00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012 #include <shogun/features/TOPFeatures.h>
00013 #include <shogun/io/SGIO.h>
00014 #include <shogun/mathematics/Math.h>
00015
00016 using namespace shogun;
00017
00018 CTOPFeatures::CTOPFeatures()
00019 {
00020 init();
00021 }
00022
00023 CTOPFeatures::CTOPFeatures(
00024 int32_t size, CHMM* p, CHMM* n, bool neglin, bool poslin)
00025 : CDenseFeatures<float64_t>(size)
00026 {
00027 init();
00028 neglinear=neglin;
00029 poslinear=poslin;
00030
00031 set_models(p,n);
00032 }
00033
00034 CTOPFeatures::CTOPFeatures(const CTOPFeatures &orig)
00035 : CDenseFeatures<float64_t>(orig)
00036 {
00037 init();
00038 pos=orig.pos;
00039 neg=orig.neg;
00040 neglinear=orig.neglinear;
00041 poslinear=orig.poslinear;
00042 }
00043
00044 CTOPFeatures::~CTOPFeatures()
00045 {
00046 SG_FREE(pos_relevant_indizes.idx_p);
00047 SG_FREE(pos_relevant_indizes.idx_q);
00048 SG_FREE(pos_relevant_indizes.idx_a_cols);
00049 SG_FREE(pos_relevant_indizes.idx_a_rows);
00050 SG_FREE(pos_relevant_indizes.idx_b_cols);
00051 SG_FREE(pos_relevant_indizes.idx_b_rows);
00052
00053 SG_FREE(neg_relevant_indizes.idx_p);
00054 SG_FREE(neg_relevant_indizes.idx_q);
00055 SG_FREE(neg_relevant_indizes.idx_a_cols);
00056 SG_FREE(neg_relevant_indizes.idx_a_rows);
00057 SG_FREE(neg_relevant_indizes.idx_b_cols);
00058 SG_FREE(neg_relevant_indizes.idx_b_rows);
00059
00060 SG_UNREF(pos);
00061 SG_UNREF(neg);
00062 }
00063
00064 void CTOPFeatures::set_models(CHMM* p, CHMM* n)
00065 {
00066 ASSERT(p && n);
00067 SG_REF(p);
00068 SG_REF(n);
00069
00070 pos=p;
00071 neg=n;
00072 set_num_vectors(0);
00073
00074 feature_matrix=SGMatrix<float64_t>();
00075
00076 if (pos && pos->get_observations())
00077 set_num_vectors(pos->get_observations()->get_num_vectors());
00078
00079 compute_relevant_indizes(p, &pos_relevant_indizes);
00080 compute_relevant_indizes(n, &neg_relevant_indizes);
00081 num_features=compute_num_features();
00082
00083 SG_DEBUG( "pos_feat=[%i,%i,%i,%i],neg_feat=[%i,%i,%i,%i] -> %i features\n", pos->get_N(), pos->get_N(), pos->get_N()*pos->get_N(), pos->get_N()*pos->get_M(), neg->get_N(), neg->get_N(), neg->get_N()*neg->get_N(), neg->get_N()*neg->get_M(),num_features) ;
00084 }
00085
00086 float64_t* CTOPFeatures::compute_feature_vector(
00087 int32_t num, int32_t &len, float64_t* target)
00088 {
00089 float64_t* featurevector=target;
00090
00091 if (!featurevector)
00092 featurevector=SG_MALLOC(float64_t, get_num_features());
00093
00094 if (!featurevector)
00095 return NULL;
00096
00097 compute_feature_vector(featurevector, num, len);
00098
00099 return featurevector;
00100 }
00101
00102 void CTOPFeatures::compute_feature_vector(
00103 float64_t* featurevector, int32_t num, int32_t& len)
00104 {
00105 int32_t i,j,p=0,x=num;
00106 int32_t idx=0;
00107
00108 float64_t posx=(poslinear) ?
00109 (pos->linear_model_probability(x)) : (pos->model_probability(x));
00110 float64_t negx=(neglinear) ?
00111 (neg->linear_model_probability(x)) : (neg->model_probability(x));
00112
00113 len=get_num_features();
00114
00115 featurevector[p++]=(posx-negx);
00116
00117
00118 if (poslinear)
00119 {
00120 for (i=0; i<pos->get_N(); i++)
00121 {
00122 for (j=0; j<pos->get_M(); j++)
00123 featurevector[p++]=exp(pos->linear_model_derivative(i, j, x)-posx);
00124 }
00125 }
00126 else
00127 {
00128 for (idx=0; idx< pos_relevant_indizes.num_p; idx++)
00129 featurevector[p++]=exp(pos->model_derivative_p(pos_relevant_indizes.idx_p[idx], x)-posx);
00130
00131 for (idx=0; idx< pos_relevant_indizes.num_q; idx++)
00132 featurevector[p++]=exp(pos->model_derivative_q(pos_relevant_indizes.idx_q[idx], x)-posx);
00133
00134 for (idx=0; idx< pos_relevant_indizes.num_a; idx++)
00135 featurevector[p++]=exp(pos->model_derivative_a(pos_relevant_indizes.idx_a_rows[idx], pos_relevant_indizes.idx_a_cols[idx], x)-posx);
00136
00137 for (idx=0; idx< pos_relevant_indizes.num_b; idx++)
00138 featurevector[p++]=exp(pos->model_derivative_b(pos_relevant_indizes.idx_b_rows[idx], pos_relevant_indizes.idx_b_cols[idx], x)-posx);
00139
00140
00141
00142
00143
00144
00145
00146
00147
00148
00149
00150
00151
00152 }
00153
00154
00155 if (neglinear)
00156 {
00157 for (i=0; i<neg->get_N(); i++)
00158 {
00159 for (j=0; j<neg->get_M(); j++)
00160 featurevector[p++]= - exp(neg->linear_model_derivative(i, j, x)-negx);
00161 }
00162 }
00163 else
00164 {
00165 for (idx=0; idx< neg_relevant_indizes.num_p; idx++)
00166 featurevector[p++]= - exp(neg->model_derivative_p(neg_relevant_indizes.idx_p[idx], x)-negx);
00167
00168 for (idx=0; idx< neg_relevant_indizes.num_q; idx++)
00169 featurevector[p++]= - exp(neg->model_derivative_q(neg_relevant_indizes.idx_q[idx], x)-negx);
00170
00171 for (idx=0; idx< neg_relevant_indizes.num_a; idx++)
00172 featurevector[p++]= - exp(neg->model_derivative_a(neg_relevant_indizes.idx_a_rows[idx], neg_relevant_indizes.idx_a_cols[idx], x)-negx);
00173
00174 for (idx=0; idx< neg_relevant_indizes.num_b; idx++)
00175 featurevector[p++]= - exp(neg->model_derivative_b(neg_relevant_indizes.idx_b_rows[idx], neg_relevant_indizes.idx_b_cols[idx], x)-negx);
00176
00177
00178
00179
00180
00181
00182
00183
00184
00185
00186
00187
00188 }
00189 }
00190
00191 float64_t* CTOPFeatures::set_feature_matrix()
00192 {
00193 int32_t len=0;
00194
00195 num_features=get_num_features();
00196 ASSERT(num_features);
00197 ASSERT(pos);
00198 ASSERT(pos->get_observations());
00199
00200 num_vectors=pos->get_observations()->get_num_vectors();
00201 SG_INFO( "allocating top feature cache of size %.2fM\n", sizeof(float64_t)*num_features*num_vectors/1024.0/1024.0);
00202 feature_matrix = SGMatrix<float64_t>(num_features,num_vectors);
00203 if (!feature_matrix.matrix)
00204 {
00205 SG_ERROR( "allocation not successful!");
00206 return NULL ;
00207 } ;
00208
00209 SG_INFO( "calculating top feature matrix\n");
00210
00211 for (int32_t x=0; x<num_vectors; x++)
00212 {
00213 if (!(x % (num_vectors/10+1)))
00214 SG_DEBUG( "%02d%%.", (int) (100.0*x/num_vectors));
00215 else if (!(x % (num_vectors/200+1)))
00216 SG_DEBUG( ".");
00217
00218 compute_feature_vector(&feature_matrix[x*num_features], x, len);
00219 }
00220
00221 SG_DONE();
00222
00223 num_vectors=get_num_vectors() ;
00224 num_features=get_num_features() ;
00225
00226 return feature_matrix.matrix;
00227 }
00228
00229 bool CTOPFeatures::compute_relevant_indizes(CHMM* hmm, T_HMM_INDIZES* hmm_idx)
00230 {
