// Copyright (C) 2014 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_SHAPE_PREDICToR_TRAINER_H_
#define DLIB_SHAPE_PREDICToR_TRAINER_H_
#include "shape_predictor_trainer_abstract.h"
#include "shape_predictor.h"
#include "../console_progress_indicator.h"
#include "../threads.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
class shape_predictor_trainer
{
/*!
This thing really only works with unsigned char or rgb_pixel images (since we assume the threshold
should be in the range [-128,128]).
!*/
public:
enum padding_mode_t
{
bounding_box_relative,
landmark_relative
};
shape_predictor_trainer (
)
{
_cascade_depth = 10;
_tree_depth = 4;
_num_trees_per_cascade_level = 500;
_nu = 0.1;
_oversampling_amount = 20;
_feature_pool_size = 400;
_lambda = 0.1;
_num_test_splits = 20;
_feature_pool_region_padding = 0;
_verbose = false;
_num_threads = 0;
_padding_mode = landmark_relative;
}
unsigned long get_cascade_depth (
) const { return _cascade_depth; }
void set_cascade_depth (
unsigned long depth
)
{
DLIB_CASSERT(depth > 0,
"\t void shape_predictor_trainer::set_cascade_depth()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t depth: " << depth
);
_cascade_depth = depth;
}
unsigned long get_tree_depth (
) const { return _tree_depth; }
void set_tree_depth (
unsigned long depth
)
{
DLIB_CASSERT(depth > 0,
"\t void shape_predictor_trainer::set_tree_depth()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t depth: " << depth
);
_tree_depth = depth;
}
unsigned long get_num_trees_per_cascade_level (
) const { return _num_trees_per_cascade_level; }
void set_num_trees_per_cascade_level (
unsigned long num
)
{
DLIB_CASSERT( num > 0,
"\t void shape_predictor_trainer::set_num_trees_per_cascade_level()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t num: " << num
);
_num_trees_per_cascade_level = num;
}
double get_nu (
) const { return _nu; }
void set_nu (
double nu
)
{
DLIB_CASSERT(0 < nu && nu <= 1,
"\t void shape_predictor_trainer::set_nu()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t nu: " << nu
);
_nu = nu;
}
std::string get_random_seed (
) const { return rnd.get_seed(); }
void set_random_seed (
const std::string& seed
) { rnd.set_seed(seed); }
unsigned long get_oversampling_amount (
) const { return _oversampling_amount; }
void set_oversampling_amount (
unsigned long amount
)
{
DLIB_CASSERT(amount > 0,
"\t void shape_predictor_trainer::set_oversampling_amount()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t amount: " << amount
);
_oversampling_amount = amount;
}
unsigned long get_feature_pool_size (
) const { return _feature_pool_size; }
void set_feature_pool_size (
unsigned long size
)
{
DLIB_CASSERT(size > 1,
"\t void shape_predictor_trainer::set_feature_pool_size()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t size: " << size
);
_feature_pool_size = size;
}
double get_lambda (
) const { return _lambda; }
void set_lambda (
double lambda
)
{
DLIB_CASSERT(lambda > 0,
"\t void shape_predictor_trainer::set_lambda()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t lambda: " << lambda
);
_lambda = lambda;
}
unsigned long get_num_test_splits (
) const { return _num_test_splits; }
void set_num_test_splits (
unsigned long num
)
{
DLIB_CASSERT(num > 0,
"\t void shape_predictor_trainer::set_num_test_splits()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t num: " << num
);
_num_test_splits = num;
}
void set_padding_mode (
padding_mode_t mode
)
{
_padding_mode = mode;
}
padding_mode_t get_padding_mode (
) const { return _padding_mode; }
double get_feature_pool_region_padding (
) const { return _feature_pool_region_padding; }
void set_feature_pool_region_padding (
double padding
)
{
DLIB_CASSERT(padding > -0.5,
"\t void shape_predictor_trainer::set_feature_pool_region_padding()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t padding: " << padding
);
_feature_pool_region_padding = padding;
}
void be_verbose (
)
{
_verbose = true;
}
void be_quiet (
)
{
_verbose = false;
}
unsigned long get_num_threads (
) const { return _num_threads; }
void set_num_threads (
unsigned long num
)
{
_num_threads = num;
}
template <typename image_array>
shape_predictor train (
const image_array& images,
const std::vector<std::vector<full_object_detection> >& objects
) const
{
using namespace impl;
DLIB_CASSERT(images.size() == objects.size() && images.size() > 0,
"\t shape_predictor shape_predictor_trainer::train()"
<< "\n\t Invalid inputs were given to this function. "
<< "\n\t images.size(): " << images.size()
<< "\n\t objects.size(): " << objects.size()
);
// make sure the objects agree on the number of parts and that there is at
// least one full_object_detection.
unsigned long num_parts = 0;
std::vector<int> part_present;
for (unsigned long i = 0; i < objects.size(); ++i)
{
for (unsigned long j = 0; j < objects[i].size(); ++j)
{
if (num_parts == 0)
{
num_parts = objects[i][j].num_parts();
DLIB_CASSERT(objects[i][j].num_parts() != 0,
"\t shape_predictor shape_predictor_trainer::train()"
<< "\n\t You can't give objects that don't have any parts to the trainer."
);
part_present.resize(num_parts);
}
else
{
DLIB_CASSERT(objects[i][j].num_parts() == num_parts,
"\t shape_predictor shape_predictor_trainer::train()"
<< "\n\t All the objects must agree on the number of parts. "
<< "\n\t objects["<<i<<"]["<<j<<"].num_parts(): " << objects[i][j].num_parts()
<< "\n\t num_parts: " << num_parts
);
}
for (unsigned long p = 0; p < objects[i][j].num_parts(); ++p)
{
if (objects[i][j].part(p) != OBJECT_PART_NOT_PRESENT)
part_present[p] = 1;
}
}
}
DLIB_CASSERT(num_parts != 0,
"\t shape_predictor shape_predictor_trainer::train()"
<< "\n\t You must give at least one full_object_detection if you want to train a shape model and it must have parts."
);
DLIB_CASSERT(sum(mat(part_present)) == (long)num_parts,
"\t shape_predictor shape_predictor_trainer::train()"
<< "\n\t Each part must appear at least once in this training data. That is, "
<< "\n\t you can't have a part that is always set to OBJECT_PART_NOT_PRESENT."
);
// creating thread pool. if num_threads <= 1, trainer should work in caller thread
thread_pool tp(_num_threads > 1 ? _num_threads : 0);
// determining the type of features used for this type of images
typedef typename std::remove_const<typename std::remove_reference<decltype(images[0])>::type>::type image_type;
typedef typename image_traits<image_type>::pixel_type pixel_type;
typedef typename pixel_traits<pixel_type>::basic_pixel_type feature_type;
rnd.set_seed(get_random_seed());
std::vector<training_sample<feature_type>> samples;
const matrix<float,0,1> initial_shape = populate_training_sample_shapes(objects, samples);
const std::vector<std::vector<dlib::vector<float,2> > > pixel_coordinates = randomly_sample_pixel_coordinates(initial_shape);
unsigned long trees_fit_so_far = 0;
console_progress_indicator pbar(get_cascade_depth()*get_num_trees_per_cascade_level());
if (_verbose)
std::cout << "Fitting trees..." << std::endl;
std::vector<std::vector<impl::regression_tree> > forests(get_cascade_depth());
// Now start doing the actual training by filling in the forests
for (unsigned long cascade = 0; cascade < get_cascade_depth(); ++cascade)
{
// Each cascade uses a different set of pixels for its features. We compute
// their representations relative to the initial shape first.
std::vector<unsigned long> anchor_idx;
std::vector<dlib::vector<float,2> > deltas;
create_shape_relative_encoding(initial_shape, pixel_coordinates[cascade], anchor_idx, deltas);
// First compute the feature_pixel_values for each training sample at this
// level of the cascade.
parallel_for(tp, 0, samples.size(), [&](unsigned long i)
{
impl::extract_feature_pixel_values(images[samples[i].image_idx], samples[i].rect,
samples[i].current_shape, initial_shape, anchor_idx,
deltas, samples[i].feature_pixel_values);
}, 1);
// Now start building the trees at this cascade level.
for (unsigned long i = 0; i < get_num_trees_per_cascade_level(); ++i)
{
forests[cascade].push_back(make_regression_tree(tp, samples, pixel_coordinates[cascade]));
if (_verbose)
{
++trees_fit_so_far;
pbar.print_status(trees_fit_so_far);
}
}
}
if (_verbose)
std::cout << "Training complete " << std::endl;
return shape_predictor(initial_shape, forests, pixel_coordinates);
}
private:
static void object_to_shape (
const full_object_detection& obj,
matrix<float,0,1>& shape,
matrix<float,0,1>& present // a mask telling which elements of #shape are present.
)
{
shape.set_size(obj.num_parts()*2);
present.set_size(obj.num_parts()*2);
const point_transform_affine tform_from_img = impl::normalizing_tform(obj.get_rect());
for (unsigned long i = 0; i < obj.num_parts(); ++i)
{
if (obj.part(i) != OBJECT_PART_NOT_PRESENT)
{
vector<float,2> p = tform_from_img(obj.part(i));
shape(2*i) = p.x();
shape(2*i+1) = p.y();
present(2*i) = 1;
present(2*i+1) = 1;
}
else
{
shape(2*i) = 0;
shape(2*i+1) = 0;
present(2*i) = 0;
present(2*i+1) = 0;
}
}
}
template<typename feature_type>
struct training_sample
{
/*!
CONVENTION
- feature_pixel_values.size() == get_feature_pool_size()
- feature_pixel_values[j] == the value of the j-th feature pool
pixel when you look it up relative to the shape in current_shape.
- target_shape == The truth shape. Stays constant during the whole
training process (except for the parts that are not present, those are
always equal to the current_shape values).
- present == 0/1 mask saying which parts of target_shape are present.
- rect == the position of the object in the image_idx-th image. All shape
coordinates are coded relative to this rectangle.
- diff_shape == temporary value for holding difference between current
shape and target shape
!*/
unsigned long image_idx;
rectangle rect;
matrix<float,0,1> target_shape;
matrix<float,0,1> present;
matrix<float,0,1> current_shape;
matrix<float,0,1> diff_shape;
std::vector<feature_type> feature_pixel_values;
void swap(training_sample& item)
{
std::swap(image_idx, item.image_idx);
std::swap(rect, item.rect);
target_shape.swap(item.target_shape);
present.swap(item.present);
current_shape.swap(item.current_shape);
diff_shape.swap(item.diff_shape);
feature_pixel_values.swap(item.feature_pixel_values);
}
};
template<typename feature_type>
impl::regression_tree make_regression_tree (
thread_pool& tp,
std::vector<training_sample<feature_type>>& samples,
const std::vector<dlib::vector<float,2> >& pixel_coordinates
) const
{
using namespace impl;
std::deque<std::pair<unsigned long, unsigned long> > parts;
parts.push_back(std::make_pair(0, (unsigned long)samples.size()));
impl::regression_tree tree;
// walk the tree in breadth first order
const unsigned long num_split_nodes = static_cast<unsigned long>(std::pow(2.0, (double)get_tree_depth())-1);
std::vector<matrix<float,0,1> > sums(num_split_nodes*2+1);
if (tp.num_threads_in_pool() > 1)
{
// Here we need to calculate shape differences and store sum of differences into sums[0]
// to make it. I am splitting samples into blocks, each block will be processed by
// separate thread, and the sum of differences of each block is stored into separate
// place in block_sums
const unsigned long num_workers = std::max(1UL, tp.num_threads_in_pool());
