Go to the documentation of this file.00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011 #include <shogun/classifier/svm/GPBTSVM.h>
00012 #include <shogun/classifier/svm/gpdt.h>
00013 #include <shogun/classifier/svm/gpdtsolve.h>
00014 #include <shogun/io/SGIO.h>
00015
00016 using namespace shogun;
00017
00018 CGPBTSVM::CGPBTSVM()
00019 : CSVM(), model(NULL)
00020 {
00021 }
00022
00023 CGPBTSVM::CGPBTSVM(float64_t C, CKernel* k, CLabels* lab)
00024 : CSVM(C, k, lab), model(NULL)
00025 {
00026 }
00027
00028 CGPBTSVM::~CGPBTSVM()
00029 {
00030 SG_FREE(model);
00031 }
00032
00033 bool CGPBTSVM::train_machine(CFeatures* data)
00034 {
00035 float64_t* solution;
00036 QPproblem prob;
00037
00038 ASSERT(kernel);
00039 ASSERT(labels && labels->get_num_labels());
00040 ASSERT(labels->is_two_class_labeling());
00041 if (data)
00042 {
00043 if (labels->get_num_labels() != data->get_num_vectors())
00044 SG_ERROR("Number of training vectors does not match number of labels\n");
00045 kernel->init(data, data);
00046 }
00047
00048 SGVector<int32_t> lab=labels->get_int_labels();
00049 prob.KER=new sKernel(kernel, lab.vlen);
00050 prob.y=lab.vector;
00051 prob.ell=lab.vlen;
00052 SG_INFO( "%d trainlabels\n", prob.ell);
00053
00054
00055 prob.delta = epsilon;
00056 prob.maxmw = kernel->get_cache_size();
00057 prob.verbosity = 0;
00058 prob.preprocess_size = -1;
00059 prob.projection_projector = -1;
00060 prob.c_const = get_C1();
00061 prob.chunk_size = get_qpsize();
00062 prob.linadd = get_linadd_enabled();
00063
00064 if (prob.chunk_size < 2) prob.chunk_size = 2;
00065 if (prob.q <= 0) prob.q = prob.chunk_size / 3;
00066 if (prob.q < 2) prob.q = 2;
00067 if (prob.q > prob.chunk_size) prob.q = prob.chunk_size;
00068 prob.q = prob.q & (~1);
00069 if (prob.maxmw < 5)
00070 prob.maxmw = 5;
00071
00072
00073 SG_INFO( "\nTRAINING PARAMETERS:\n");
00074 SG_INFO( "\tNumber of training documents: %d\n", prob.ell);
00075 SG_INFO( "\tq: %d\n", prob.chunk_size);
00076 SG_INFO( "\tn: %d\n", prob.q);
00077 SG_INFO( "\tC: %lf\n", prob.c_const);
00078 SG_INFO( "\tkernel type: %d\n", prob.ker_type);
00079 SG_INFO( "\tcache size: %dMb\n", prob.maxmw);
00080 SG_INFO( "\tStopping tolerance: %lf\n", prob.delta);
00081
00082
00083 if (prob.preprocess_size == -1)
00084 prob.preprocess_size = (int32_t) ( (float64_t)prob.chunk_size * 1.5 );
00085
00086 if (prob.projection_projector == -1)
00087 {
00088 if (prob.chunk_size <= 20) prob.projection_projector = 0;
00089 else prob.projection_projector = 1;
00090 }
00091
00092
00093 solution = SG_MALLOC(float64_t, prob.ell);
00094 prob.gpdtsolve(solution);
00095
00096
00097 CSVM::set_objective(prob.objective_value);
00098
00099 int32_t num_sv=0;
00100 int32_t bsv=0;
00101 int32_t i=0;
00102 int32_t k=0;
00103
00104 for (i = 0; i < prob.ell; i++)
00105 {
00106 if (solution[i] > prob.DELTAsv)
00107 {
00108 num_sv++;
00109 if (solution[i] > (prob.c_const - prob.DELTAsv)) bsv++;
00110 }
00111 }
00112
00113 create_new_model(num_sv);
00114 set_bias(prob.bee);
00115
00116 SG_INFO("SV: %d BSV = %d\n", num_sv, bsv);
00117
00118 for (i = 0; i < prob.ell; i++)
00119 {
00120 if (solution[i] > prob.DELTAsv)
00121 {
00122 set_support_vector(k, i);
00123 set_alpha(k++, solution[i]*labels->get_label(i));
00124 }
00125 }
00126
00127 delete prob.KER;
00128 lab.free_vector();
00129 SG_FREE(solution);
00130
00131 return true;
00132 }