msmo.m

很好用的统计模式识别工具箱

function [malpha,mbias,nsv,kercnt] = msmo(X,I,ker,arg,C,epsilon,tolerance,info)
% MSMO solves multiclass SVM problem by the SMO.
% [malpha,mbias,nsv,kercnt] = msmo(X,I,ker,arg,C,epsilon,tolerance,info)
%
% MSMO solves multiclass Support Vector Problem problem 
%  by the use of Sequential Minimal Optimizer.
%  The input K-class problem is decomposed onto 
%  K binary problems: i-th class against remaining (K-1)
%  classes. The binary problems are solved by the use 
%  of the SMO which results to the K decision rules.
%  
% Mandatory inputs:
%  X [NxL] contains L training N-dimensional patterns.
%  I [1xL] labels of the patterns; integers from 1 to K.
%  ker [string] kernel identifier (see "help smoc").
%  arg [...] arguments of given kernel (see "help smoc").
%  C [1x1] trade-off between maximal margin and training error.
%
% Optional inputs:
%  epsilon [real] tolerance of KKT-conditions fulfilment 
%    (default is 0.001).
%  tolerance [real] minimal change of optimized Lagrangeians 
%    (default is 0.001).
%  info [int] if info is 1 then prints which decision rule is
%    being built (default is 0).
%
% Output: 
%  malpha [KxL] found Lagrange multipliers.
%  mbias [Kx1] found biases.
%  nsv [int] number of SVs
%  kercnt [int] number of kernel evaluations.
%  
% For instance malpha{i} and mbias{i} are the multipliers and
% bias for i-th binary problem: i-th class against the others.
%
% See also MSVMCLASS, SMO, SVM.
%

% Statistical Pattern Recognition Toolbox, Vojtech Franc, Vaclav Hlavac
% (c) Czech Technical University Prague, http://cmp.felk.cvut.cz
% Modifications
%  21-Oct-2001, V.Franc
%  18-September-2001, V. Franc, created


% -- Processing of the input arguments ----------------

if nargin < 5,
   error('Not enough input arguments.');
end

if nargin < 6,
   epsilon = 0.001;
end

if nargin < 7,
   tolerance = 0.001;
end

if nargin < 8, 
  info = 0;
end


% -- Init variables -------------------------------

K = max( I );    % maximal class label
L= size(X,2);

% preallocate variables
malpha=zeros(K,L);
mbias=zeros( K,1 );

% -- Build decision rules -------------------------------

kercnt = 0;

for i= 1:K,        

   if info == 1,
      disp( sprintf('Building rule %d-th of %d.', i, K));
   end
  
   Idich=multi2dicho( I, i );
   
   [malpha(i,:),mbias(i),nsv,akercnt]=smo(X,Idich,ker,arg,C,epsilon,tolerance);
   
   kercnt = kercnt + akercnt;
   
end

nsv = length( find( sum(malpha)));

return;