const unsigned long num = samples.size();
const unsigned long block_size = std::max(1UL, (num + num_workers - 1) / num_workers);
std::vector<matrix<float,0,1> > block_sums(num_workers);
parallel_for(tp, 0, num_workers, [&](unsigned long block)
{
const unsigned long block_begin = block * block_size;
const unsigned long block_end = std::min(num, block_begin + block_size);
for (unsigned long i = block_begin; i < block_end; ++i)
{
samples[i].diff_shape = samples[i].target_shape - samples[i].current_shape;
block_sums[block] += samples[i].diff_shape;
}
}, 1);
// now calculate the total result from separate blocks
for (unsigned long i = 0; i < block_sums.size(); ++i)
sums[0] += block_sums[i];
}
else
{
// synchronous implementation
for (unsigned long i = 0; i < samples.size(); ++i)
{
samples[i].diff_shape = samples[i].target_shape - samples[i].current_shape;
sums[0] += samples[i].diff_shape;
}
}
for (unsigned long i = 0; i < num_split_nodes; ++i)
{
std::pair<unsigned long,unsigned long> range = parts.front();
parts.pop_front();
const impl::split_feature split = generate_split(tp, samples, range.first,
range.second, pixel_coordinates, sums[i], sums[left_child(i)],
sums[right_child(i)]);
tree.splits.push_back(split);
const unsigned long mid = partition_samples(split, samples, range.first, range.second);
parts.push_back(std::make_pair(range.first, mid));
parts.push_back(std::make_pair(mid, range.second));
}
// Now all the parts contain the ranges for the leaves so we can use them to
// compute the average leaf values.
matrix<float,0,1> present_counts(samples[0].target_shape.size());
tree.leaf_values.resize(parts.size());
for (unsigned long i = 0; i < parts.size(); ++i)
{
// Get the present counts for each dimension so we can divide each
// dimension by the number of observations we have on it to find the mean
// displacement in each leaf.
present_counts = 0;
for (unsigned long j = parts[i].first; j < parts[i].second; ++j)
present_counts += samples[j].present;
present_counts = dlib::reciprocal(present_counts);
if (parts[i].second != parts[i].first)
tree.leaf_values[i] = pointwise_multiply(present_counts,sums[num_split_nodes+i]*get_nu());
else
tree.leaf_values[i] = zeros_matrix(samples[0].target_shape);
// now adjust the current shape based on these predictions
parallel_for(tp, parts[i].first, parts[i].second, [&](unsigned long j)
{
samples[j].current_shape += tree.leaf_values[i];
// For parts that aren't present in the training data, we just make
// sure that the target shape always matches and therefore gives zero
// error. So this makes the algorithm simply ignore non-present
// landmarks.
for (long k = 0; k < samples[j].present.size(); ++k)
{
// if this part is not present
if (samples[j].present(k) == 0)
samples[j].target_shape(k) = samples[j].current_shape(k);
}
}, 1);
}
return tree;
}
impl::split_feature randomly_generate_split_feature (
const std::vector<dlib::vector<float,2> >& pixel_coordinates
) const
{
const double lambda = get_lambda();
impl::split_feature feat;
double accept_prob;
do
{
feat.idx1 = rnd.get_random_32bit_number()%get_feature_pool_size();
feat.idx2 = rnd.get_random_32bit_number()%get_feature_pool_size();
const double dist = length(pixel_coordinates[feat.idx1]-pixel_coordinates[feat.idx2]);
accept_prob = std::exp(-dist/lambda);
}
while(feat.idx1 == feat.idx2 || !(accept_prob > rnd.get_random_double()));
feat.thresh = (rnd.get_random_double()*256 - 128)/2.0;
return feat;
}
template<typename feature_type>
impl::split_feature generate_split (
thread_pool& tp,
const std::vector<training_sample<feature_type>>& samples,
unsigned long begin,
unsigned long end,
const std::vector<dlib::vector<float,2> >& pixel_coordinates,
const matrix<float,0,1>& sum,
matrix<float,0,1>& left_sum,
matrix<float,0,1>& right_sum
) const
{
// generate a bunch of random splits and test them and return the best one.
const unsigned long num_test_splits = get_num_test_splits();
// sample the random features we test in this function
std::vector<impl::split_feature> feats;
feats.reserve(num_test_splits);
for (unsigned long i = 0; i < num_test_splits; ++i)
feats.push_back(randomly_generate_split_feature(pixel_coordinates));
std::vector<matrix<float,0,1> > left_sums(num_test_splits);
std::vector<unsigned long> left_cnt(num_test_splits);
const unsigned long num_workers = std::max(1UL, tp.num_threads_in_pool());
const unsigned long block_size = std::max(1UL, (num_test_splits + num_workers - 1) / num_workers);
// now compute the sums of vectors that go left for each feature
parallel_for(tp, 0, num_workers, [&](unsigned long block)
{
const unsigned long block_begin = block * block_size;
const unsigned long block_end = std::min(block_begin + block_size, num_test_splits);
for (unsigned long j = begin; j < end; ++j)
{
for (unsigned long i = block_begin; i < block_end; ++i)
{
if ((float)samples[j].feature_pixel_values[feats[i].idx1] - (float)samples[j].feature_pixel_values[feats[i].idx2] > feats[i].thresh)
{
left_sums[i] += samples[j].diff_shape;
++left_cnt[i];
}
}
}
}, 1);
// now figure out which feature is the best
double best_score = -1;
unsigned long best_feat = 0;
matrix<float,0,1> temp;
for (unsigned long i = 0; i < num_test_splits; ++i)
{
// check how well the feature splits the space.
double score = 0;
unsigned long right_cnt = end-begin-left_cnt[i];
if (left_cnt[i] != 0 && right_cnt != 0)
{
temp = sum - left_sums[i];
score = dot(left_sums[i],left_sums[i])/left_cnt[i] + dot(temp,temp)/right_cnt;
if (score > best_score)
{
best_score = score;
best_feat = i;
}
}
}
left_sums[best_feat].swap(left_sum);
if (left_sum.size() != 0)
{
right_sum = sum - left_sum;
}
else
{
right_sum = sum;
left_sum = zeros_matrix(sum);
}
return feats[best_feat];
}
template<typename feature_type>
unsigned long partition_samples (
const impl::split_feature& split,
std::vector<training_sample<feature_type>>& samples,
unsigned long begin,
unsigned long end
) const
{
// splits samples based on split (sorta like in quick sort) and returns the mid
// point. make sure you return the mid in a way compatible with how we walk
// through the tree.
unsigned long i = begin;
for (unsigned long j = begin; j < end; ++j)
{
if ((float)samples[j].feature_pixel_values[split.idx1] - (float)samples[j].feature_pixel_values[split.idx2] > split.thresh)
{
samples[i].swap(samples[j]);
++i;
}
}
return i;
}
template<typename feature_type>
matrix<float,0,1> populate_training_sample_shapes(
const std::vector<std::vector<full_object_detection> >& objects,
std::vector<training_sample<feature_type>>& samples
) const
{
samples.clear();
matrix<float,0,1> mean_shape;
matrix<float,0,1> count;
// first fill out the target shapes
for (unsigned long i = 0; i < objects.size(); ++i)
{
for (unsigned long j = 0; j < objects[i].size(); ++j)
{
training_sample<feature_type> sample;
sample.image_idx = i;
sample.rect = objects[i][j].get_rect();
object_to_shape(objects[i][j], sample.target_shape, sample.present);
for (unsigned long itr = 0; itr < get_oversampling_amount(); ++itr)
samples.push_back(sample);
mean_shape += sample.target_shape;
count += sample.present;
}
}
mean_shape = pointwise_multiply(mean_shape,reciprocal(count));
// now go pick random initial shapes
for (unsigned long i = 0; i < samples.size(); ++i)
{
if ((i%get_oversampling_amount()) == 0)
{
// The mean shape is what we really use as an initial shape so always
// include it in the training set as an example starting shape.
samples[i].current_shape = mean_shape;
}
else
{
samples[i].current_shape.set_size(0);
matrix<float,0,1> hits(mean_shape.size());
hits = 0;
int iter = 0;
// Pick a few samples at random and randomly average them together to
// make the initial shape. Note that we make sure we get at least one
// observation (i.e. non-OBJECT_PART_NOT_PRESENT) on each part
// location.
while(min(hits) == 0 || iter < 2)
{
++iter;
const unsigned long rand_idx = rnd.get_random_32bit_number()%samples.size();
const double alpha = rnd.get_random_double()+0.1;
samples[i].current_shape += alpha*samples[rand_idx].target_shape;
hits += alpha*samples[rand_idx].present;
}
samples[i].current_shape = pointwise_multiply(samples[i].current_shape, reciprocal(hits));
}
}
for (unsigned long i = 0; i < samples.size(); ++i)
{
for (long k = 0; k < samples[i].present.size(); ++k)
{
// if this part is not present
if (samples[i].present(k) == 0)
samples[i].target_shape(k) = samples[i].current_shape(k);
}
}
return mean_shape;
}
void randomly_sample_pixel_coordinates (
std::vector<dlib::vector<float,2> >& pixel_coordinates,
const double min_x,
const double min_y,
const double max_x,
const double max_y
) const
/*!
ensures
- #pixel_coordinates.size() == get_feature_pool_size()
- for all valid i:
- pixel_coordinates[i] == a point in the box defined by the min/max x/y arguments.
!*/
{
pixel_coordinates.resize(get_feature_pool_size());
for (unsigned long i = 0; i < get_feature_pool_size(); ++i)
{
pixel_coordinates[i].x() = rnd.get_random_double()*(max_x-min_x) + min_x;
pixel_coordinates[i].y() = rnd.get_random_double()*(max_y-min_y) + min_y;
}
}
std::vector<std::vector<dlib::vector<float,2> > > randomly_sample_pixel_coordinates (
const matrix<float,0,1>& initial_shape
) const
{
const double padding = get_feature_pool_region_padding();
// Figure out the bounds on the object shapes. We will sample uniformly
// from this box.
matrix<float> temp = reshape(initial_shape, initial_shape.size()/2, 2);
double min_x = min(colm(temp,0))-padding;
double min_y = min(colm(temp,1))-padding;
double max_x = max(colm(temp,0))+padding;
double max_y = max(colm(temp,1))+padding;
if (get_padding_mode() == bounding_box_relative)
{
min_x = std::min(0.0, min_x);
min_y = std::min(0.0, min_y);
max_x = std::max(1.0, max_x);
max_y = std::max(1.0, max_y);
}
min_x -= padding;
min_y -= padding;
max_x += padding;
max_y += padding;
std::vector<std::vector<dlib::vector<float,2> > > pixel_coordinates;
pixel_coordinates.resize(get_cascade_depth());
for (unsigned long i = 0; i < get_cascade_depth(); ++i)
randomly_sample_pixel_coordinates(pixel_coordinates[i], min_x, min_y, max_x, max_y);
return pixel_coordinates;
}
mutable dlib::rand rnd;
unsigned long _cascade_depth;
unsigned long _tree_depth;
unsigned long _num_trees_per_cascade_level;
double _nu;
unsigned long _oversampling_amount;
unsigned long _feature_pool_size;
double _lambda;
unsigned long _num_test_splits;
double _feature_pool_region_padding;
bool _verbose;
unsigned long _num_threads;
padding_mode_t _padding_mode;
};
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_SHAPE_PREDICToR_TRAINER_H_