libsvm简易说明

@nlpr.ia.ac.cn§§vSVM§ANNSVM§SVM§SVMvSVM§LibSVM§LibSVM§PythonLibSVM§——§——§v§n§§v§ANN&SVM1122(,),(,),(,)nnxyxyxyL*(,)fxw()(,(,))(,)RwLyfxwdFxy=∫(SRM)(ERM)SVMANN§——§——vSVM§§§SVM§——Lagrangeywxb=⋅+margin2/w=*{,},1,...,{1,1},diiiixyinyxR=∈-∈,111min()2..()1(1,2,...,)nnijijijiijiiiyyxxstywxbinααα==⋅-⋅≥=∑∑12min||||2..()1(1,2,...,)iiwstywxbin⋅≥=211min(||||)2..()1(1,2,...,),0niiiiiiwstywxbnCiξξξ=+⋅+≥-=≥∑****1()sgn{}sgn{()}niiiifxwxbyxxbα==⋅+=⋅+∑§§NO!****1(),()()sgn{}sgn{(()())}niiiihxgxfhwhbyxxbϕαϕϕ====⋅+=⋅+∑()()ixxϕϕ⋅§§§vSVM(()())(,)()ijijijxxKxxxxϕϕφ⋅==⋅sigmoidRBFPolynomialLinear()dTijxxrγ+(,)TijijKxxxx=2exp()ijxxγ--tanh()Tijxxrγ+()1111i1min,2..00,1,nnnijijijjijjniiiyyKxxstyilCαααααα====-=≤≤=∑∑∑∑L***11()sgn((,))),(,){|0}lliiijiiijjiifxyKxxbbyyKxxjCjααα===+=-∈∑∑(~cjlin/libsvm)§DifferentSVMformulations§Efficientmulti-classclassification§Crossvalidationformodelselection§Probabilityestimates§WeightedSVMforunbalanceddata§BothC++andJavasources§GUIdemonstratingSVMclassificationandregression§Python,R(alsoSplus),MATLAB,Perl,Ruby,Weka,CommonLISPandLabVIEWinterfaces.C#.NETcodeisavailable.§Automaticmodelselectionwhichcangeneratecontourofcrossvaliationaccuracy.vLibSVM§•Win32+python+pnuplot•Linux+python+pnuplot§•TrainingSet•TestSet§SVMSVM——LibSVM§labelfeature1:value1index2:value2...•labelfeaturevalue•value0§iris.trUCI/IrisPlant,4features,3classesvcheckdata.pypythoncheckdata.pyiris.tr.txt11:-0.5555562:0.53:-0.6949154:-0.7531:-0.1666672:-0.3333333:0.389834:0.91666721:-0.3333332:-0.753:0.01694914:-4.03573e-0811:-0.8333333:-0.8644074:-0.91666711:-0.6111112:0.08333333:-0.8644074:-0.91666731:0.6111112:0.3333333:0.7288134:131:0.2222223:0.389834:0.58333321:0.2222222:-0.3333333:0.2203394:0.16666721:-0.2222222:-0.3333333:0.1864414:-4.03573e-08…§§vv§§§,*,,ijijijixxx=∑,*,mijjijjjxxMm-=-,*,ijjijjxxxS-=§svm-scale[options]filenameoptions:•[-1,1]-llower-uupper•-sscalefile•-rscalefile•-ylowerupper§•scale[lower,upper]•§svm-scale-siris.scaleiris.trainsvm-scale-l-0.8-u0.8-siris.scaleiris.trainiris.train.scaledsvm-scale-riris.scaleiris.test§svm-train[options]filename[modelfile]options•-ssvm_type:settypeofSVM•-tkernel_type:settypeofkernelfunction•-vn:n-foldcrossvalidationmode•-ggamma():setgammainkernelfunction(default1/k)•-ccost:settheparameterCofC-SVC,epsilon-SVR,andnu-SVR(default1)•-bprobabilityestimates:whethertotrainaSVCorSVRmodelforprobabilityestimates,0or1(default0)•-mcachesize:setcachememorysizeinMB(default100)Epsilon-SVR3One-classSVM2one-classSVM1Nu-SVRC-SVC40(default)sigmoid3radialbasisfunction(RBF)2(default)polynomial1linear0㧧(C-SVC)(RBF)10-fold,c=100,g=0.01,§(C-SVC)10-fold,c=100§v§libSVM§svm-train-v10iris.trainiris.modelsvm-train-t0–v10–c100–g0.01iris.trainiris.modelsvm-trainiris.trainsvm-train–b1-t0–v10–c100–g0.01iris.trainiris.model(Cross-validationandGrid-search)§&RBF(linear)§python[options]grid.pytrainingfilename§options•-svmtrainpathname•-gnuplotpathname•-outpathname•-pngpathname•-log2cbegin,end,step•-log2gbegin,end,step•additionalsameoptionsforsvm-train§§linear——gamma:-log2g1,1,1pythongrid.pyiris.trainpythongrid.py–svmtraind:\libsvm\svm-train.exe–gnuplotd:\gnuplot\bin\pgnuplot.exe-pngd:\iris.gird.png–log2c-8,8,2–log2g8,-8,-2–v10iris.trainpythongrid.py–log2c-8,8,2–log2g1,1,1–t0–v10iris.train[options]testfilemodelfileresultfileoptions§-bprobability:-b0-b1v§-b1§vv§SVR,one-classSVM§-bsvm-train§svm-train–b1iris.trainiris.modelsvm-predictiris.testiris.modeliris.resultsvm-predict–b1iris.testiris.modeliris.resultsvm-scale-riris.train.scaleiris.testiris.test.scaled§labelv§grid.pyàgridregression.py§-c-gà-c-g–pv§-s3epsilon-SVR§-t2RBF-t0linear§pv§RateàMSE§svm-scale-y-11regression.train.scaledregression.modelsvm-train–s3-c100–g0.01–p0.1regression.train.scaledregression.modelpythongridregression.py–log2c-8,8,2–log2g8,-8,-2–log2p-8,8,2–v10regression.trainsvm-predictregression.testregression.modelregression.result§svmc.pydlibsvm\windows\python\python§svm.py§cross_validation.pyv§§svm_type,kernel_type,gamma,C…§fromsvmimport*param=svm_parameter(svm_type=C_SVC,kernel_type=LINEAR)param.kernel_type=RBFListLable=[1,-1]ListValue=[[1,0,1],[-1,0,-1]]#ListValue=[{1:1,3:1},{1:-1,3:-1}]prob=svm_problem(ListLabel,ListValue)§§§mod=svm_model(prob,param)r=mod.predict([1,1,1])d=mod.predict_values([1,1,1])prd,prb=m.predict_probability([1,1,1])mod.save(‘modelfile’)mod2=svm_model(‘modelfile’)target=cross_validation(prob,param,n)