Daniel@0: function net = netinit(net, prior)
Daniel@0: %NETINIT Initialise the weights in a network.
Daniel@0: %
Daniel@0: %	Description
Daniel@0: %
Daniel@0: %	NET = NETINIT(NET, PRIOR) takes a network data structure NET and sets
Daniel@0: %	the weights and biases by sampling from a Gaussian distribution. If
Daniel@0: %	PRIOR is a scalar, then all of the parameters (weights and biases)
Daniel@0: %	are sampled from a single isotropic Gaussian with inverse variance
Daniel@0: %	equal to PRIOR. If PRIOR is a data structure of the kind generated by
Daniel@0: %	MLPPRIOR, then the parameters are sampled from multiple Gaussians
Daniel@0: %	according to their groupings (defined by the INDEX field) with
Daniel@0: %	corresponding variances (defined by the ALPHA field).
Daniel@0: %
Daniel@0: %	See also
Daniel@0: %	MLPPRIOR, NETUNPAK, RBFPRIOR
Daniel@0: %
Daniel@0: 
Daniel@0: %	Copyright (c) Ian T Nabney (1996-2001)
Daniel@0: 
Daniel@0: if isstruct(prior)
Daniel@0:     if (isfield(net, 'mask'))
Daniel@0: 	if find(sum(prior.index, 2)) ~= find(net.mask)
Daniel@0: 	    error('Index does not match mask');
Daniel@0: 	end
Daniel@0: 	sig = sqrt(prior.index*prior.alpha);
Daniel@0: 	% Weights corresponding to zeros in mask will not be used anyway
Daniel@0: 	% Set their priors to one to avoid division by zero
Daniel@0: 	sig = sig + (sig == 0);  
Daniel@0: 	sig = 1./sqrt(sig);
Daniel@0:     else
Daniel@0: 	sig = 1./sqrt(prior.index*prior.alpha);
Daniel@0:     end
Daniel@0:     w = sig'.*randn(1, net.nwts); 
Daniel@0: elseif size(prior) == [1 1]
Daniel@0:   w = randn(1, net.nwts).*sqrt(1/prior);
Daniel@0: else
Daniel@0:   error('prior must be a scalar or a structure');
Daniel@0: end  
Daniel@0: 
Daniel@0: if (isfield(net, 'mask'))
Daniel@0:     w = w(logical(net.mask));
Daniel@0: end
Daniel@0: net = netunpak(net, w);
Daniel@0: