wolffd@0: function [means, covs, weights, ll] = gmmem_multi_restart(K, data, varargin)
wolffd@0: % GMMEM_MULTI_RESTART Multiple restart wrapper for gmmem_kpm
wolffd@0: % function [means, covs, weights, ll] = gmmem_multi_restart(K, data, varargin)
wolffd@0: %
wolffd@0: % Input:
wolffd@0: % K = number of mixture components
wolffd@0: % data(i,:) is the i'th example (feature vector)
wolffd@0: %
wolffd@0: % Output:
wolffd@0: % The parameters for the k'th mixture component, k=1:K, are
wolffd@0: %   means(k,:), covs(:,:,k) and weights(k)
wolffd@0: %
wolffd@0: % [ ... ] = gmmem_multi_restart(..., 'param1',val1, 'param2',val2, ...)
wolffd@0: % allows you to specify optional parameter name/value pairs.
wolffd@0: % Parameters are below [default value in brackets]
wolffd@0: %
wolffd@0: % 'nrestarts' - number of EM restarts [2]
wolffd@0: % 'cov_type' - 'full', 'diag' or 'spherical' ['full']
wolffd@0: % 'init_cov' - the initial covariance matrix [0.1*cov(data) for each k]
wolffd@0: % 'init_means' - [] means sample from randn(); otherwise, use
wolffd@0: %               init_means(k,:,r) for the k'th comp. on the r'th restart [ [] ]
wolffd@0: % 'restartfn'  - this function, if non-empty,  will be called before/after every restart
wolffd@0: %               (e.g., to display the parameters as they evolve) [ [] ]
wolffd@0: %               The fn is called as fn(mix{r}, data, restart_num, niter, outerfnargs)
wolffd@0: %               where niter is the number of iterations performed (0 initially)
wolffd@0: % 'restartfnargs' - additional arguments to be passed to restartfn [ {} ]
wolffd@0: %
wolffd@0: % Optional arguments for gmmem_kpm are passed through.
wolffd@0: %
wolffd@0: % Written by Kevin P Murphy, 30 Dec 2002
wolffd@0: 
wolffd@0: [ndata nfeatures] = size(data);
wolffd@0: 
wolffd@0: %Cinit = repmat(0.1*diag(diag(cov(data))), [1 1 K]);
wolffd@0: Cinit = repmat(0.1*cov(data), [1 1 K]);
wolffd@0: 
wolffd@0: [nrestarts, init_cov, init_means, cov_type, ...
wolffd@0:  restartfn, restartfnargs, unused_args] = ...
wolffd@0:     process_options(varargin, ...
wolffd@0: 	'nrestarts', 2, 'init_cov',  Cinit,  'init_means', [], ...
wolffd@0: 	'cov_type', 'full', 'restartfn', [], 'restartfnargs', {});
wolffd@0: 
wolffd@0: mix = cell(1, nrestarts);
wolffd@0: cost = inf*ones(1,nrestarts);
wolffd@0: 
wolffd@0: for r=1:nrestarts
wolffd@0:   mix{r} = gmm(nfeatures, K, cov_type); % random centers
wolffd@0:   if ~isempty(init_means), mix{r}.centres = init_means(:,:,r); end
wolffd@0:   mix{r}.covars = init_cov;
wolffd@0:   if ~isempty(restartfn)
wolffd@0:     feval(restartfn, mix{r}, data, r, 0, restartfnargs{:});
wolffd@0:   end
wolffd@0:   [mix{r}, niter, ll] = gmmem_kpm(mix{r}, data, unused_args{:}); 
wolffd@0:   cost(r) = -ll; %-sum(log(gmmprob(mix{r}, data)));
wolffd@0:   if ~isempty(restartfn)
wolffd@0:     feval(restartfn, mix{r}, data, r, niter, restartfnargs{:});
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: [nll, bestr] = min(cost);
wolffd@0: fprintf('best r = %d\n', bestr);
wolffd@0: ll = -nll;
wolffd@0: means = mix{bestr}.centres;
wolffd@0: covs = mix{bestr}.covars;
wolffd@0: weights = mix{bestr}.priors;