Skip to content

GitLab

Commit bbeac285 authored by Evgeniy Fominov's avatar Evgeniy Fominov Committed by Davis E. King
Browse files

Shape predictor trainer optimizations (#126) 

* Shape predictor trainer optimizations

* Fixed performance leak in single thread mode & made VS2010 support
parent 6eb5bd80
Show whitespace changes
Inline Side-by-side
Showing with 133 additions and 40 deletions
dlib/image_processing/shape_predictor.h
View file @ bbeac285
......@@ -80,7 +80,7 @@ namespace dlib
i = 0;
while (i < splits.size())
{
if (feature_pixel_values[splits[i].idx1] - feature_pixel_values[splits[i].idx2] > splits[i].thresh)
if ((float)feature_pixel_values[splits[i].idx1] - (float)feature_pixel_values[splits[i].idx2] > splits[i].thresh)
i = left_child(i);
else
i = right_child(i);
......@@ -235,7 +235,7 @@ namespace dlib
// ------------------------------------------------------------------------------------
template <typename image_type>
template <typename image_type, typename feature_type>
void extract_feature_pixel_values (
const image_type& img_,
const rectangle& rect,
......@@ -243,7 +243,7 @@ namespace dlib
const matrix<float,0,1>& reference_shape,
const std::vector<unsigned long>& reference_pixel_anchor_idx,
const std::vector<dlib::vector<float,2> >& reference_pixel_deltas,
std::vector<float>& feature_pixel_values
std::vector<feature_type>& feature_pixel_values
)
/*!
requires
......@@ -453,6 +453,7 @@ namespace dlib
_num_test_splits = 20;
_feature_pool_region_padding = 0;
_verbose = false;
_num_threads = 0;
}
unsigned long get_cascade_depth (
......@@ -605,6 +606,15 @@ namespace dlib
_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,
......@@ -661,13 +671,17 @@ namespace dlib
<< "\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> samples;
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);
......@@ -688,17 +702,17 @@ namespace dlib
// First compute the feature_pixel_values for each training sample at this
// level of the cascade.
for (unsigned long i = 0; i < samples.size(); ++i)
parallel_for(tp, 0, samples.size(), [&](unsigned long i)
{
extract_feature_pixel_values(images[samples[i].image_idx], samples[i].rect,
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(samples, pixel_coordinates[cascade]));
forests[cascade].push_back(make_regression_tree(tp, samples, pixel_coordinates[cascade]));
if (_verbose)
{
......@@ -745,6 +759,7 @@ namespace dlib
}
}
template<typename feature_type>
struct training_sample
{
/*!
......@@ -760,6 +775,8 @@ namespace dlib
- 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;
......@@ -768,7 +785,8 @@ namespace dlib
matrix<float,0,1> present;
matrix<float,0,1> current_shape;
std::vector<float> feature_pixel_values;
matrix<float,0,1> diff_shape;
std::vector<feature_type> feature_pixel_values;
void swap(training_sample& item)
{
......@@ -777,12 +795,15 @@ namespace dlib
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 (
std::vector<training_sample>& samples,
thread_pool& tp,
std::vector<training_sample<feature_type>>& samples,
const std::vector<dlib::vector<float,2> >& pixel_coordinates
) const
{
......@@ -795,15 +816,49 @@ namespace dlib
// 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 of 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)
sums[0] += samples[i].target_shape - samples[i].current_shape;
{
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(samples, range.first,
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);
......@@ -833,7 +888,7 @@ namespace dlib
tree.leaf_values[i] = zeros_matrix(samples[0].target_shape);
// now adjust the current shape based on these predictions
for (unsigned long j = parts[i].first; j < parts[i].second; ++j)
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
......@@ -846,7 +901,7 @@ namespace dlib
if (samples[j].present(k) == 0)
samples[j].target_shape(k) = samples[j].current_shape(k);
}
}
}, 1);
}
return tree;
......@@ -873,8 +928,10 @@ namespace dlib
return feat;
}
template<typename feature_type>
impl::split_feature generate_split (
const std::vector<training_sample>& samples,
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,
......@@ -896,24 +953,33 @@ namespace dlib
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
matrix<float,0,1> temp;
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)
{
temp = samples[j].target_shape-samples[j].current_shape;
for (unsigned long i = 0; i < num_test_splits; ++i)
for (unsigned long i = block_begin; i < block_end; ++i)
{
if (samples[j].feature_pixel_values[feats[i].idx1] - samples[j].feature_pixel_values[feats[i].idx2] > feats[i].thresh)
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] += temp;
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.
......@@ -944,9 +1010,10 @@ namespace dlib
return feats[best_feat];
}
template<typename feature_type>
unsigned long partition_samples (
const impl::split_feature& split,
std::vector<training_sample>& samples,
std::vector<training_sample<feature_type>>& samples,
unsigned long begin,
unsigned long end
) const
......@@ -958,7 +1025,7 @@ namespace dlib
unsigned long i = begin;
for (unsigned long j = begin; j < end; ++j)
{
if (samples[j].feature_pixel_values[split.idx1] - samples[j].feature_pixel_values[split.idx2] > split.thresh)
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;
......@@ -969,9 +1036,10 @@ namespace dlib
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>& samples
std::vector<training_sample<feature_type>>& samples
) const
{
samples.clear();
......@@ -982,7 +1050,7 @@ namespace dlib
{
for (unsigned long j = 0; j < objects[i].size(); ++j)
{
training_sample sample;
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);
......@@ -1099,6 +1167,7 @@ namespace dlib
unsigned long _num_test_splits;
double _feature_pool_region_padding;
bool _verbose;
unsigned long _num_threads;
};
// ----------------------------------------------------------------------------------------
......
dlib/image_processing/shape_predictor_abstract.h
View file @ bbeac285
......@@ -148,6 +148,7 @@ namespace dlib
- #get_num_test_splits() == 20
- #get_feature_pool_region_padding() == 0
- #get_random_seed() == ""
- #get_num_threads() == 0
- This object will not be verbose
!*/
......@@ -367,6 +368,26 @@ namespace dlib
- #get_num_test_splits() == num
!*/
unsigned long get_num_threads (
) const;
/*!
ensures
- When running training process, it is possible to make some parts of it parallel
using CPU threads with #parallel_for() extension and creating #thread_pool internally
When get_num_threads() == 0, trainer will not create threads and all processing will
be done in the calling thread
!*/
void set_num_threads (
unsigned long num
);
/*!
requires
- num >= 0
ensures
- #get_num_threads() == num
!*/
void be_verbose (
);
/*!
......
examples/train_shape_predictor_ex.cpp
View file @ bbeac285
......@@ -108,6 +108,9 @@ int main(int argc, char** argv)
trainer.set_nu(0.05);
trainer.set_tree_depth(2);
// some parts of training process can be parellelized.
// Trainer will use this count of threads when possible
trainer.set_num_threads(2);
// Tell the trainer to print status messages to the console so we can
// see how long the training will take.
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment