decision_tree.cc

   1 /*
   2  *  folded-ctf is an implementation of the folded hierarchy of
   3  *  classifiers for object detection, developed by Francois Fleuret
   4  *  and Donald Geman.
   5  *
   6  *  Copyright (c) 2008 Idiap Research Institute, http://www.idiap.ch/
   7  *  Written by Francois Fleuret <francois.fleuret@idiap.ch>
   8  *
   9  *  This file is part of folded-ctf.
  10  *
  11  *  folded-ctf is free software: you can redistribute it and/or modify
  12  *  it under the terms of the GNU General Public License version 3 as
  13  *  published by the Free Software Foundation.
  14  *
  15  *  folded-ctf is distributed in the hope that it will be useful, but
  16  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  17  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  *  General Public License for more details.
  19  *
  20  *  You should have received a copy of the GNU General Public License
  21  *  along with folded-ctf.  If not, see <http://www.gnu.org/licenses/>.
  22  *
  23  */
  24
  25 #include "decision_tree.h"
  26 #include "fusion_sort.h"
  27
  28 DecisionTree::DecisionTree() {
  29   _feature_index = -1;
  30   _threshold = 0;
  31   _weight = 0;
  32   _subtree_greater = 0;
  33   _subtree_lesser = 0;
  34 }
  35
  36 DecisionTree::~DecisionTree() {
  37   if(_subtree_lesser)
  38     delete _subtree_lesser;
  39   if(_subtree_greater)
  40     delete _subtree_greater;
  41 }
  42
  43 int DecisionTree::nb_leaves() {
  44   if(_subtree_lesser ||_subtree_greater)
  45     return _subtree_lesser->nb_leaves() + _subtree_greater->nb_leaves();
  46   else
  47     return 1;
  48 }
  49
  50 int DecisionTree::depth() {
  51   if(_subtree_lesser ||_subtree_greater)
  52     return 1 + max(_subtree_lesser->depth(), _subtree_greater->depth());
  53   else
  54     return 1;
  55 }
  56
  57 scalar_t DecisionTree::response(SampleSet *sample_set, int n_sample) {
  58   if(_subtree_lesser && _subtree_greater) {
  59     if(sample_set->feature_value(n_sample, _feature_index) < _threshold)
  60       return _subtree_lesser->response(sample_set, n_sample);
  61     else
  62       return _subtree_greater->response(sample_set, n_sample);
  63   } else {
  64     return _weight;
  65   }
  66 }
  67
  68 void DecisionTree::pick_best_split(SampleSet *sample_set, scalar_t *loss_derivatives) {
  69
  70   int nb_samples = sample_set->nb_samples();
  71
  72   scalar_t *responses = new scalar_t[nb_samples];
  73   int *indexes = new int[nb_samples];
  74   int *sorted_indexes = new int[nb_samples];
  75
  76   scalar_t max_abs_sum = 0;
  77   _feature_index = -1;
  78
  79   for(int f = 0; f < sample_set->nb_features(); f++) {
  80     scalar_t sum = 0;
  81
  82     for(int s = 0; s < nb_samples; s++) {
  83       indexes[s] = s;
  84       responses[s] = sample_set->feature_value(s, f);
  85       sum += loss_derivatives[s];
  86     }
  87
  88     indexed_fusion_sort(nb_samples, indexes, sorted_indexes, responses);
  89
  90     int t, u = sorted_indexes[0];
  91     for(int s = 0; s < nb_samples - 1; s++) {
  92       t = u;
  93       u = sorted_indexes[s + 1];
  94       sum -= 2 * loss_derivatives[t];
  95
  96       if(responses[t] < responses[u] && abs(sum) > max_abs_sum) {
  97         max_abs_sum = abs(sum);
  98         _feature_index = f;
  99         _threshold = (responses[t] + responses[u])/2;
 100       }
 101     }
 102   }
 103
 104   delete[] indexes;
 105   delete[] sorted_indexes;
 106   delete[] responses;
 107 }
 108
 109 void DecisionTree::train(LossMachine *loss_machine,
