ZF/OpenFace
1
0
Fork 0
mirror of https://gitcode.com/gh_mirrors/ope/OpenFace.git synced 2026-08-02 00:27:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Updating dlib and linking to it differently, increasing compilation speed and reducing dependency issues with it.

Browse Source
This commit is contained in:
Tadas Baltrusaitis
2018-06-08 16:48:21 +01:00
parent 8bcc5c46ca
commit cd24af000f
724 changed files with 91433 additions and 92254 deletions
Download Patch File Download Diff File Expand all files Collapse all files

738
lib/3rdParty/dlib/include/dlib/random_forest/random_forest_regression.h vendored Normal file
View File

@@ -0,0 +1,738 @@
// Copyright (C) 2018 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_RANdOM_FOREST_REGRESSION_H_
#define DLIB_RANdOM_FOREST_REGRESSION_H_
#include "random_forest_regression_abstract.h"
#include <vector>
#include "../matrix.h"
#include <algorithm>
#include "../threads.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
class dense_feature_extractor
{
public:
typedef uint32_t feature;
typedef matrix<double,0,1> sample_type;
dense_feature_extractor(
) = default;
void setup (
const std::vector<sample_type>& x,
const std::vector<double>& y
)
{
DLIB_CASSERT(x.size() > 0);
DLIB_CASSERT(x.size() == y.size());
for (auto& el : x)
DLIB_CASSERT(el.size() == x[0].size(), "All the vectors in a training set have to have the same dimensionality.");
DLIB_CASSERT(x[0].size() != 0, "The vectors can't be empty.");
num_feats = x[0].size();
}
void get_random_features (
dlib::rand& rnd,
size_t num,
std::vector<feature>& feats
) const
{
DLIB_ASSERT(max_num_feats() != 0);
num = std::min(num, num_feats);
feats.clear();
for (size_t i = 0; i < num_feats; ++i)
feats.push_back(i);
// now pick num features at random
for (size_t i = 0; i < num; ++i)
{
auto idx = rnd.get_integer_in_range(i,num_feats);
std::swap(feats[i], feats[idx]);
}
feats.resize(num);
}
double extract_feature_value (
const sample_type& item,
const feature& f
) const
{
DLIB_ASSERT(max_num_feats() != 0);
return item(f);
}
size_t max_num_feats (
) const
{
return num_feats;
}
friend void serialize(const dense_feature_extractor& item, std::ostream& out)
{
serialize("dense_feature_extractor", out);
serialize(item.num_feats, out);
}
friend void deserialize(dense_feature_extractor& item, std::istream& in)
{
check_serialized_version("dense_feature_extractor", in);
deserialize(item.num_feats, in);
}
private:
size_t num_feats = 0;
};
// ----------------------------------------------------------------------------------------
template <
typename feature_extractor
>
struct internal_tree_node
{
uint32_t left;
uint32_t right;
float split_threshold;
typename feature_extractor::feature split_feature;
};
template <typename feature_extractor>
void serialize(const internal_tree_node<feature_extractor>& item, std::ostream& out)
{
serialize(item.left, out);
serialize(item.right, out);
serialize(item.split_threshold, out);
serialize(item.split_feature, out);
}
template <typename feature_extractor>
void deserialize(internal_tree_node<feature_extractor>& item, std::istream& in)
{
deserialize(item.left, in);
deserialize(item.right, in);
deserialize(item.split_threshold, in);
deserialize(item.split_feature, in);
}
// ----------------------------------------------------------------------------------------
template <
typename feature_extractor = dense_feature_extractor
>
class random_forest_regression_function
{
public:
typedef feature_extractor feature_extractor_type;
typedef typename feature_extractor::sample_type sample_type;
random_forest_regression_function(
) = default;
random_forest_regression_function (
feature_extractor_type&& fe_,
std::vector<std::vector<internal_tree_node<feature_extractor>>>&& trees_,
std::vector<std::vector<float>>&& leaves_
) :
fe(std::move(fe_)),
trees(std::move(trees_)),
leaves(std::move(leaves_))
{
DLIB_ASSERT(trees.size() > 0);
DLIB_ASSERT(trees.size() == leaves.size(), "Every set of tree nodes has to have leaves");
#ifdef ENABLE_ASSERTS
for (size_t i = 0; i < trees.size(); ++i)
{
DLIB_ASSERT(trees[i].size() > 0, "A tree can't have 0 leaves.");
for (auto& node : trees[i])
{
DLIB_ASSERT(trees[i].size()+leaves[i].size() > node.left, "left node index in tree is too big. There is no associated tree node or leaf.");
DLIB_ASSERT(trees[i].size()+leaves[i].size() > node.right, "right node index in tree is too big. There is no associated tree node or leaf.");
}
}
#endif
}
size_t get_num_trees(
) const
{
return trees.size();
}
const std::vector<std::vector<internal_tree_node<feature_extractor>>>& get_internal_tree_nodes (
) const { return trees; }
const std::vector<std::vector<float>>& get_tree_leaves (
) const { return leaves; }
const feature_extractor_type& get_feature_extractor (
) const { return fe; }
double operator() (
const sample_type& x
) const
{
DLIB_ASSERT(get_num_trees() > 0);
double accum = 0;
for (size_t i = 0; i < trees.size(); ++i)
{
auto& tree = trees[i];
// walk the tree to the leaf
uint32_t idx = 0;
while(idx < tree.size())
{
auto feature_value = fe.extract_feature_value(x, tree[idx].split_feature);
if (feature_value < tree[idx].split_threshold)
idx = tree[idx].left;
else
idx = tree[idx].right;
}
// compute leaf index
accum += leaves[i][idx-tree.size()];
}
return accum/trees.size();
}
friend void serialize(const random_forest_regression_function& item, std::ostream& out)
{
serialize("random_forest_regression_function", out);
serialize(item.fe, out);
serialize(item.trees, out);
serialize(item.leaves, out);
}
friend void deserialize(random_forest_regression_function& item, std::istream& in)
{
check_serialized_version("random_forest_regression_function", in);
deserialize(item.fe, in);
deserialize(item.trees, in);
deserialize(item.leaves, in);
}
private:
/*!
CONVENTION
- trees.size() == leaves.size()
- Any .left or .right index in trees that is larger than the number of
nodes in the tree references a leaf. Moreover, the index of the leaf is
computed by subtracting the number of nodes in the tree.
!*/
feature_extractor_type fe;
// internal nodes of trees
std::vector<std::vector<internal_tree_node<feature_extractor>>> trees;
// leaves of trees
std::vector<std::vector<float>> leaves;
};
// ----------------------------------------------------------------------------------------
template <
typename feature_extractor = dense_feature_extractor
>
class random_forest_regression_trainer
{
public:
typedef feature_extractor feature_extractor_type;
typedef random_forest_regression_function<feature_extractor> trained_function_type;
typedef typename feature_extractor::sample_type sample_type;
random_forest_regression_trainer (
) = default;
const feature_extractor_type& get_feature_extractor (
) const
{
return fe_;
}
void set_feature_extractor (
const feature_extractor_type& feat_extractor
)
{
fe_ = feat_extractor;
}
void set_seed (
const std::string& seed
)
{
random_seed = seed;
}
const std::string& get_random_seed (
) const
{
return random_seed;
}
size_t get_num_trees (
) const
{
return num_trees;
}
void set_num_trees (
size_t num
)
{
DLIB_CASSERT(num > 0);
num_trees = num;
}
void set_feature_subsampling_fraction (
double frac
)
{
DLIB_CASSERT(0 < frac && frac <= 1);
feature_subsampling_frac = frac;
}
double get_feature_subsampling_frac(
) const
{
return feature_subsampling_frac;
}
void set_min_samples_per_leaf (
size_t num
)
{
DLIB_ASSERT(num > 0);
min_samples_per_leaf = num;
}
size_t get_min_samples_per_leaf(
) const
{
return min_samples_per_leaf;
}
void be_verbose (
)
{
verbose = true;
}
void be_quiet (
)
{
verbose = false;
}
trained_function_type train (
const std::vector<sample_type>& x,
const std::vector<double>& y
) const
{
std::vector<double> junk;
return do_train(x,y,junk,false);
}
trained_function_type train (
const std::vector<sample_type>& x,
const std::vector<double>& y,
std::vector<double>& oob_values
) const
{
return do_train(x,y,oob_values,true);
}
private:
trained_function_type do_train (
const std::vector<sample_type>& x,
const std::vector<double>& y,
std::vector<double>& oob_values,
bool compute_oob_values
) const
{
DLIB_CASSERT(x.size() == y.size());
DLIB_CASSERT(x.size() > 0);
feature_extractor_type fe = fe_;
fe.setup(x,y);
DLIB_CASSERT(fe.max_num_feats() != 0);
std::vector<std::vector<internal_tree_node<feature_extractor>>> all_trees(num_trees);
std::vector<std::vector<float>> all_leaves(num_trees);
const double sumy = sum(mat(y));
const size_t feats_per_node = std::max(1.0,std::round(fe.max_num_feats()*feature_subsampling_frac));
// Each tree couldn't have more than this many interior nodes. It might
// end up having less though. We need to know this value because the way
// we mark a left or right pointer in a tree as pointing to a leaf is by
// making its index larger than the number of interior nodes in the tree.
// But we don't know the tree's size before we finish building it. So we
// will use max_num_nodes as a proxy during tree construction and then go
// back and fix it once a tree's size is known.
const uint32_t max_num_nodes = y.size();
std::vector<uint32_t> oob_hits;
if (compute_oob_values)
{
oob_values.resize(y.size());
oob_hits.resize(y.size());
}
std::mutex m;
// Calling build_tree(i) creates the ith tree and stores the results in
// all_trees and all_leaves.
auto build_tree = [&](long i)
{
dlib::rand rnd(random_seed + std::to_string(i));
auto& tree = all_trees[i];
auto& leaves = all_leaves[i];
// Check if there are fewer than min_samples_per_leaf and if so then
// don't make any tree. Just average the things and be done.
if (y.size() <= min_samples_per_leaf)
{
leaves.push_back(sumy/y.size());
return;
}
// pick a random bootstrap of the data.
std::vector<std::pair<float,uint32_t>> idxs(y.size());
for (auto& idx : idxs)
idx = std::make_pair(0,rnd.get_integer(y.size()));
// We are going to use ranges_to_process as a stack that tracks which
// range of samples we are going to split next.
std::vector<range_t> ranges_to_process;
// start with the root of the tree, i.e. the entire range of training
// samples.
ranges_to_process.emplace_back(sumy,0,y.size());
// push an unpopulated root node into the tree. We will populate it
// when we process its corresponding range.
tree.emplace_back();
std::vector<typename feature_extractor::feature> feats;
while(ranges_to_process.size() > 0)
{
// Grab the next range/node to process.
const auto range = ranges_to_process.back();
ranges_to_process.pop_back();
// Get the split features we will consider at this node.
fe.get_random_features(rnd, feats_per_node, feats);
// Then find the best split
auto best_split = find_best_split_among_feats(fe, range, feats, x, y, idxs);
range_t left_split(best_split.left_sum, range.begin, best_split.split_idx);
range_t right_split(best_split.right_sum, best_split.split_idx, range.end);
DLIB_ASSERT(left_split.begin < left_split.end);
DLIB_ASSERT(right_split.begin < right_split.end);
// Now that we know the split we can populate the parent node we popped
// from ranges_to_process.
tree[range.tree_idx].split_threshold = best_split.split_threshold;
tree[range.tree_idx].split_feature = best_split.split_feature;
// If the left split is big enough to make a new interior leaf
// node. We also stop splitting if all the samples went into this node.
// This could happen if the features are all uniform so there just
// isn't any way to split them anymore.
if (left_split.size() > min_samples_per_leaf && right_split.size() != 0)
{
// allocate an interior leaf node for it.
left_split.tree_idx = tree.size();
tree.emplace_back();
// set the pointer in the parent node to the newly allocated
// node.
tree[range.tree_idx].left = left_split.tree_idx;
ranges_to_process.emplace_back(left_split);
}
else
{
// Add to leaves. Don't forget to set the pointer in the
// parent node to the newly allocated leaf node.
tree[range.tree_idx].left = leaves.size() + max_num_nodes;
leaves.emplace_back(left_split.avg());
}
// If the right split is big enough to make a new interior leaf
// node. We also stop splitting if all the samples went into this node.
// This could happen if the features are all uniform so there just
// isn't any way to split them anymore.
if (right_split.size() > min_samples_per_leaf && left_split.size() != 0)
{
// allocate an interior leaf node for it.
right_split.tree_idx = tree.size();
tree.emplace_back();
// set the pointer in the parent node to the newly allocated
// node.
tree[range.tree_idx].right = right_split.tree_idx;
ranges_to_process.emplace_back(right_split);
}
else
{
// Add to leaves. Don't forget to set the pointer in the
// parent node to the newly allocated leaf node.
tree[range.tree_idx].right = leaves.size() + max_num_nodes;
leaves.emplace_back(right_split.avg());
}
} // end while (still building tree)
// Fix the leaf pointers in the tree now that we know the correct
// tree.size() value.
DLIB_CASSERT(max_num_nodes >= tree.size());
const auto offset = max_num_nodes - tree.size();
for (auto& n : tree)
{
if (n.left >= max_num_nodes)
n.left -= offset;
if (n.right >= max_num_nodes)
n.right -= offset;
}
if (compute_oob_values)
{
std::sort(idxs.begin(), idxs.end(),
[](const std::pair<float,uint32_t>& a, const std::pair<float,uint32_t>& b) {return a.second<b.second; });
std::lock_guard<std::mutex> lock(m);
size_t j = 0;
for (size_t i = 0; i < oob_values.size(); ++i)
{
// check if i is in idxs
while(j < idxs.size() && i > idxs[j].second)
++j;
// i isn't in idxs so it's an oob sample and we should process it.
if (j == idxs.size() || idxs[j].second != i)
{
oob_hits[i]++;
// walk the tree to find the leaf value for this oob sample
uint32_t idx = 0;
while(idx < tree.size())
{
auto feature_value = fe.extract_feature_value(x[i], tree[idx].split_feature);
if (feature_value < tree[idx].split_threshold)
idx = tree[idx].left;
else
idx = tree[idx].right;
}
oob_values[i] += leaves[idx-tree.size()];
}
}
}
};
if (verbose)
parallel_for_verbose(0, num_trees, build_tree);
else
parallel_for(0, num_trees, build_tree);
if (compute_oob_values)
{
double meanval = 0;
double cnt = 0;
for (size_t i = 0; i < oob_values.size(); ++i)
{
if (oob_hits[i] != 0)
{
oob_values[i] /= oob_hits[i];
meanval += oob_values[i];
++cnt;
}
}
// If there are some elements that didn't get hits, we set their oob values
// to the mean oob value.
if (cnt != 0)
{
const double typical_value = meanval/cnt;
for (size_t i = 0; i < oob_values.size(); ++i)
{
if (oob_hits[i] == 0)
oob_values[i] = typical_value;
}
}
}
return trained_function_type(std::move(fe), std::move(all_trees), std::move(all_leaves));
}
struct range_t
{
range_t(
double sumy,
uint32_t begin,
uint32_t end
) : sumy(sumy), begin(begin), end(end), tree_idx(0) {}
double sumy;
uint32_t begin;
uint32_t end;
// Every range object corresponds to an entry in a tree. This tells you the
// tree node that owns the range.
uint32_t tree_idx;
uint32_t size() const { return end-begin; }
double avg() const { return sumy/size(); }
};
struct best_split_details
{
double score = -std::numeric_limits<double>::infinity();
double left_sum;
double right_sum;
uint32_t split_idx;
double split_threshold;
typename feature_extractor::feature split_feature;
bool operator < (const best_split_details& rhs) const
{
return score < rhs.score;
}
};
static best_split_details find_best_split (
const range_t& range,
const std::vector<double>& y,
const std::vector<std::pair<float,uint32_t>>& idxs
)
/*!
requires
- max(mat(idxs)) < y.size()
- range.sumy == sum of y[idxs[j].second] for all valid j in range [range.begin, range.end).
ensures
- finds a threshold T such that there exists an i satisfying the following:
- y[idxs[j].second] < T for all j <= i
- y[idxs[j].second] > T for all j > i
Therefore, the threshold T partitions the contents of y into two groups,
relative to the ordering established by idxs. Moreover the partitioning
of y values into two groups has the additional requirement that it is
optimal in the sense that the sum of the squared deviations from each
partition's mean is minimized.
!*/
{
size_t best_i = range.begin;
double best_score = -1;
double left_sum = 0;
double best_left_sum = y[idxs[range.begin].second];
const auto size = range.size();
size_t left_size = 0;
for (size_t i = range.begin; i+1 < range.end; ++i)
{
++left_size;
left_sum += y[idxs[i].second];
// Don't split here because the next element has the same feature value so
// we can't *really* split here.
if (idxs[i].first==idxs[i+1].first)
continue;
const double right_sum = range.sumy-left_sum;
const double score = left_sum*left_sum/left_size + right_sum*right_sum/(size-left_size);
if (score > best_score)
{
best_score = score;
best_i = i;
best_left_sum = left_sum;
}
}
best_split_details result;
result.score = best_score;
result.left_sum = best_left_sum;
result.right_sum = range.sumy-best_left_sum;
result.split_idx = best_i+1; // one past the end of the left range
result.split_threshold = (idxs[best_i].first+idxs[best_i+1].first)/2;
return result;
}
static best_split_details find_best_split_among_feats(
const feature_extractor& fe,
const range_t& range,
const std::vector<typename feature_extractor::feature>& feats,
const std::vector<sample_type>& x,
const std::vector<double>& y,
std::vector<std::pair<float,uint32_t>>& idxs
)
{
auto compare_first = [](const std::pair<float,uint32_t>& a, const std::pair<float,uint32_t>& b) { return a.first<b.first; };
best_split_details best;
for (auto& feat : feats)
{
// Extract feature values for this feature and sort the indexes based on
// that feature so we can then find the best split.
for (auto i = range.begin; i < range.end; ++i)
idxs[i].first = fe.extract_feature_value(x[idxs[i].second], feat);
std::sort(idxs.begin()+range.begin, idxs.begin()+range.end, compare_first);
auto split = find_best_split(range, y, idxs);
if (best < split)
{
best = split;
best.split_feature = feat;
}
}
// resort idxs based on winning feat
for (auto i = range.begin; i < range.end; ++i)
idxs[i].first = fe.extract_feature_value(x[idxs[i].second], best.split_feature);
std::sort(idxs.begin()+range.begin, idxs.begin()+range.end, compare_first);
return best;
}
std::string random_seed;
size_t num_trees = 1000;
double feature_subsampling_frac = 1.0/3.0;
size_t min_samples_per_leaf = 5;
feature_extractor_type fe_;
bool verbose = false;
};
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_RANdOM_FOREST_REGRESSION_H_

460
lib/3rdParty/dlib/include/dlib/random_forest/random_forest_regression_abstract.h vendored Normal file
View File

@@ -0,0 +1,460 @@
// Copyright (C) 2018 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#undef DLIB_RANdOM_FOREST_REGRESION_ABSTRACT_H_
#ifdef DLIB_RANdOM_FOREST_REGRESION_ABSTRACT_H_
#include <vector>
#include "../matrix.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
class dense_feature_extractor
{
/*!
WHAT THIS OBJECT REPRESENTS
This object is a tool for extracting features from objects. In particular,
it is designed to be used with the random forest regression tools discussed
below.
This particular feature extract does almost nothing since it works on
vectors in R^n and simply selects elements from each vector. However, the
tools below are templated and allow you to design your own feature extractors
that operate on whatever object types you create. So for example, maybe
you want to perform regression on images rather than vectors. Moreover,
your feature extraction could be more complex. Maybe you are selecting
differences between pairs of pixels in an image or doing something
involving geometric transforms during feature extraction. Any of these
kinds of more complex feature extraction patterns can be realized with the
random forest tools by implementing your own feature extractor object and
using it with the random forest objects.
Therefore, you should consider this dense_feature_extractor as an example
that documents the interface as well as the simple default extractor for
use with dense vectors.
THREAD SAFETY
It is safe to call const members of this object from multiple threads. ANY
USER DEFINED FEATURE EXTRACTORS MUST ALSO MEET THIS GUARONTEE AS WELL SINCE
IT IS ASSUMED BY THE RANDOM FOREST TRAINING ROUTINES.
!*/
public:
typedef uint32_t feature;
typedef matrix<double,0,1> sample_type;
dense_feature_extractor(
);
/*!
ensures
- #max_num_feats() == 0
!*/
void setup (
const std::vector<sample_type>& x,
const std::vector<double>& y
);
/*!
requires
- x.size() == y.size()
- x.size() > 0
- x[0].size() > 0
- all the vectors in x have the same dimensionality.
ensures
- Configures this feature extractor to work on the given training data.
For dense feature extractors all we do is record the dimensionality of
the training vectors.
- #max_num_feats() == x[0].size()
(In general, setup() sets max_num_feats() to some non-zero value so that
the other methods of this object can then be called. The point of setup()
is to allow a feature extractor to gather whatever statistics it needs from
training data. That is, more complex feature extraction strategies my
themselves be trained from data.)
!*/
void get_random_features (
dlib::rand& rnd,
size_t num,
std::vector<feature>& feats
) const;
/*!
requires
- max_num_feats() != 0
ensures
- #feats.size() == min(num, max_num_feats())
- This function randomly identifies num features and stores them into feats.
These feature objects can then be used with extract_feature_value() to
obtain a value from any particular sample_type object. This value is the
"feature value" used by a decision tree algorithm to deice how to split
and traverse trees.
- The above two conditions define the behavior of get_random_features() in
general. For this specific implementation of the feature extraction interface
this function selects num integer values from the range [0, max_num_feats()),
without replacement. These values are stored into feats.
!*/
double extract_feature_value (
const sample_type& item,
const feature& f
) const;
/*!
requires
- #max_num_feats() != 0
- f was produced from a call to get_random_features().
ensures
- Extracts the feature value corresponding to f. For this simple dense
feature extractor this simply means returning item(f). But in general
you can design feature extractors that do something more complex.
!*/
size_t max_num_feats (
) const;
/*!
ensures
- returns the number of distinct features this object might extract. That is,
a feature extractor essentially defines a mapping from sample_type objects to
vectors in R^max_num_feats().
!*/
};
void serialize(const dense_feature_extractor& item, std::ostream& out);
void deserialize(dense_feature_extractor& item, std::istream& in);
/*!
provides serialization support
!*/
// ----------------------------------------------------------------------------------------
template <
typename feature_extractor
>
struct internal_tree_node
{
/*!
WHAT THIS OBJECT REPRESENTS
This object is an internal node in a regression tree. See the code of
random_forest_regression_function to see how it is used to create a tree.
!*/
uint32_t left;
uint32_t right;
float split_threshold;
typename feature_extractor::feature split_feature;
};
template <typename feature_extractor>
void serialize(const internal_tree_node<feature_extractor>& item, std::ostream& out);
template <typename feature_extractor>
void deserialize(internal_tree_node<feature_extractor>& item, std::istream& in);
/*!
provides serialization support
!*/
// ----------------------------------------------------------------------------------------
template <
typename feature_extractor = dense_feature_extractor
>
class random_forest_regression_function
{
/*!
REQUIREMENTS ON feature_extractor
feature_extractor must be dense_feature_extractor or a type with a
compatible interface.
WHAT THIS OBJECT REPRESENTS
This object represents a regression forest. This is a collection of
decision trees that take an object as input and each vote on a real value
to associate with the object. The final real value output is the average
of all the votes from each of the trees.
!*/
public:
typedef feature_extractor feature_extractor_type;
typedef typename feature_extractor::sample_type sample_type;
random_forest_regression_function(
);
/*!
ensures
- #num_trees() == 0
!*/
random_forest_regression_function (
feature_extractor_type&& fe_,
std::vector<std::vector<internal_tree_node<feature_extractor>>>&& trees_,
std::vector<std::vector<float>>&& leaves_
);
/*!
requires
- trees.size() > 0
- trees.size() = leaves.size()
- for all valid i:
- leaves[i].size() > 0
- trees[i].size()+leaves[i].size() > the maximal left or right index values in trees[i].
(i.e. each left or right value must index to some existing internal tree node or leaf node).
ensures
- #get_internal_tree_nodes() == trees_
- #get_tree_leaves() == leaves_
- #get_feature_extractor() == fe_
!*/
size_t get_num_trees(
) const;
/*!
ensures
- returns the number of trees in this regression forest.
!*/
const std::vector<std::vector<internal_tree_node<feature_extractor>>>& get_internal_tree_nodes (
) const;
/*!
ensures
- returns the internal tree nodes that define the regression trees.
- get_internal_tree_nodes().size() == get_num_trees()
!*/
const std::vector<std::vector<float>>& get_tree_leaves (
) const;
/*!
ensures
- returns the tree leaves that define the regression trees.
- get_tree_leaves().size() == get_num_trees()
!*/
const feature_extractor_type& get_feature_extractor (
) const;
/*!
ensures
- returns the feature extractor used by the trees.
!*/
double operator() (
const sample_type& x
) const;
/*!
requires
- get_num_trees() > 0
ensures
- Maps x to a real value and returns the value. To do this, we find the
get_num_trees() leaf values associated with x and then return the average
of these leaf values.
!*/
};
void serialize(const random_forest_regression_function& item, std::ostream& out);
void deserialize(random_forest_regression_function& item, std::istream& in);
/*!
provides serialization support
!*/
// ----------------------------------------------------------------------------------------
template <
typename feature_extractor = dense_feature_extractor
>
class random_forest_regression_trainer
{
/*!
REQUIREMENTS ON feature_extractor
feature_extractor must be dense_feature_extractor or a type with a
compatible interface.
WHAT THIS OBJECT REPRESENTS
This object implements Breiman's classic random forest regression
algorithm. The algorithm learns to map objects, nominally vectors in R^n,
into the reals. It essentially optimizes the mean squared error by fitting
a bunch of decision trees, each of which vote on the output value of the
regressor. The final prediction is obtained by averaging all the
predictions.
For more information on the algorithm see:
Breiman, Leo. "Random forests." Machine learning 45.1 (2001): 5-32.
!*/
public:
typedef feature_extractor feature_extractor_type;
typedef random_forest_regression_function<feature_extractor> trained_function_type;
typedef typename feature_extractor::sample_type sample_type;
random_forest_regression_trainer (
);
/*!
ensures
- #get_min_samples_per_leaf() == 5
- #get_num_trees() == 1000
- #get_feature_subsampling_frac() == 1.0/3.0
- #get_feature_extractor() == a default initialized feature extractor.
- #get_random_seed() == ""
- this object is not verbose.
!*/
const feature_extractor_type& get_feature_extractor (
) const;
/*!
ensures
- returns the feature extractor used when train() is invoked.
!*/
void set_feature_extractor (
const feature_extractor_type& feat_extractor
);
/*!
ensures
- #get_feature_extractor() == feat_extractor
!*/
void set_seed (
const std::string& seed
);
/*!
ensures
- #get_random_seed() == seed
!*/
const std::string& get_random_seed (
) const;
/*!
ensures
- A central part of this algorithm is random selection of both training
samples and features. This function returns the seed used to initialized
the random number generator used for these random selections.
!*/
size_t get_num_trees (
) const;
/*!
ensures
- Random forests built by this object will contain get_num_trees() trees.
!*/
void set_num_trees (
size_t num
);
/*!
requires
- num > 0
ensures
- #get_num_trees() == num
!*/
void set_feature_subsampling_fraction (
double frac
);
/*!
requires
- 0 < frac <= 1
ensures
- #get_feature_subsampling_frac() == frac
!*/
double get_feature_subsampling_frac(
) const;
/*!
ensures
- When we build trees, at each node we don't look at all the available
features. We consider only get_feature_subsampling_frac() fraction of
them, selected at random.
!*/
void set_min_samples_per_leaf (
size_t num
);
/*!
requires
- num > 0
ensures
- #get_min_samples_per_leaf() == num
!*/
size_t get_min_samples_per_leaf(
) const;
/*!
ensures
- When building trees, each leaf node in a tree will contain at least
get_min_samples_per_leaf() samples. This means that the output votes of
each tree are averages of at least get_min_samples_per_leaf() y values.
!*/
void be_verbose (
);
/*!
ensures
- This object will print status messages to standard out so that the
progress of training can be tracked..
!*/
void be_quiet (
);
/*!
ensures
- this object will not print anything to standard out
!*/
random_forest_regression_function<feature_extractor> train (
const std::vector<sample_type>& x,
const std::vector<double>& y,
std::vector<double>& oob_values
) const;
/*!
requires
- x.size() == y.size()
- x.size() > 0
- Running following code:
auto fe = get_feature_extractor()
fe.setup(x,y);
Must be valid and result in fe.max_num_feats() != 0
ensures
- This function fits a regression forest to the given training data. The
goal being to regress x to y in the mean squared sense. It therefore
fits regression trees and returns the resulting random_forest_regression_function
RF, which will have the following properties:
- RF.get_num_trees() == get_num_trees()
- for all valid i:
- RF(x[i]) should output a value close to y[i]
- RF.get_feature_extractor() will be a copy of this->get_feature_extractor()
that has been configured by a call the feature extractor's setup() routine.
To run the algorithm we need to use a feature extractor. We obtain a
valid feature extractor by making a copy of get_feature_extractor(), then
invoking setup(x,y) on it. This feature extractor is what is used to fit
the trees and is also the feature extractor stored in the returned random
forest.
- #oob_values.size() == y.size()
- for all valid i:
- #oob_values[i] == the "out of bag" prediction for y[i]. It is
calculated by computing the average output from trees not trained on
y[i]. This is similar to a leave-one-out cross-validation prediction
of y[i] and can be used to estimate the generalization error of the
regression forest.
- Training uses all the available CPU cores.
!*/
random_forest_regression_function<feature_extractor> train (
const std::vector<sample_type>& x,
const std::vector<double>& y
) const;
/*!
requires
- x.size() == y.size()
- x.size() > 0
- Running following code:
auto fe = get_feature_extractor()
fe.setup(x,y);
Must be valid and result in fe.max_num_feats() != 0
ensures
- This function is identical to train(x,y,oob_values) except that the
oob_values are not calculated.
!*/
};
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_RANdOM_FOREST_REGRESION_ABSTRACT_H_