tweak kmeans to the extreme

chapter09/kmeans.m

@@ -1,12 +1,12 @@
-function [label, m, energy] = kmeans(X, init)
+function [label, mu, energy] = kmeans(X, init)
 % Perform kmeans clustering.
 % Input:
 %   X: d x n data matrix
 %   init: k number of clusters or label (1 x n vector)
 % Output:
 %   label: 1 x n cluster label
+%   mu: d x k center of clusters
 %   energy: optimization target value
-%   model: trained model structure
 % Written by Mo Chen ([email protected]).
 n = size(X,2);
 idx = 1:n;
@@ -19,8 +19,7 @@
 end
 while any(label ~= last)
     [~,~,last(:)] = unique(label);   % remove empty clusters
-    E = sparse(idx,last,1);  % transform label into indicator matrix
-    m = X*(E./sum(E,1));    % compute centers 
-    [val,label] = min(dot(m,m,1)'/2-m'*X,[],1); % assign labels
+    mu = X*normalize(sparse(idx,last,1),1);    % compute centers 
+    [val,label] = min(dot(mu,mu,1)'/2-mu'*X,[],1); % assign labels
 end
 energy = dot(X(:),X(:),1)+2*sum(val);

chapter09/kmeansPred.m

@@ -1,4 +1,4 @@
-function [label, energy] = kmeansPred(m, X)
+function [label, energy] = kmeansPred(mu, X)
 % Prediction for kmeans clusterng
 % Input:
 %   model: dx k cluster center matrix
@@ -7,5 +7,5 @@
 %   label: 1 x n cluster label
 %   energy: optimization target value
 % Written by Mo Chen ([email protected]).
-[val,label] = min(dot(X,X,1)+dot(m,m,1)'-2*m'*X,[],1); % assign labels
+[val,label] = min(dot(X,X,1)+dot(mu,mu,1)'-2*mu'*X,[],1); % assign labels
 energy = sum(val);

chapter09/litekmeans.m

@@ -0,0 +1,17 @@
+function [label, mu] = litekmeans(X, k)
+n = size(X,2);
+last = zeros(1,n);
+label = ceil(k*rand(1,n));
+while any(label ~= last)
+    [~,~,last(:)] = unique(label);            % remove empty clusters
+    mu = X*normalize(sparse(1:n,last,1),1);    % compute cluster centers 
+    [~,label] = min(dot(mu,mu,1)'/2-mu'*X,[],1); % assign sample labels
+end
+% Perform kmeans clustering.
+% Input:
+%   X: d x n data matrix
+%   k: number of clusters
+% Output:
+%   label: 1 x n cluster label
+%   mu: d x k center of clusters
+% Written by Mo Chen ([email protected]).

common/normalize.m

@@ -8,4 +8,4 @@
     if isempty(dim), dim = 1; end
 end
 s = sum(X,dim);
-Y = bsxfun(@times,X,1./s);
+Y = X./s;