tuningkerex7.m

关于svm的例程

%
%  SVM Classification 2D examples
%
%
%  Optimizing the kernel parameters by means of of gradient descent
%  with regards to the leave one out error from R^2w^2.
%
%  See O. Chapelle, V. Vapnik  
%
%  20/08/00 AR


close all
clear all
clc

chemin='\code\data\breast\';
data='breast-cancer';
number='34';
ext='.asc';

xapp=load([chemin data '_train_data_' number ext]) ;
yapp=load([chemin data '_train_labels_' number ext]) ;
xtest=load([chemin data '_test_data_' number ext]) ;
ytest=load([chemin data '_test_labels_' number ext]) ;







[nbapp,nbvar]=size(xapp);
nbtest=length(ytest);


%-----------------------------------------------------
%   Learning and Learning Parameters

c=1000000;
C=1;
lambda =1e-10;
kerneloption=[ones(1,nbvar) C];
kerneloption1=kerneloption;
kernel='gaussian';
verbose = 0;


eta=0.01;
nbiter=50;

%[xsup,w,b,pos,kerneloptiontuned,nberreur,dTold,kerpar]=svmclasstuned(xapp,yapp,c,lambda,kernel,kerneloption,verbose,xtest,ytest,...
%     nbiter);
%
[xsup,w,b,pos,kerneloptiontuned,r2w2,dif]=svmclasstunedr2w2(xapp,yapp,c,lambda,kernel,...
    kerneloption,verbose,nbiter);

% [xsup,w,b,pos,kerneloptiontuned]=svmclasstunedspanbound(xapp,yapp,c,lambda,kernel,...
% kerneloption,eta,verbose,nbiter);


ypredtuned= svmval(xtest,xsup,w,b,kernel,kerneloptiontuned);
[xsup1,w1,b1,pos]=svmclass(xapp,yapp,c,lambda,kernel,kerneloption1,0);
ypredapp = svmval(xtest,xsup1,w1,b1,kernel,kerneloption1);                                           
erruntuned=sum( (-ypredapp.*ytest)> 0);
errtuned=sum( -(ypredtuned.*ytest)>0);
fprintf('test error for tuned hyperparam.: %f  for untuned hyperparam : %f\n',errtuned/nbtest,erruntuned/nbtest);