Daniel@0: function [mu, Sigma, weights] = mixgauss_init(M, data, cov_type, method)
Daniel@0: % MIXGAUSS_INIT Initial parameter estimates for a mixture of Gaussians
Daniel@0: % function [mu, Sigma, weights] = mixgauss_init(M, data, cov_type. method)
Daniel@0: %
Daniel@0: % INPUTS:
Daniel@0: % data(:,t) is the t'th example
Daniel@0: % M = num. mixture components
Daniel@0: % cov_type = 'full', 'diag' or 'spherical'
Daniel@0: % method = 'rnd' (choose centers randomly from data) or 'kmeans' (needs netlab)
Daniel@0: %
Daniel@0: % OUTPUTS:
Daniel@0: % mu(:,k) 
Daniel@0: % Sigma(:,:,k) 
Daniel@0: % weights(k)
Daniel@0: 
Daniel@0: if nargin < 4, method = 'kmeans'; end
Daniel@0: 
Daniel@0: [d T] = size(data);
Daniel@0: data = reshape(data, d, T); % in case it is data(:, t, sequence_num)
Daniel@0: 
Daniel@0: switch method
Daniel@0:  case 'rnd', 
Daniel@0:   C = cov(data');
Daniel@0:   Sigma = repmat(diag(diag(C))*0.5, [1 1 M]);
Daniel@0:   % Initialize each mean to a random data point
Daniel@0:   indices = randperm(T);
Daniel@0:   mu = data(:,indices(1:M));
Daniel@0:   weights = normalise(ones(M,1));
Daniel@0:  case 'kmeans',
Daniel@0:   mix = gmm(d, M, cov_type);
Daniel@0:   options = foptions;
Daniel@0:   max_iter = 5;
Daniel@0:   options(1) = -1; % be quiet!
Daniel@0:   options(14) = max_iter;
Daniel@0:   mix = gmminit(mix, data', options);
Daniel@0:   mu = reshape(mix.centres', [d M]);
Daniel@0:   weights = mix.priors(:);
Daniel@0:   for m=1:M
Daniel@0:     switch cov_type
Daniel@0:      case 'diag',
Daniel@0:       Sigma(:,:,m) = diag(mix.covars(m,:));
Daniel@0:      case 'full',
Daniel@0:       Sigma(:,:,m) = mix.covars(:,:,m);
Daniel@0:      case 'spherical',
Daniel@0:       Sigma(:,:,m) = mix.covars(m) * eye(d);
Daniel@0:     end
Daniel@0:   end
Daniel@0: end
Daniel@0: