function [weights,means,covars]= kmeans_init(data,K,numtries,permseed)
% KMEANS_INIT: [weights,means,covars]= kmeans_init(data,K,numtries,permseed)
%
%    Use k-means to find a good starting point for EM for Gaussian
%	mixture modeling.
%
%
% INPUT:
% -----
% data - matrix with feature vectors as rows
% k       - number of Gaussian components in the mixture
% numtries - number of different random starting points to try for kmeans
% permseed - random seed initializer
%
%-------------------------------------------------------------------------------

% size of data matrix
[n d] = size(data);

% Initialize
joints = zeros(n,K);
epsilon = 0.0001;


 %t0 = clock;
 for i=1:numtries
	% Initialize means and sigmas using k-means
	[labels,means,kseed] = kmeans(data, K,permseed*i);
        seed(i) = kseed;
        [weights,covars] = kmeans_params(K,data,labels,n,d,epsilon);

	% Compute conditional likelihood p(x|wi) of validation data 
	%   using the means and covariance matrices of the mixture we just found
	for k = 1:K
		covar_k = covars(:,((k-1)*d+1):(k*d));
		joints(:,k) = pgauss_chol(data, means(k,:), covar_k) * weights(k);
	end
        f_x = sum(joints');
	ll(i) = sum(log(f_x)');
  end
     %fprintf('time for kmeans, K=%d: %f\n',K,etime(clock,t0));
 
%  Pick the kmeans partition with the highest likelihood - a hack
%  to try to avoid local minima
        [llmaxi maxi] = max(ll);
	[labels,means,kseed] = kmeans(data, K,seed(maxi));
        [weights,covars] = kmeans_params(K,data,labels,n,d,epsilon);