 110                          SampleSet *sample_set,
 111                          scalar_t *current_responses,
 112                          scalar_t *loss_derivatives,
 113                          int depth) {
 114
 115   if(_subtree_lesser || _subtree_greater || _feature_index >= 0) {
 116     cerr << "You can not re-train a tree." << endl;
 117     abort();
 118   }
 119
 120   int nb_samples = sample_set->nb_samples();
 121
 122   int nb_pos = 0, nb_neg = 0;
 123   for(int s = 0; s < sample_set->nb_samples(); s++) {
 124     if(sample_set->label(s) > 0) nb_pos++;
 125     else if(sample_set->label(s) < 0) nb_neg++;
 126   }
 127
 128   (*global.log_stream) << "Training tree" << endl;
 129   (*global.log_stream) << "  nb_samples " << nb_samples << endl;
 130   (*global.log_stream) << "  depth " << depth << endl;
 131   (*global.log_stream) << "  nb_pos = " << nb_pos << endl;
 132   (*global.log_stream) << "  nb_neg = " << nb_neg << endl;
 133
 134   if(depth >= global.tree_depth_max)
 135     (*global.log_stream) << "  Maximum depth reached." << endl;
 136   if(nb_pos < min_nb_samples_for_split)
 137     (*global.log_stream) << "  Not enough positive samples." << endl;
 138   if(nb_neg < min_nb_samples_for_split)
 139     (*global.log_stream) << "  Not enough negative samples." << endl;
 140
 141   if(depth < global.tree_depth_max &&
 142      nb_pos >= min_nb_samples_for_split &&
 143      nb_neg >= min_nb_samples_for_split) {
 144
 145     pick_best_split(sample_set, loss_derivatives);
 146
 147     if(_feature_index >= 0) {
 148       int indexes[nb_samples];
 149       scalar_t *parted_current_responses = new scalar_t[nb_samples];
 150       scalar_t *parted_loss_derivatives = new scalar_t[nb_samples];
 151
 152       int nb_lesser = 0, nb_greater = 0;
 153       int nb_lesser_pos = 0, nb_lesser_neg = 0, nb_greater_pos = 0, nb_greater_neg = 0;
 154
 155       for(int s = 0; s < nb_samples; s++) {
 156         if(sample_set->feature_value(s, _feature_index) < _threshold) {
 157           indexes[nb_lesser] = s;
 158           parted_current_responses[nb_lesser] = current_responses[s];
 159           parted_loss_derivatives[nb_lesser] = loss_derivatives[s];
 160
 161           if(sample_set->label(s) > 0)
 162             nb_lesser_pos++;
 163           else if(sample_set->label(s) < 0)
 164             nb_lesser_neg++;
 165
 166           nb_lesser++;
 167         } else {
 168           nb_greater++;
 169
 170           indexes[nb_samples - nb_greater] = s;
 171           parted_current_responses[nb_samples - nb_greater] = current_responses[s];
 172           parted_loss_derivatives[nb_samples - nb_greater] = loss_derivatives[s];
 173
 174           if(sample_set->label(s) > 0)
 175             nb_greater_pos++;
 176           else if(sample_set->label(s) < 0)
 177             nb_greater_neg++;
 178         }
 179       }
 180
 181       if((nb_lesser_pos >= min_nb_samples_for_split ||
 182           nb_lesser_neg >= min_nb_samples_for_split) &&
 183          (nb_greater_pos >= min_nb_samples_for_split ||
 184           nb_greater_neg >= min_nb_samples_for_split)) {
 185
 186         _subtree_lesser = new DecisionTree();
 187
 188         {
 189           SampleSet sub_sample_set(sample_set, nb_lesser, indexes);
 190
 191           _subtree_lesser->train(loss_machine,
 192                                  &sub_sample_set,
 193                                  parted_current_responses,
 194                                  parted_loss_derivatives,
 195                                  depth + 1);
 196         }
 197
 198         _subtree_greater = new DecisionTree();
 199
 200         {
 201           SampleSet sub_sample_set(sample_set, nb_greater, indexes + nb_lesser);
 202
 203           _subtree_greater->train(loss_machine,
 204                                   &sub_sample_set,
 205                                   parted_current_responses + nb_lesser,
 206                                   parted_loss_derivatives + nb_lesser,
 207                                   depth + 1);
 208         }
 209       }
 210
 211       delete[] parted_current_responses;
 212       delete[] parted_loss_derivatives;
 213     } else {
 214       (*global.log_stream) << "Could not find a feature for split." << endl;
 215     }
 216   }
 217
 218   if(!(_subtree_greater && _subtree_lesser)) {
 219     scalar_t *tmp_responses = new scalar_t[nb_samples];
 220     for(int s = 0; s < nb_samples; s++)
 221       tmp_responses[s] = 1;
 222
 223     _weight = loss_machine->optimal_weight(sample_set, tmp_responses, current_responses);
 224
 225     const scalar_t max_weight = 10.0;
 226
 227     if(_weight > max_weight) {
 228       _weight = max_weight;
 229     } else if(_weight < - max_weight) {
 230       _weight = - max_weight;
 231     }
 232
 233     (*global.log_stream) << "  _weight " << _weight << endl;
 234
 235     delete[] tmp_responses;
 236   }
 237 }
 238
 239 void DecisionTree::train(LossMachine *loss_machine,
 240                  SampleSet *sample_set,
 241                  scalar_t *current_responses) {
 242
 243   scalar_t *loss_derivatives = new scalar_t[sample_set->nb_samples()];
 244
 245   loss_machine->get_loss_derivatives(sample_set, current_responses, loss_derivatives);
 246
 247   train(loss_machine, sample_set, current_responses, loss_derivatives, 0);
 248
 249   delete[] loss_derivatives;
 250 }
 251
 252 //////////////////////////////////////////////////////////////////////
 253
 254 void DecisionTree::tag_used_features(bool *used) {
 255   if(_subtree_lesser && _subtree_greater) {
 256     used[_feature_index] = true;
 257     _subtree_lesser->tag_used_features(used);
 258     _subtree_greater->tag_used_features(used);
 259   }
 260 }
 261
 262 void DecisionTree::re_index_features(int *new_indexes) {
 263   if(_subtree_lesser && _subtree_greater) {
 264     _feature_index = new_indexes[_feature_index];
 265     _subtree_lesser->re_index_features(new_indexes);
 266     _subtree_greater->re_index_features(new_indexes);
 267   }
 268 }
 269
 270 //////////////////////////////////////////////////////////////////////
 271
 272 void DecisionTree::read(istream *is) {
 273   if(_subtree_lesser || _subtree_greater) {
 274     cerr << "You can not read in an existing tree." << endl;
 275     abort();
 276   }
 277
 278   read_var(is, &_feature_index);
 279   read_var(is, &_threshold);
 280   read_var(is, &_weight);
 281
 282   int split;
 283   read_var(is, &split);
 284
 285   if(split) {
 286     _subtree_lesser = new DecisionTree();
 287     _subtree_lesser->read(is);
 288     _subtree_greater = new DecisionTree();
 289     _subtree_greater->read(is);
 290   }
 291 }
 292
 293 void DecisionTree::write(ostream *os) {
 294
 295   write_var(os, &_feature_index);
 296   write_var(os, &_threshold);
 297   write_var(os, &_weight);
 298
 299   int split;
 300   if(_subtree_lesser && _subtree_greater) {
 301     split = 1;
 302     write_var(os, &split);
 303     _subtree_lesser->write(os);
 304     _subtree_greater->write(os);
 305   } else {
 306     split = 0;
 307     write_var(os, &split);
 308   }
 309 }