00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011 #include "lib/common.h"
00012
00013 #ifdef HAVE_LAPACK
00014 #include "classifier/Classifier.h"
00015 #include "classifier/LinearClassifier.h"
00016 #include "classifier/LDA.h"
00017 #include "features/Labels.h"
00018 #include "lib/Mathematics.h"
00019 #include "lib/lapack.h"
00020
00021 CLDA::CLDA(DREAL gamma)
00022 : CLinearClassifier(), m_gamma(gamma)
00023 {
00024 }
00025
00026 CLDA::CLDA(DREAL gamma, CRealFeatures* traindat, CLabels* trainlab)
00027 : CLinearClassifier(), m_gamma(gamma)
00028 {
00029 set_features(traindat);
00030 set_labels(trainlab);
00031 }
00032
00033
00034 CLDA::~CLDA()
00035 {
00036 }
00037
00038 bool CLDA::train()
00039 {
00040 ASSERT(labels);
00041 ASSERT(features);
00042 INT num_train_labels=0;
00043 INT* train_labels=labels->get_int_labels(num_train_labels);
00044 ASSERT(train_labels);
00045
00046 INT num_feat=features->get_num_features();
00047 INT num_vec=features->get_num_vectors();
00048 ASSERT(num_vec==num_train_labels);
00049
00050 INT* classidx_neg=new INT[num_vec];
00051 INT* classidx_pos=new INT[num_vec];
00052
00053 INT i=0;
00054 INT j=0;
00055 INT num_neg=0;
00056 INT num_pos=0;
00057 for (i=0; i<num_train_labels; i++)
00058 {
00059 if (train_labels[i]==-1)
00060 classidx_neg[num_neg++]=i;
00061 else if (train_labels[i]==+1)
00062 classidx_pos[num_pos++]=i;
00063 else
00064 {
00065 SG_ERROR( "found label != +/- 1 bailing...");
00066 return false;
00067 }
00068 }
00069
00070 if (num_neg<=0 && num_pos<=0)
00071 {
00072 SG_ERROR( "whooooo ? only a single class found\n");
00073 return false;
00074 }
00075
00076 delete[] w;
00077 w=new DREAL[num_feat];
00078 w_dim=num_feat;
00079
00080 DREAL* mean_neg=new DREAL[num_feat];
00081 memset(mean_neg,0,num_feat*sizeof(DREAL));
00082
00083 DREAL* mean_pos=new DREAL[num_feat];
00084 memset(mean_pos,0,num_feat*sizeof(DREAL));
00085
00086 DREAL* scatter=new DREAL[num_feat*num_feat];
00087 DREAL* buffer=new DREAL[num_feat*CMath::max(num_neg, num_pos)];
00088
00089
00090 for (i=0; i<num_neg; i++)
00091 {
00092 INT vlen;
00093 bool vfree;
00094 double* vec=features->get_feature_vector(classidx_neg[i], vlen, vfree);
00095 ASSERT(vec);
00096
00097 for (j=0; j<vlen; j++)
00098 {
00099 mean_neg[j]+=vec[j];
00100 buffer[num_feat*i+j]=vec[j];
00101 }
00102
00103 features->free_feature_vector(vec, classidx_neg[i], vfree);
00104 }
00105
00106 for (j=0; j<num_feat; j++)
00107 mean_neg[j]/=num_neg;
00108
00109 for (i=0; i<num_neg; i++)
00110 {
00111 for (j=0; j<num_feat; j++)
00112 buffer[num_feat*i+j]-=mean_neg[j];
00113 }
00114 cblas_dgemm(CblasColMajor, CblasNoTrans, CblasTrans, num_feat, num_feat, num_neg, 1.0, buffer, num_feat, buffer, num_feat, 0, scatter, num_feat);
00115
00116
00117 for (i=0; i<num_pos; i++)
00118 {
00119 INT vlen;
00120 bool vfree;
00121 double* vec=features->get_feature_vector(classidx_pos[i], vlen, vfree);
00122 ASSERT(vec);
00123
00124 for (j=0; j<vlen; j++)
00125 {
00126 mean_pos[j]+=vec[j];
00127 buffer[num_feat*i+j]=vec[j];
00128 }
00129
00130 features->free_feature_vector(vec, classidx_pos[i], vfree);
00131 }
00132
00133 for (j=0; j<num_feat; j++)
00134 mean_pos[j]/=num_pos;
00135
00136 for (i=0; i<num_pos; i++)
00137 {
00138 for (j=0; j<num_feat; j++)
00139 buffer[num_feat*i+j]-=mean_pos[j];
00140 }
00141 cblas_dgemm(CblasColMajor, CblasNoTrans, CblasTrans, num_feat, num_feat, num_pos, 1.0/(num_train_labels-1), buffer, num_feat, buffer, num_feat, 1.0/(num_train_labels-1), scatter, num_feat);
00142
00143 DREAL trace=CMath::trace(scatter, num_feat, num_feat);
00144
00145 double s=1.0-m_gamma;
00146
00147 for (i=0; i<num_feat*num_feat; i++)
00148 scatter[i]*=s;
00149
00150 for (i=0; i<num_feat; i++)
00151 scatter[i*num_feat+i]+= trace*m_gamma/num_feat;
00152
00153 DREAL* inv_scatter= CMath::pinv(scatter, num_feat, num_feat, NULL);
00154
00155 DREAL* w_pos=buffer;
00156 DREAL* w_neg=&buffer[num_feat];
00157
00158 cblas_dsymv(CblasColMajor, CblasUpper, num_feat, 1.0, inv_scatter, num_feat, mean_pos, 1, 0, w_pos, 1);
00159 cblas_dsymv(CblasColMajor, CblasUpper, num_feat, 1.0, inv_scatter, num_feat, mean_neg, 1, 0, w_neg, 1);
00160
00161 bias=0.5*(CMath::dot(w_neg, mean_neg, num_feat)-CMath::dot(w_pos, mean_pos, num_feat));
00162 for (i=0; i<num_feat; i++)
00163 w[i]=w_pos[i]-w_neg[i];
00164
00165 #ifdef DEBUG_LDA
00166 SG_PRINT("bias: %f\n", bias);
00167 CMath::display_vector(w, num_feat, "w");
00168 CMath::display_vector(w_pos, num_feat, "w_pos");
00169 CMath::display_vector(w_neg, num_feat, "w_neg");
00170 CMath::display_vector(mean_pos, num_feat, "mean_pos");
00171 CMath::display_vector(mean_neg, num_feat, "mean_neg");
00172 #endif
00173
00174 delete[] train_labels;
00175 delete[] mean_neg;
00176 delete[] mean_pos;
00177 delete[] scatter;
00178 delete[] inv_scatter;
00179 delete[] classidx_neg;
00180 delete[] classidx_pos;
00181 delete[] buffer;
00182 return true;
00183 }
00184 #endif