songsenand
/
cgal
mirror of https://github.com/CGAL/cgal
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
cgal/Classification/include/CGAL/Classification/Trainer.h

692 lines
22 KiB
C++
Raw Blame History

// Copyright (c) 2017 GeometryFactory Sarl (France).
// All rights reserved.
//
// This file is part of CGAL (www.cgal.org).
// You can redistribute it and/or modify it under the terms of the GNU
// General Public License as published by the Free Software Foundation,
// either version 3 of the License, or (at your option) any later version.
//
// Licensees holding a valid commercial license may use this file in
// accordance with the commercial license agreement provided with the software.
//
// This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
// WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
//
// $URL$
// $Id$
//
// Author(s) : Simon Giraudot
#ifndef CGAL_CLASSIFICATION_TRAINER_H
#define CGAL_CLASSIFICATION_TRAINER_H
#include <CGAL/Classification/Feature_base.h>
#include <CGAL/Classification/Label.h>
#include <CGAL/Classifier.h>
//#define CGAL_CLASSTRAINING_VERBOSE
#if defined(CGAL_CLASSTRAINING_VERBOSE)
#define CGAL_CLASSTRAINING_CERR std::cerr
#else
#define CGAL_CLASSTRAINING_CERR std::ostream(0)
#endif
namespace CGAL {
namespace Classification {
/*!
\ingroup PkgClassification
\brief Training algorithm to set up weights and effects of the
features used for classification.
Each label must have ben given a small set of user-defined inliers to
provide the training algorithm with a ground truth (see `set_inlier()`
and `set_inliers()`).
This methods estimates the set of feature weights and of
[effects](@ref Classification::Feature::Effect) that make the
classifier succeed in correctly classifying the sets of inliers given
by the user. These parameters are directly modified within the
`Classification::Feature_base` and `Classification::Label` objects.
\tparam ItemRange model of `ConstRange`. Its iterator type is
`RandomAccessIterator`.
\tparam ItemMap model of `ReadablePropertyMap` whose key
type is the value type of the iterator of `ItemRange` and value type is
the type of the items that are classified.
*/
template <typename ItemRange, typename ItemMap>
class Trainer
{
public:
typedef CGAL::Classifier<ItemRange, ItemMap> Classifier;
typedef typename Classification::Label_handle Label_handle;
typedef typename Classification::Feature_handle Feature_handle;
private:
Classifier* m_classifier;
std::vector<std::vector<std::size_t> > m_training_sets;
std::vector<double> m_precision;
std::vector<double> m_recall;
std::vector<double> m_iou; // intersection over union
public:
/// \name Constructor
/// @{
Trainer (Classifier& classifier)
: m_classifier (&classifier)
, m_training_sets (classifier.number_of_labels())
{
}
/// @}
/// \name Inliers
/// @{
/*!
\brief Adds the item at position `index` as an inlier of
`label` for the training algorithm.
\note This inlier is only used for training. There is no guarantee
that the item at position `index` will be classified as `label`
after calling `run()`, `run_with_local_smoothing()` or
`run_with_graphcut()`.
\return `true` if the inlier was correctly added, `false`
otherwise (if `label` was not found).
*/
bool set_inlier (Label_handle label, std::size_t index)
{
std::size_t label_idx = (std::size_t)(-1);
for (std::size_t i = 0; i < m_classifier->number_of_labels(); ++ i)
if (m_classifier->label(i) == label)
{
label_idx = i;
break;
}
if (label_idx == (std::size_t)(-1))
return false;
if (label_idx >= m_training_sets.size())
m_training_sets.resize (label_idx + 1);
m_training_sets[label_idx].push_back (index);
return true;
}
/*!
\brief Adds the items at positions `indices` as inliers of
`label` for the training algorithm.
\note These inliers are only used for training. There is no
guarantee that the items at positions `indices` will be classified
as `label` after calling `run()`,
`run_with_local_smoothing()` or `run_with_graphcut()`.
\tparam IndexRange range of `std::size_t`, model of `ConstRange`.
