function [model,y] = kmeans(X,num_centers,Init_centers)% KMEANS K-means clustering algorithm.% % Synopsis:%  [model,y] = kmeans(X,num_centers)%  [model,y] = kmeans(X,num_centers,Init_centers)%% Description:%  [model,y] = kmeans(X,num_centers) runs K-means clustering %   where inital centers are randomly selected from the %   input vectors X. The output are found centers stored in %   structure model.%   %  [model,y] = kmeans(X,num_centers,Init_centers) uses%   init_centers as the starting point.%% Input:%  X [dim x num_data] Input vectors.%  num_centers [1x1] Number of centers.%  Init_centers [1x1] Starting point of the algorithm.%    % Output:%  model [struct] Found clustering:%   .X [dim x num_centers] Found centers.%%   .y [1 x num_centers] Implicitly added labels 1..num_centers.%   .t [1x1] Number of iterations.%   .MsErr [1xt] Mean-Square error at each iteration.%%  y [1 x num_data] Labels assigned to data according to %   the nearest center.%% Example:%  data = load('riply_trn');%  [model,data.y] = kmeans( data.X, 4 );%  figure; ppatterns(data); %  ppatterns(model,12); pboundary( model );%% See also %  EMGMM, KNNCLASS.%% (c) Statistical Pattern Recognition Toolbox, (C) 1999-2003,% Written by Vojtech Franc and Vaclav Hlavac,% <a href="http://www.cvut.cz">Czech Technical University Prague</a>,% <a href="http://www.feld.cvut.cz">Faculty of Electrical engineering</a>,% <a href="http://cmp.felk.cvut.cz">Center for Machine Perception</a>% Modifications:% 12-may-2004, VF[dim,num_data] = size(X);% random inicialization of class centers%-----------------------------------------------if nargin < 3,  inx=randperm(num_data);  model.X = X(:,inx(1:num_centers));  model.y = 1:num_centers;  model.K = 1;endmodel.fun = 'knnclass';old_y = zeros(1,num_data);t = 0;% main loop%-------------------------while 1,    t = t+1;    % classificitation  y = knnclass( X, model );    % computation of class centers  err = 0;  for i=1:num_centers,    inx = find(y == i);    if ~isempty(inx),            % compute approximation error      err = err + sum(sum((X(:,inx) - model.X(:,i)*ones(1,length(inx)) ).^2));            % compute new centers      model.X(:,i) = sum(X(:,inx),2)/length(inx);    end  end  % Number of iterations and Mean-Square Error   model.t = t;  model.MsErr(t) = err/num_data;    if sum( abs(y - old_y) ) == 0,    return;  end  old_y = y;endreturn;% EOF