd@0: mixes = net.mdnmixes;
wolffd@0: 
wolffd@0: ncentres = mixes.ncentres;	% Number of components in mixture model
wolffd@0: dim_target = mixes.dim_target;	% Dimension of targets
wolffd@0: nparams = mixes.nparams;	% Number of parameters in mixture model
wolffd@0: 
wolffd@0: % Propagate forwards through MLP
wolffd@0: [y, z, a] = mlpfwd(mlpnet, x);
wolffd@0: 
wolffd@0: % Compute the postion for each parameter in the whole
wolffd@0: % matrix.  Used to define the mixparams structure
wolffd@0: mixcoeff  = [1:1:ncentres];
wolffd@0: centres   = [ncentres+1:1:(ncentres*(1+dim_target))];
wolffd@0: variances = [(ncentres*(1+dim_target)+1):1:nparams];
wolffd@0: 
wolffd@0: % Convert output values into mixture model parameters
wolffd@0: 
wolffd@0: % Use softmax to calculate priors
wolffd@0: % Prevent overflow and underflow: use same bounds as glmfwd
wolffd@0: % Ensure that sum(exp(y), 2) does not overflow
wolffd@0: maxcut = log(realmax) - log(ncentres);
wolffd@0: % Ensure that exp(y) > 0
wolffd@0: mincut = log(realmin);
wolffd@0: temp = min(y(:,1:ncentres), maxcut);
wolffd@0: temp = max(temp, mincut);
wolffd@0: temp = exp(temp);
wolffd@0: mixpriors = temp./(sum(temp, 2)*ones(1,ncentres));
wolffd@0: 
wolffd@0: % Centres are just copies of network outputs
wolffd@0: mixcentres =  y(:,(ncentres+1):ncentres*(1+dim_target));
wolffd@0: 
wolffd@0: % Variances are exp of network outputs
wolffd@0: mixwidths = exp(y(:,(ncentres*(1+dim_target)+1):nparams));
wolffd@0: 
wolffd@0: % Now build up all the mixture model weight vectors
wolffd@0: ndata = size(x, 1);
wolffd@0: 
wolffd@0: % Return parameters
wolffd@0: mixparams.type = mixes.type;
wolffd@0: mixparams.ncentres = mixes.ncentres;
wolffd@0: mixparams.dim_target = mixes.dim_target;
wolffd@0: mixparams.nparams = mixes.nparams;
wolffd@0: 
wolffd@0: mixparams.mixcoeffs = mixpriors;
wolffd@0: mixparams.centres   = mixcentres;
wolffd@0: mixparams.covars      = mixwidths;
wolffd@0: