LibLinear.cpp

Go to the documentation of this file.
00001 /*
00002  * This program is free software; you can redistribute it and/or modify
00003  * it under the terms of the GNU General Public License as published by
00004  * the Free Software Foundation; either version 3 of the License, or
00005  * (at your option) any later version.
00006  *
00007  * Written (W) 1999-2009 Soeren Sonnenburg
00008  * Copyright (C) 1999-2009 Fraunhofer Institute FIRST and Max-Planck-Society
00009  */
00010 #include "lib/config.h"
00011 
00012 #ifdef HAVE_LAPACK
00013 #include "lib/io.h"
00014 #include "classifier/svm/LibLinear.h"
00015 #include "classifier/svm/SVM_linear.h"
00016 #include "classifier/svm/Tron.h"
00017 #include "features/DotFeatures.h"
00018 
00019 CLibLinear::CLibLinear(LIBLINEAR_LOSS l)
00020 : CLinearClassifier()
00021 {
00022     loss=l;
00023     use_bias=false;
00024     C1=1;
00025     C2=1;
00026 }
00027 
00028 CLibLinear::CLibLinear(
00029     float64_t C, CDotFeatures* traindat, CLabels* trainlab)
00030 : CLinearClassifier(), C1(C), C2(C), use_bias(true), epsilon(1e-5)
00031 {
00032     set_features(traindat);
00033     set_labels(trainlab);
00034     loss=LR;
00035 }
00036 
00037 
00038 CLibLinear::~CLibLinear()
00039 {
00040 }
00041 
00042 bool CLibLinear::train()
00043 {
00044     ASSERT(labels);
00045     ASSERT(features);
00046     ASSERT(labels->is_two_class_labeling());
00047 
00048     int32_t num_train_labels=labels->get_num_labels();
00049     int32_t num_feat=features->get_dim_feature_space();
00050     int32_t num_vec=features->get_num_vectors();
00051 
00052     ASSERT(num_vec==num_train_labels);
00053     delete[] w;
00054     if (use_bias)
00055         w=new float64_t[num_feat+1];
00056     else
00057         w=new float64_t[num_feat+0];
00058     w_dim=num_feat;
00059 
00060     problem prob;
00061     if (use_bias)
00062     {
00063         prob.n=w_dim+1;
00064         memset(w, 0, sizeof(float64_t)*(w_dim+1));
00065     }
00066     else
00067     {
00068         prob.n=w_dim;
00069         memset(w, 0, sizeof(float64_t)*(w_dim+0));
00070     }
00071     prob.l=num_vec;
00072     prob.x=features;
00073     prob.y=new int[prob.l];
00074     prob.use_bias=use_bias;
00075 
00076     for (int32_t i=0; i<prob.l; i++)
00077         prob.y[i]=labels->get_int_label(i);
00078 
00079     SG_INFO( "%d training points %d dims\n", prob.l, prob.n);
00080 
00081     function *fun_obj=NULL;
00082 
00083     switch (loss)
00084     {
00085         case LR:
00086             fun_obj=new l2_lr_fun(&prob, get_C1(), get_C2());
00087             break;
00088         case L2:
00089             fun_obj=new l2loss_svm_fun(&prob, get_C1(), get_C2());
00090             break;
00091         default:
00092             SG_ERROR("unknown loss\n");
00093             break;
00094     }
00095 
00096     if (fun_obj)
00097     {
00098         CTron tron_obj(fun_obj, epsilon);
00099         tron_obj.tron(w);
00100         float64_t sgn=prob.y[0];
00101 
00102         for (int32_t i=0; i<w_dim; i++)
00103             w[i]*=sgn;
00104 
00105         if (use_bias)
00106             set_bias(sgn*w[w_dim]);
00107         else
00108             set_bias(0);
00109 
00110         delete fun_obj;
00111     }
00112 
00113     delete[] prob.y;
00114 
00115     return true;
00116 }
00117 #endif //HAVE_LAPACK

SHOGUN Machine Learning Toolbox - Documentation