global.cc

   1
   2 ///////////////////////////////////////////////////////////////////////////
   3 // This program is free software: you can redistribute it and/or modify  //
   4 // it under the terms of the version 3 of the GNU General Public License //
   5 // as published by the Free Software Foundation.                         //
   6 //                                                                       //
   7 // This program is distributed in the hope that it will be useful, but   //
   8 // WITHOUT ANY WARRANTY; without even the implied warranty of            //
   9 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU      //
  10 // General Public License for more details.                              //
  11 //                                                                       //
  12 // You should have received a copy of the GNU General Public License     //
  13 // along with this program. If not, see <http://www.gnu.org/licenses/>.  //
  14 //                                                                       //
  15 // Written by Francois Fleuret, (C) IDIAP                                //
  16 // Contact <francois.fleuret@idiap.ch> for comments & bug reports        //
  17 ///////////////////////////////////////////////////////////////////////////
  18
  19 #include <string.h>
  20
  21 #include "global.h"
  22
  23 Global global;
  24
  25 Global::Global() {
  26   log_stream = 0;
  27 }
  28
  29 Global::~Global() {
  30   delete log_stream;
  31 }
  32
  33 void Global::init_parser(ParamParser *parser) {
  34   // The nice level of the process
  35   parser->add_association("niceness", "15", false);
  36
  37   // Seed to initialize the random generator
  38   parser->add_association("random-seed", "0", false);
  39
  40   // Should the pictures be b&w
  41   parser->add_association("pictures-for-article", "no", false);
  42
  43   // The name of the image pool to use
  44   parser->add_association("pool-name", "", false);
  45   // The name of the test image pool to use
  46   parser->add_association("test-pool-name", "", false);
  47   // From where to load or where to save the detector
  48   parser->add_association("detector-name", "default.det", false);
  49   // Where to put the generated files
  50   parser->add_association("result-path", "/tmp/", false);
  51
  52   // What kind of loss for the boosting
  53   parser->add_association("loss-type", "exponential", false);
  54
  55   // How many images to produce/process
  56   parser->add_association("nb-images", "-1", false);
  57   // What is the number of the feature to show in the images
  58   parser->add_association("material-feature-nb", "-1", false);
  59
  60   // What is the maximum tree depth
  61   parser->add_association("tree-depth-max", "1", false);
  62   // What is the proportion of negative cells we actually use during training
  63   parser->add_association("proportion-negative-cells-for-training", "0.025", false);
  64   // How many negative samples to sub-sample for boosting every classifier
  65   parser->add_association("nb-negative-samples-per-positive", "10", false);
  66   // How many features we will look at for boosting optimization
  67   parser->add_association("nb-features-for-boosting-optimization", "10000", false);
  68   // Do we allow head-belly registration
  69   parser->add_association("force-head-belly-independence", "no", false);
  70   // How many weak-learners in every classifier
  71   parser->add_association("nb-weak-learners-per-classifier", "100", false);
  72   // How many classifiers per level
  73   parser->add_association("nb-classifiers-per-level", "25", false);
  74   // How many levels
  75   parser->add_association("nb-levels", "2", false);
  76
  77   // Proportion of images from the pool to use for training
  78   parser->add_association("proportion-for-train", "0.5", false);
  79   // Proportion of images from the pool to use for validation
  80   parser->add_association("proportion-for-validation", "0.25", false);
  81   // Proportion of images from the pool to use for test (negative
  82   // means everything else)
  83   parser->add_association("proportion-for-test", "0.25", false);
  84   // During training, should we write the ROC curve estimated on the
  85   // validation set (which cost a bit of computation)
  86   parser->add_association("write-validation-rocs", "no", false);
  87
  88   // Should we write down the PNGs for the results of the parsing
  89   parser->add_association("write-parse-images", "no", false);
  90
  91   // Should we write down the PNGs for the tags
  92   parser->add_association("write-tag-images", "no", false);
  93
  94   // What is the wanted true overall positive rate
  95   parser->add_association("wanted-true-positive-rate", "0.75", false);
  96   // How many rates to try for the sequence of tests
  97   parser->add_association("nb-wanted-true-positive-rates", "10", false);
  98
  99   // What is the minimum radius of the heads to detect. This is used
 100   // as the reference size.