*/
template <class IndexRange>
bool set_inliers (Label_handle label,
IndexRange indices)
{
std::size_t label_idx = (std::size_t)(-1);
for (std::size_t i = 0; i < m_classifier->number_of_labels(); ++ i)
if (m_classifier->label(i) == label)
{
label_idx = i;
break;
}
if (label_idx == (std::size_t)(-1))
return false;
if (label_idx >= m_training_sets.size())
m_training_sets.resize (label_idx + 1);
std::copy (indices.begin(), indices.end(), std::back_inserter (m_training_sets[label_idx]));
return true;
}
/*!
\brief Resets inlier sets used for training.
*/
void reset_inlier_sets()
{
std::vector<std::vector<std::size_t > >().swap (m_training_sets);
}
/// @}
/// \name Training
/// @{
/*!
\brief Runs the training algorithm.
All the `Classification::Label` and `Classification::Feature`
necessary for classification should have been added before running
this function. After training, the user can call `run()`,
`run_with_local_smoothing()` or `run_with_graphcut()` to compute
the classification using the estimated parameters.
\param nb_tests number of tests to perform. Higher values may
provide the user with better results at the cost of a higher
computation time. Using a value of at least 10 times the number of
features is advised.
\return minimum ratio (over all labels) of provided
ground truth items correctly classified using the best
configuration found.
*/
double train (std::size_t nb_tests = 300)
{
if (m_training_sets.empty())
return 0.;
for (std::size_t i = 0; i < m_classifier->number_of_labels(); ++ i)
if (m_training_sets.size() <= i || m_training_sets[i].empty())
std::cerr << "WARNING: \"" << m_classifier->label(i)->name() << "\" doesn't have a training set." << std::endl;
std::vector<double> best_weights (m_classifier->number_of_features(), 1.);
struct Feature_training
{
bool skipped;
double wmin;
double wmax;
double factor;
};
std::vector<Feature_training> att_train;
std::size_t nb_trials = 100;
double wmin = 1e-5, wmax = 1e5;
double factor = std::pow (wmax/wmin, 1. / (double)nb_trials);
std::size_t att_used = 0;
for (std::size_t j = 0; j < m_classifier->number_of_features(); ++ j)
{
Feature_handle att = m_classifier->feature(j);
best_weights[j] = att->weight();
std::size_t nb_useful = 0;
double min = (std::numeric_limits<double>::max)();
double max = -(std::numeric_limits<double>::max)();
att->set_weight(wmin);
for (std::size_t i = 0; i < 100; ++ i)
{
estimate_feature_effect(att);
if (feature_useful(att))
{
CGAL_CLASSTRAINING_CERR << "#";
nb_useful ++;
min = (std::min) (min, att->weight());
max = (std::max) (max, att->weight());
}
else
CGAL_CLASSTRAINING_CERR << "-";
att->set_weight(factor * att->weight());
}
CGAL_CLASSTRAINING_CERR << std::endl;
CGAL_CLASSTRAINING_CERR << att->name() << " useful in "
<< nb_useful << "% of the cases, in interval [ "
<< min << " ; " << max << " ]" << std::endl;
att_train.push_back (Feature_training());
att_train.back().skipped = false;
att_train.back().wmin = min / factor;
att_train.back().wmax = max * factor;
if (nb_useful < 2)
{
att_train.back().skipped = true;
att->set_weight(0.);
best_weights[j] = att->weight();
}
else if (best_weights[j] == 1.)
{
att->set_weight(0.5 * (att_train.back().wmin + att_train.back().wmax));
best_weights[j] = att->weight();
++ att_used;
}
else
{
att->set_weight(best_weights[j]);
++ att_used;
}
estimate_feature_effect(att);
}
std::size_t nb_trials_per_feature = 1 + (std::size_t)(nb_tests / (double)(att_used));
CGAL_CLASSIFICATION_CERR << "Trials = " << nb_tests << ", features = " << att_used
<< ", trials per att = " << nb_trials_per_feature << std::endl;
for (std::size_t i = 0; i < att_train.size(); ++ i)
if (!(att_train[i].skipped))
att_train[i].factor = std::pow (att_train[i].wmax / att_train[i].wmin,
1. / (double)nb_trials_per_feature);
double best_score = compute_worst_score(0.);
double best_confidence = compute_worst_confidence(0.);
CGAL_CLASSIFICATION_CERR << "TRAINING GLOBALLY: Best score evolution: " << std::endl;
CGAL_CLASSIFICATION_CERR << 100. * best_score << "% (found at initialization)" << std::endl;
std::size_t current_att_changed = 0;
for (std::size_t i = 0; i < att_used; ++ i)
{
while (att_train[current_att_changed].skipped)
{
++ current_att_changed;
if (current_att_changed == m_classifier->number_of_features())
current_att_changed = 0;
}
std::size_t nb_used = 0;
for (std::size_t j = 0; j < m_classifier->number_of_features(); ++ j)
{
if (j == current_att_changed)
continue;
m_classifier->feature(j)->set_weight(best_weights[j]);
estimate_feature_effect(m_classifier->feature(j));
if (feature_useful(m_classifier->feature(j)))
nb_used ++;
}
Feature_handle current_att = m_classifier->feature(current_att_changed);
const Feature_training& tr = att_train[current_att_changed];
current_att->set_weight(tr.wmin);
for (std::size_t j = 0; j < nb_trials_per_feature; ++ j)
{
estimate_feature_effect(current_att);
double worst_confidence = compute_worst_confidence(best_confidence);
double worst_score = compute_worst_score(best_score);
if (worst_score > best_score
&& worst_confidence > best_confidence)
{
best_score = worst_score;
best_confidence = worst_confidence;
CGAL_CLASSIFICATION_CERR << 100. * best_score << "% (found at iteration "
<< (i * nb_trials_per_feature) + j << "/" << nb_tests << ", "
<< nb_used + (feature_useful(current_att) ? 1 : 0)
<< "/" << m_classifier->number_of_features() << " feature(s) used)" << std::endl;
for (std::size_t k = 0; k < m_classifier->number_of_features(); ++ k)
{
Feature_handle att = m_classifier->feature(k);
best_weights[k] = att->weight();
}
}
current_att->set_weight(current_att->weight() * tr.factor);
}
++ current_att_changed;
}
for (std::size_t i = 0; i < best_weights.size(); ++ i)
{
Feature_handle att = m_classifier->feature(i);
att->set_weight(best_weights[i]);
}
estimate_features_effects();
CGAL_CLASSIFICATION_CERR << std::endl << "Best score found is at least " << 100. * best_score
<< "% of correct classification" << std::endl;
std::size_t nb_removed = 0;
for (std::size_t i = 0; i < best_weights.size(); ++ i)
{
Feature_handle att = m_classifier->feature(i);
CGAL_CLASSTRAINING_CERR << "FEATURE " << att->name() << ": " << best_weights[i] << std::endl;
att->set_weight(best_weights[i]);
Classification::Feature::Effect side = m_classifier->label(0)->feature_effect(att);
bool to_remove = true;
for (std::size_t j = 0; j < m_classifier->number_of_labels(); ++ j)
{
Label_handle clabel = m_classifier->label(j);
if (clabel->feature_effect(att) == Classification::Feature::FAVORING)
CGAL_CLASSTRAINING_CERR << " * Favored for ";
else if (clabel->feature_effect(att) == Classification::Feature::PENALIZING)
CGAL_CLASSTRAINING_CERR << " * Penalized for ";
else
CGAL_CLASSTRAINING_CERR << " * Neutral for ";
if (clabel->feature_effect(att) != side)
to_remove = false;
CGAL_CLASSTRAINING_CERR << clabel->name() << std::endl;
}
if (to_remove)
{
CGAL_CLASSTRAINING_CERR << " -> Useless! Should be removed" << std::endl;
++ nb_removed;
}
}
CGAL_CLASSIFICATION_CERR << nb_removed
<< " feature(s) out of " << m_classifier->number_of_features() << " are useless" << std::endl;
compute_precision_recall();
return best_score;
}
/// @}
/// \name Evaluation
/// @{
/*!
\brief Returns the precision of the training for the given label.
Precision is the number of true positives divided by the sum of
the true positives and the false positives.
*/
double precision (Label_handle label)
{
std::size_t label_idx = (std::size_t)(-1);
for (std::size_t i = 0; i < m_classifier->number_of_labels(); ++ i)
if (m_classifier->label(i) == label)
{
label_idx = i;
break;
}
if (label_idx == (std::size_t)(-1))
return 0.;
return m_precision[label_idx];
}
/*!
\brief Returns the recall of the training for the given label.
Recall is the number of true positives divided by the sum of
the true positives and the false negatives.
*/
double recall (Label_handle label)
{
std::size_t label_idx = (std::size_t)(-1);
for (std::size_t i = 0; i < m_classifier->number_of_labels(); ++ i)
if (m_classifier->label(i) == label)
{
label_idx = i;
break;
}
if (label_idx == (std::size_t)(-1))
return 0.;
return m_recall[label_idx];
}
/*!
\brief Returns the \f$F_1\f$ score of the training for the given label.
\f$F_1\f$ score is the harmonic mean of `precision()` and `recall()`:
\f[
F_1 = 2 \times \frac{precision \times recall}{precision + recall}
\f]
*/
double f1_score (Label_handle label)
{
std::size_t label_idx = (std::size_t)(-1);
for (std::size_t i = 0; i < m_classifier->number_of_labels(); ++ i)
if (m_classifier->label(i) == label)
{
label_idx = i;
break;
}
if (label_idx == (std::size_t)(-1))
return 0.;
return 2. * (m_precision[label_idx] * m_recall[label_idx])
/ (m_precision[label_idx] + m_recall[label_idx]);
}
/*!
\brief Returns the intersection over union of the training for the
given label.
Intersection over union is the number of true positives divided by
the sum of the true positives, of the false positives and of the
false negatives.
*/
double IoU (Label_handle label)
{
std::size_t label_idx = (std::size_t)(-1);
for (std::size_t i = 0; i < m_classifier->number_of_labels(); ++ i)
if (m_classifier->label(i) == label)
{
label_idx = i;
break;
}
if (label_idx == (std::size_t)(-1))
return 0.;
return m_iou[label_idx];
}
/// @}
/// \cond SKIP_IN_MANUAL
Label_handle training_label_of (std::size_t index) const
{
for (std::size_t i = 0; i < m_training_sets.size(); ++ i)
if (std::find (m_training_sets[i].begin(), m_training_sets[i].end(), index) != m_training_sets[i].end())
return m_classifier->label(i);
return Label_handle();
}
/// \endcond
private:
void estimate_features_effects()
{
for (std::size_t i = 0; i < m_classifier->number_of_features(); ++ i)
estimate_feature_effect (m_classifier->feature(i));
}
void estimate_feature_effect (Feature_handle att)
{
std::vector<double> mean (m_classifier->number_of_labels(), 0.);
for (std::size_t j = 0; j < m_classifier->number_of_labels(); ++ j)
{
for (std::size_t k = 0; k < m_training_sets[j].size(); ++ k)
{
double val = att->normalized(m_training_sets[j][k]);
mean[j] += val;
}
mean[j] /= m_training_sets[j].size();
}
std::vector<double> sd (m_classifier->number_of_labels(), 0.);
for (std::size_t j = 0; j < m_classifier->number_of_labels(); ++ j)
{
Label_handle clabel = m_classifier->label(j);
for (std::size_t k = 0; k < m_training_sets[j].size(); ++ k)
{
double val = att->normalized(m_training_sets[j][k]);
sd[j] += (val - mean[j]) * (val - mean[j]);
}
sd[j] = std::sqrt (sd[j] / m_training_sets[j].size());
if (mean[j] - sd[j] > 0.5)
clabel->set_feature_effect (att, Classification::Feature::FAVORING);
else if (mean[j] + sd[j] < 0.5)
clabel->set_feature_effect (att, Classification::Feature::PENALIZING);
else
clabel->set_feature_effect (att, Classification::Feature::NEUTRAL);
}
}
double compute_worst_score (double lower_bound)
{
double worst_score = 1.;
for (std::size_t j = 0; j < m_classifier->number_of_labels(); ++ j)
{
std::size_t nb_okay = 0;
for (std::size_t k = 0; k < m_training_sets[j].size(); ++ k)
{
std::size_t nb_class_best=0;
double val_class_best = (std::numeric_limits<double>::max)();
for(std::size_t l = 0; l < m_classifier->number_of_labels(); ++ l)
{
double value = m_classifier->classification_value (m_classifier->label(l),
m_training_sets[j][k]);
if(val_class_best > value)
{
val_class_best = value;
nb_class_best = l;
}
}
if (nb_class_best == j)
nb_okay ++;
}
double score = nb_okay / (double)(m_training_sets[j].size());
if (score < worst_score)
worst_score = score;
if (worst_score < lower_bound)
return worst_score;
}
return worst_score;
}
double compute_worst_confidence (double lower_bound)
{
double worst_confidence = (std::numeric_limits<double>::max)();
for (std::size_t j = 0; j < m_classifier->number_of_labels(); ++ j)
{
double confidence = 0.;
for (std::size_t k = 0; k < m_training_sets[j].size(); ++ k)
{
std::vector<std::pair<double, std::size_t> > values;
for(std::size_t l = 0; l < m_classifier->number_of_labels(); ++ l)
values.push_back (std::make_pair (m_classifier->classification_value (m_classifier->label(l),
m_training_sets[j][k]),
l));
std::sort (values.begin(), values.end());
if (values[0].second == j)
confidence += values[1].first - values[0].first;
else
{
// for(std::size_t l = 0; l < values.size(); ++ l)
// if (values[l].second == j)
// {
// confidence += values[0].first - values[l].first;
// break;
// }
}
}
confidence /= (double)(m_training_sets[j].size() * m_classifier->number_of_features());
if (confidence < worst_confidence)
worst_confidence = confidence;
if (worst_confidence < lower_bound)
return worst_confidence;
}
return worst_confidence;
}
bool feature_useful (Feature_handle att)
{
Classification::Feature::Effect side = m_classifier->label(0)->feature_effect(att);
for (std::size_t k = 1; k < m_classifier->number_of_labels(); ++ k)
if (m_classifier->label(k)->feature_effect(att) != side)
return true;
return false;
}
void compute_precision_recall ()
{
std::vector<std::size_t> true_positives (m_classifier->number_of_labels());
std::vector<std::size_t> false_positives (m_classifier->number_of_labels());
std::vector<std::size_t> false_negatives (m_classifier->number_of_labels());
for (std::size_t j = 0; j < m_classifier->number_of_labels(); ++ j)
{
for (std::size_t k = 0; k < m_training_sets[j].size(); ++ k)
{
std::size_t nb_class_best=0;
double val_class_best = (std::numeric_limits<double>::max)();
for(std::size_t l = 0; l < m_classifier->number_of_labels(); ++ l)
{
double value = m_classifier->classification_value (m_classifier->label(l),
m_training_sets[j][k]);
if(val_class_best > value)
{
val_class_best = value;
nb_class_best = l;
}
}
if (nb_class_best == j)
++ true_positives[j];
else
{
++ false_negatives[j];
for(std::size_t l = 0; l < m_classifier->number_of_labels(); ++ l)
if (nb_class_best == l)
++ false_positives[l];
}
}
}
m_precision.clear();
m_recall.clear();
for (std::size_t j = 0; j < m_classifier->number_of_labels(); ++ j)
{
m_precision.push_back (true_positives[j] / double(true_positives[j] + false_positives[j]));
m_recall.push_back (true_positives[j] / double(true_positives[j] + false_negatives[j]));
m_iou.push_back (true_positives[j] / double(true_positives[j] + false_positives[j] + false_negatives[j]));
}
}
};
} // namespace Classification
} // namespace CGAL
#endif // CGAL_CLASSIFICATION_TRAINER_H
Reference in New Issue View Git Blame Copy Permalink