multicSVMmul.zip_SVM多分类matlab_svm多分类_多分类_多分类SVM

function [oput_par]=multicSVMmul(X,pre_param,burnin,num,space,dataname) %%% learn the supervised factor loading and score %%% X: PxN, the observed data matrix, each column corresponds a data sample %%% Y: Nx1, the label vector %%% pre_param: predefined hyperparameters %%% pre_param.K: predifined number of %%% pre_param.L: predefined cost loss %%% pre_param.C: trade off parameter %%% Written by Bo Chen %%Initialize the paramters if min(pre_param.trainv) == -1 pre_param.trainv(pre_param.trainv==-1) = 2; pre_param.testv(pre_param.testv==-1) = 2; end randn('state',0); rand('state',0); [P,XN]=size(X);%%%size of text data trX=X(:,pre_param.trindv); K=size(trX,1); Y=pre_param.trainv; M = max(Y); N=length(Y); oput_par.eta=randn(K+1,M); oput_par.S=X; % Initialize the SVM parameters yind = sub2ind([N,M],(1:N)',Y); bY = -ones(N,M); % N x M bY(yind) = 1; PX=[trX;ones(1,N)]'*oput_par.eta; oput_par.Lres=pre_param.L-PX.*bY; oput_par.invlamda = 1./abs( pre_param.C*oput_par.Lres); oput_par.etasigma=gamrnd(pre_param.etasigmaa*ones(K+1,M),1/pre_param.etasigmab); maxit = burnin + num*space; iter = 0; numcol=0; oldDW=0; oput_par.avertrainacc=zeros(1,maxit); trtime = 0; tetime = 0; Nte = length(pre_param.teindv); while (iter<maxit) tic iter = iter + 1; %% sample augumented parameters for m=1:M [oput_par]=sampleLamdaFS(pre_param,oput_par,bY(:,m),m); [oput_par]=samplesvmetamul(pre_param,oput_par,bY(:,m),m); oput_par.Lres(:,m)=pre_param.L-[trX;ones(1,N)]'*oput_par.eta(:,m).*bY(:,m); end [oput_par]=samplesvmetasigma(pre_param,oput_par); trtime = trtime + toc; %%% DW=[X;ones(1,XN)]'*oput_par.eta; oput_par.Lmse(iter)=sum(max(oput_par.Lres,[],2)); [~, ty] = max(DW, [], 2); oput_par.trainacc(iter)= mean(ty(pre_param.trindv)==pre_param.trainv); oput_par.testacc(iter)=mean(ty(pre_param.teindv)==pre_param.testv); ndx = iter - burnin; test = mod(ndx,space); oput_par.avtestacc(iter)=0; tic if (ndx>0) && (test==0) DW=[X(:,pre_param.teindv);ones(1,Nte)]'*oput_par.eta; oldDW=oldDW+DW; numcol=numcol+1; oput_par.collecteta{numcol}=oput_par.eta; [~, ty] = max(oldDW, [], 2); oput_par.avtestacc(iter) = mean(ty==pre_param.testv); end tetime = tetime + toc; % if mod(iter,20)==0 % fprintf('%s: iter %d Lmse=%g tr=%g trav=%g te=%g teav=%g time=%g\n',... % dataname,iter,oput_par.Lmse(iter),oput_par.trainacc(iter),oput_par.avertrainacc(iter),oput_par.testacc(iter),... % oput_par.avtestacc(iter),toc); % end % fprintf('%s: iter %d liklihood=%g mse=%g tracc=%g time=%g\n',dataname,iter,liklihood,oput_par.mse(iter),oput_par.trainacc(iter),toc); end oput_par.trtime = trtime; oput_par.tetime = tetime; oput_par.avtrtime = oput_par.trtime/(burnin+num*space); oput_par.avtetime = oput_par.tetime/(num); xx = oput_par.avtestacc(oput_par.avtestacc>0); oput_par.acc = xx(end);