00231 int32_t i=0;
00232 int32_t j=0;
00233
00234 hmm_idx->num_p=0;
00235 hmm_idx->num_q=0;
00236 hmm_idx->num_a=0;
00237 hmm_idx->num_b=0;
00238
00239 for (i=0; i<hmm->get_N(); i++)
00240 {
00241 if (hmm->get_p(i)>CMath::ALMOST_NEG_INFTY)
00242 hmm_idx->num_p++;
00243
00244 if (hmm->get_q(i)>CMath::ALMOST_NEG_INFTY)
00245 hmm_idx->num_q++;
00246
00247 for (j=0; j<hmm->get_N(); j++)
00248 {
00249 if (hmm->get_a(i,j)>CMath::ALMOST_NEG_INFTY)
00250 hmm_idx->num_a++;
00251 }
00252
00253 for (j=0; j<pos->get_M(); j++)
00254 {
00255 if (hmm->get_b(i,j)>CMath::ALMOST_NEG_INFTY)
00256 hmm_idx->num_b++;
00257 }
00258 }
00259
00260 if (hmm_idx->num_p > 0)
00261 {
00262 hmm_idx->idx_p=SG_MALLOC(int32_t, hmm_idx->num_p);
00263 ASSERT(hmm_idx->idx_p);
00264 }
00265
00266 if (hmm_idx->num_q > 0)
00267 {
00268 hmm_idx->idx_q=SG_MALLOC(int32_t, hmm_idx->num_q);
00269 ASSERT(hmm_idx->idx_q);
00270 }
00271
00272 if (hmm_idx->num_a > 0)
00273 {
00274 hmm_idx->idx_a_rows=SG_MALLOC(int32_t, hmm_idx->num_a);
00275 hmm_idx->idx_a_cols=SG_MALLOC(int32_t, hmm_idx->num_a);
00276 ASSERT(hmm_idx->idx_a_rows);
00277 ASSERT(hmm_idx->idx_a_cols);
00278 }
00279
00280 if (hmm_idx->num_b > 0)
00281 {
00282 hmm_idx->idx_b_rows=SG_MALLOC(int32_t, hmm_idx->num_b);
00283 hmm_idx->idx_b_cols=SG_MALLOC(int32_t, hmm_idx->num_b);
00284 ASSERT(hmm_idx->idx_b_rows);
00285 ASSERT(hmm_idx->idx_b_cols);
00286 }
00287
00288
00289 int32_t idx_p=0;
00290 int32_t idx_q=0;
00291 int32_t idx_a=0;
00292 int32_t idx_b=0;
00293
00294 for (i=0; i<hmm->get_N(); i++)
00295 {
00296 if (hmm->get_p(i)>CMath::ALMOST_NEG_INFTY)
00297 {
00298 ASSERT(idx_p < hmm_idx->num_p);
00299 hmm_idx->idx_p[idx_p++]=i;
00300 }
00301
00302 if (hmm->get_q(i)>CMath::ALMOST_NEG_INFTY)
00303 {
00304 ASSERT(idx_q < hmm_idx->num_q);
00305 hmm_idx->idx_q[idx_q++]=i;
00306 }
00307
00308 for (j=0; j<hmm->get_N(); j++)
00309 {
00310 if (hmm->get_a(i,j)>CMath::ALMOST_NEG_INFTY)
00311 {
00312 ASSERT(idx_a < hmm_idx->num_a);
00313 hmm_idx->idx_a_rows[idx_a]=i;
00314 hmm_idx->idx_a_cols[idx_a++]=j;
00315 }
00316 }
00317
00318 for (j=0; j<pos->get_M(); j++)
00319 {
00320 if (hmm->get_b(i,j)>CMath::ALMOST_NEG_INFTY)
00321 {
00322 ASSERT(idx_b < hmm_idx->num_b);
00323 hmm_idx->idx_b_rows[idx_b]=i;
00324 hmm_idx->idx_b_cols[idx_b++]=j;
00325 }
00326 }
00327 }
00328
00329 return true;
00330 }
00331
00332 int32_t CTOPFeatures::compute_num_features()
00333 {
00334 int32_t num=0;
00335
00336 if (pos && neg)
00337 {
00338 num+=1;
00339
00340 if (poslinear)
00341 num+=pos->get_N()*pos->get_M();
00342 else
00343 {
00344 num+= pos_relevant_indizes.num_p + pos_relevant_indizes.num_q + pos_relevant_indizes.num_a + pos_relevant_indizes.num_b;
00345 }
00346
00347 if (neglinear)
00348 num+=neg->get_N()*neg->get_M();
00349 else
00350 {
00351 num+= neg_relevant_indizes.num_p + neg_relevant_indizes.num_q + neg_relevant_indizes.num_a + neg_relevant_indizes.num_b;
00352 }
00353
00354
00355
00356
00357 }
00358 return num;
00359 }
00360
00361 void CTOPFeatures::init()
00362 {
00363 pos = NULL;
00364 neg = NULL;
00365 neglinear = false;
00366 poslinear = false;
00367
00368 memset(&pos_relevant_indizes, 0, sizeof(pos_relevant_indizes));
00369 memset(&neg_relevant_indizes, 0, sizeof(neg_relevant_indizes));
00370
00371 unset_generic();
00372
00373
00374
00375 m_parameters->add(&neglinear, "neglinear", "If negative HMM is a LinearHMM");
00376 m_parameters->add(&poslinear, "poslinear", "If positive HMM is a LinearHMM");
00377 }