bsvm2.m~

一个工具包

function [model]=msvm2( data, options )
% MSVM2 Multiclass SVM with 2-soft margin trained by single-class optimizer.
%
% Synopsis:
%  model = mssvm( data, options) 
%
% Description:
%  This function implements transformation of multi-class SVM problem
%  to the single-class SVM which is consequently solved by a seleceted
%  single-class SVM optimizers (e.g., Kozinec, SMO, KOZSMO).
%  
%  The transformation is based on the Kersler's construction [DHS01] which 
%  can be simply expresed in terms of kernel functions [Franc02a]. 
%  The resulting multi-class SVM has the same architecture as the one 
%  yieled by one-against-all decomposition. In constrast to the one-against-all 
%  decomposition all rules (discriminant functions) are trained simultanously.
%
%  See 'help oaasvmclass' for info about classification.
%
% Inputs:
%  data.X [dim x num_data] Training patterns.
%  data.y [1 x num_data] Labels of training patterns. The integers from
%   1 to max_class are assumed as the labels.
%  options.ker [string] Kernel identifier. See 'help kernel'.
%  options.arg [...] Arguments of selected kernel.
%  options.C [positiv real] Regularization constant.
%  options.solver [string] KOZINEC, SMO, KOZSMO
%  options.tmax [int]
%  options.tolabs [real] 
%  options.tolrel [real]
%
% Output:
%  model.rule [cell num_class x 1] Description of SVM discriminat functions.
%  model.num_classes [int] Number of classes.
%  model.num_rules [int] Number of classification rules == num_classes.
%  model.kercnt [int] Total number of used kernel evaluations.
%  model.cputime [int] Used CPU time.
%  model.options [struct] Copy of used options.
%  model.nsv [int] Number of unique support vectors.
%  model.sv.X [dim x nsv] Support vectors stored as columns.
%  model.sv.y [1 x nsv] Labels of Support vectors (not used for 
%   classification).
%  model.classifier [string] Name of classifier.
%
% Example:
%  data=load('threeclusters1');
%  model=msvm2(data,{'ker','linear','C',10})
%  ytrn = oaasvmclass( data.X, model);
%  trnerr=cerror(data.y,ytrn)
%  figure; ppatterns(data); pmsvm(model);
%
% See also 
%  OAASVMCLASS, OAOSVM, OAOSVMCLASS, PMSVM.
%

% About: 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:
% 23-jan-2003, VF

tic;
model.classifier = 'oaasvmclass';

% process input argumens
data=c2s(data);
if nargin < 2, options=[]; else options=c2s(options); end

if ~isfield(options,'ker'), options.ker='linear'; end
if ~isfield(options,'arg'), options.arg=1; end
if ~isfield(options,'C'), options.C=inf; end
if ~isfield(options,'tmax'), options.tmax=inf; end
if ~isfield(options,'tolabs'), options.tolabs=0; end
if ~isfield(options,'tolrel'), options.tolrel=0.001; end
if ~isfield(options,'solver'), options.solver='kozinec'; end

%---------------------------------

[dim,num_data]=size(data.X);
num_classes=max(data.y);

%---------------------------------
model.name = 'One-Against-All SVM classifier';

[ker,arg]=kerarg(options.ker,options.arg);

switch upper(options.solver)
   case 'KOZINEC', solver = 0;   
   case 'SMO', solver = 1;
   case 'KOZSMO', solver = 2;
   case 'NPA', solver = 3;
   otherwise error('Unknown solver.');
end

[Alpha,b,exitflag,kercnt,trnerr,t,UB,LB,History] = msvm2ccode(data.X,data.y,...
    ker,arg,options.C,solver,options.tmax,options.tolabs, options.tolrel);
 
model.num_classes = num_classes;
model.num_rules = num_classes;
model.rule = cell(model.num_rules,1);
model.exitflag = exitflag;
model.trnerr = trnerr;
model.kercnt = kercnt;
model.t = t;
model.UB = UB;
model.LB = LB;
model.LB_History = History(1,:);
model.UB_History = History(2,:);

for i=1:num_classes,
  model.rule{i}.b = b(i);
  inx=find(Alpha(i,:)~=0);
  model.rule{i}.Alpha = Alpha(i,inx);
  model.rule{i}.sv.inx = inx;
  lab=data.y(inx);
  lab2=lab; lab2(find(lab==i))=1;lab2(find(lab~=i))=2;
  model.rule{i}.Alpha = model.rule{i}.Alpha(:).*label12topm(lab2(:));
  model.rule{i}.options = options;
end

model.options = options;
model.sv = data;

model.nsv = length(find(sum(Alpha)));  
inx_sv = find(sum(abs(Alpha)));  

% takes only support vectors (maintain only one copy od support 
% vectors) and change indices
model.nsv = length(inx_sv);
inx_tmp = zeros(1,num_data);
inx_tmp(inx_sv) = [1:model.nsv];
model.sv.X = data.X(:,inx_sv);
model.sv.y = data.y(:,inx_sv);
for i=1:model.num_rules,
  model.rule{i}.sv.inx = inx_tmp(model.rule{i}.sv.inx);
end

model.cputime = toc;

return;

% EOF