SVM实现MNIST数据集分类

function [xapp,yapp,xtest,ytest,xtest1,xtest2]=dataset(n,nbapp,nbtest,sigma) % Usage % [xapp,yapp,xtest,ytest,xtest1,xtest2]=dataset(String,nbapp,nbtest,sigma) % % % % 'Mixture' : 2 classes % First class is a a mixture of gaussian % which takes in sandwiches the second class % % % 'Gaussian' : is a classical dataset with two gaussianly distributed % classes with similar variances but different means % % 'Checkers': Checkers % % 'MultiClassCheckers' % % 'Clowns' : The Clowns % % 'CosExp': sign( 0.95*cos(0.5*(exp(x)-1)))); for x in [0...4] % xtest1,xtest2 meshgrid of xtest xapp=[]; yapp=[]; xtest=[]; ytest=[]; xtest1=[]; xtest2=[]; switch lower(n) case 'mixture' %------------------Data Learn Generation ------------------------- if nargin <4 sigma=0.4; end; nbapp=round(nbapp/3); x1=sigma*randn(1,nbapp)+0.3; x2=sigma*randn(1,nbapp)-0.3; x3=sigma*randn(1,nbapp)-1; y1=sigma*randn(1,nbapp)+0.5; y2=sigma*randn(1,nbapp)-0.5; y3=sigma*randn(1,nbapp)-1; xapp=[x1 x2 x3; y1 y2 y3]'; yapp=[ones(1,nbapp) -ones(1,nbapp) ones(1,nbapp)]'; %----------------Data Test Generation --------------------------- nbtest=round(nbtest/3); xt=sigma*randn(1,nbtest)+0.3; xt2=sigma*randn(1,nbtest)-0.3; xt3=sigma*randn(1,nbtest)-1; yt=sigma*randn(1,nbtest)+0.5+0.7*xt; yt2=sigma*randn(1,nbtest)-0.5; yt3=sigma*randn(1,nbtest)-1-0.7*xt; xtest=[xt xt2 xt3; yt yt2 yt3]'; ytest=[ones(1,nbtest) -ones(1,nbtest) ones(1,nbtest)]'; return case 'gaussian' nbapp=round(nbapp/2); nbtest=round(nbtest/2); if nargin <4 sigma=0.2; end; x1=sigma*randn(1,nbapp)+0.3; x2=sigma*randn(1,nbapp)-0.3; y1=sigma*randn(1,nbapp)+0.5; y2=sigma*randn(1,nbapp)-0.5; xapp=[x1 x2; y1 y2]'; yapp=[ones(1,nbapp) -ones(1,nbapp)]'; x1=sigma*randn(1,nbtest)+0.3; x2=sigma*randn(1,nbtest)-0.3; y1=sigma*randn(1,nbtest)+0.5; y2=sigma*randn(1,nbtest)-0.5; xtest=[x1 x2; y1 y2]'; ytest=[ones(1,nbtest) -ones(1,nbtest)]'; case 'checkers' xapp=[]; yapp=[]; nb=floor(nbapp/16); for i=-2:1; for j=-2:1; xapp=[xapp; [i+rand(nb,1) j+rand(nb,1)]]; %(2*rem((i+j+4),2)-1) yapp=[yapp; (2*rem((i+j+4),2)-1)*ones(nb,1)]; end; end; xtest=[]; ytest=[]; nb=floor(nbtest/16); for i=-2:1; for j=-2:1; xtest=[xtest; [i+rand(nb,1) j+rand(nb,1)]]; %(2*rem((i+j+4),2)-1) ytest=[ytest; (2*rem((i+j+4),2)-1)*ones(nb,1)]; end; end; % [xt1,xt2]=meshgrid(linspace(-2,2,sqrt(nbtest))); % [na,nb]=size(xt1); % xtest1=reshape(xt1,na*nb,1); % xtest2=reshape(xt2,na*nb,1); % xtest=[xtest1 xtest2]; % ytest=zeros(length(xtest),1); % for i=-2:1 % for j=-2:1 % pos=find( xtest(:,1)>=i & xtest(:,1)< i+1 & xtest(:,2)>=j & xtest(:,2)< j+1 ); % ytest(pos)=(2*rem((i+j+4),2)-1)*ones(length(pos),1); % end; % end; % xtest1=reshape(xt1,na,nb); % xtest2=reshape(xt2,na,nb); case 'clowns' nbapp=round(nbapp/2); nbtest=round(nbtest/2); x1 = (6*rand(nbapp,1)-3); x2 = x1.^2 + randn(nbapp,1); x0 = sigma*randn(nbapp,2)+(ones(nbapp,1)*[0 6]); xapp = [x0;[x1 x2]]; xapp = (xapp-ones(2*nbapp,1)*mean(xapp))*diag(1./std(xapp)); yapp = [ones(nbapp,1); -ones(nbapp,1)]; if nbtest>0 x1 = (6*rand(nbtest,1)-3); x2 = x1.^2 + randn(nbtest,1); x0 = sigma*randn(nbtest,2)+(ones(nbtest,1)*[0 6]); xtest = [x0;[x1 x2]]; xtest = (xtest-ones(2*nbtest,1)*mean(xtest))*diag(1./std(xtest)); ytest = [ones(nbtest,1); -ones(nbtest,1)]; end; case 'cosexp' maxx=5; freq=0.7; nbtest=floor(sqrt(nbtest)); [xi]=rand(nbapp,2); yapp= sign(cos(0.5*(exp(freq*maxx*xi(:,1))))-(2*xi(:,2)-1)); xapp = [maxx*xi(:,1) (2*xi(:,2)-1)]; [xtest1 xtest2] = meshgrid(linspace(0,maxx,nbtest),linspace(-1,1,nbtest)); nn = length(xtest1); xtest = [reshape(xtest1 ,nn*nn,1) reshape(xtest2 ,nn*nn,1)]; ytest = sign(cos(0.5*(exp(freq*maxx*xtest(:,1))))-(2*xtest(:,2)-1)); case 'multiclasscheckers' xapp=[]; yapp=[]; nb=floor(nbapp/16); % keyboard for i=-2:1; for j=-2:1; xapp=[xapp; [i+rand(nb,1) j+rand(nb,1)]]; if rem(abs(i),2)==0 & rem(abs(j),2)==0 yapp=[yapp; 1*ones(nb,1)]; elseif rem(abs(i),2)==1 & rem(abs(j),2)==0 yapp=[yapp; 2*ones(nb,1)]; elseif rem(abs(i),2)==0 & rem(abs(j),2)==1 yapp=[yapp; 3*ones(nb,1)]; elseif rem(abs(i),2)==1 & rem(abs(j),2)==1 yapp=[yapp; 4*ones(nb,1)]; end; end; end; xtest=[]; [xt1,xt2]=meshgrid(linspace(-2,2,sqrt(nbtest))); [na,nb]=size(xt1); xtest1=reshape(xt1,na*nb,1); xtest2=reshape(xt2,na*nb,1); xtest=[xtest1 xtest2]; ytest=zeros(length(xtest),1); for i=-2:1 for j=-2:1 pos=find( xtest(:,1)>=i & xtest(:,1)< i+1 & xtest(:,2)>=j & xtest(:,2)< j+1 ); %ytest(pos)=(2*rem((i+j+4),2)-1)*ones(length(pos),1); if rem(abs(i),2)==0 & rem(abs(j),2)==0 ytest(pos)=1*ones(length(pos),1); elseif rem(abs(i),2)==1 & rem(abs(j),2)==0 ytest(pos)=2*ones(length(pos),1); elseif rem(abs(i),2)==0 & rem(abs(j),2)==1 ytest(pos)=3*ones(length(pos),1); elseif rem(abs(i),2)==1 & rem(abs(j),2)==1 ytest(pos)=4*ones(length(pos),1); end; end; end; xtest1=reshape(xt1,na,nb); xtest2=reshape(xt2,na,nb); case 'multiclassgaussian' mean1=[1 1]; mean2=[-1 1]; mean3= [0 -1]; x1=sigma*randn(nbapp,2)+ones(nbapp,1)*mean1; y1= ones(nbapp,1); x2=sigma*randn(nbapp,2)+ones(nbapp,1)*mean2; y2=2*ones(nbapp,1); x3=sigma*randn(nbapp,2)+ones(nbapp,1)*mean3; y3=3*ones(nbapp,1); Ytarget=[1 2 3]; xapp=[x1;x2;x3]; yapp=[y1;y2;y3]; nbapp=size(xapp,1); %----------------Data Test Generation --------------------------- x1=sigma*randn(nbtest,2)+ones(nbtest,1)*mean1; y1= ones(nbtest,1); x2=sigma*randn(nbtest,2)+ones(nbtest,1)*mean2; y2=2*ones(nbtest,1); x3=sigma*randn(nbtest,2)+ones(nbtest,1)*mean3; y3=3*ones(nbtest,1); Ytarget=[1 2 3]; xtest=[x1;x2;x3]; ytest=[y1;y2;y3]; nbapp=size(xapp,1); case 'westonnonlinear' % dataset used in the Weston FeatSel Paper