 101   parser->add_association("min-head-radius", "25", false);
 102   // What is the maximum size of the heads to detect.
 103   parser->add_association("max-head-radius", "200", false);
 104   // How many translation cell for one scale when generating the "top
 105   // level" cells for an image.
 106   parser->add_association("root-cell-nb-xy-per-radius", "5", false);
 107
 108   // What is the minimum size of the windows
 109   parser->add_association("pi-feature-window-min-size", "0.1", false);
 110
 111   // How many scales between two powers of two for the multi-scale
 112   // images
 113   parser->add_association("nb-scales-per-power-of-two", "5", false);
 114
 115   // Should we display a progress bar for lengthy operations
 116   parser->add_association("progress-bar", "yes", false);
 117 }
 118
 119 void Global::read_parser(ParamParser *parser) {
 120   niceness = parser->get_association_int("niceness");
 121   random_seed = parser->get_association_int("random-seed");
 122   pictures_for_article = parser->get_association_bool("pictures-for-article");
 123
 124   strncpy(pool_name, parser->get_association("pool-name"), buffer_size);
 125   strncpy(test_pool_name, parser->get_association("test-pool-name"), buffer_size);
 126   strncpy(detector_name, parser->get_association("detector-name"), buffer_size);
 127   strncpy(result_path, parser->get_association("result-path"), buffer_size);
 128
 129   char buffer[buffer_size];
 130   sprintf(buffer, "%s/log", result_path);
 131   log_stream = new ofstream(buffer);
 132
 133   char *l = parser->get_association("loss-type");
 134   if(strcmp(l, "exponential") == 0)
 135     loss_type = LOSS_EXPONENTIAL;
 136   else if(strcmp(l, "ev-regularized") == 0)
 137     loss_type = LOSS_EV_REGULARIZED;
 138   else if(strcmp(l, "hinge") == 0)
 139     loss_type = LOSS_HINGE;
 140   else if(strcmp(l, "logistic") == 0)
 141     loss_type = LOSS_LOGISTIC;
 142   else {
 143     cerr << "Unknown loss type." << endl;
 144     exit(1);
 145   }
 146
 147   nb_images = parser->get_association_int("nb-images");
 148   material_feature_nb = parser->get_association_int("material-feature-nb");
 149   tree_depth_max = parser->get_association_int("tree-depth-max");
 150   nb_weak_learners_per_classifier = parser->get_association_int("nb-weak-learners-per-classifier");
 151   nb_classifiers_per_level = parser->get_association_int("nb-classifiers-per-level");
 152   nb_levels = parser->get_association_int("nb-levels");
 153   proportion_negative_cells_for_training = parser->get_association_scalar("proportion-negative-cells-for-training");
 154   nb_negative_samples_per_positive = parser->get_association_int("nb-negative-samples-per-positive");
 155   nb_features_for_boosting_optimization = parser->get_association_int("nb-features-for-boosting-optimization");
 156   force_head_belly_independence = parser->get_association_bool("force-head-belly-independence");
 157   proportion_for_train = parser->get_association_scalar("proportion-for-train");
 158   proportion_for_validation = parser->get_association_scalar("proportion-for-validation");
 159   proportion_for_test = parser->get_association_scalar("proportion-for-test");
 160   write_validation_rocs = parser->get_association_bool("write-validation-rocs");
 161   write_parse_images = parser->get_association_bool("write-parse-images");
 162   write_tag_images = parser->get_association_bool("write-tag-images");
 163   wanted_true_positive_rate = parser->get_association_scalar("wanted-true-positive-rate");
 164   nb_wanted_true_positive_rates = parser->get_association_int("nb-wanted-true-positive-rates");
 165
 166   min_head_radius = parser->get_association_scalar("min-head-radius");
 167   max_head_radius = parser->get_association_scalar("max-head-radius");
 168   root_cell_nb_xy_per_radius = parser->get_association_int("root-cell-nb-xy-per-radius");
 169
 170   pi_feature_window_min_size = parser->get_association_scalar("pi-feature-window-min-size");
 171
 172   nb_scales_per_power_of_two = parser->get_association_int("nb-scales-per-power-of-two");
 173
 174   bar.set_visible(parser->get_association_bool("progress-bar"));
 175 }