Instance segmentation (#1918)
* Add instance segmentation example - first version of training code
* Add MMOD options; get rid of the cache approach, and instead load all MMOD rects upfront
* Improve console output
* Set filter count
* Minor tweaking
* Inference - first version, at least compiles!
* Ignore overlapped boxes
* Ignore even small instances
* Set overlaps_ignore
* Add TODO remarks
* Revert "Set overlaps_ignore"
This reverts commit 65adeff1f89af62b10c691e7aa86c04fc358d03e.
* Set result size
* Set label image size
* Take ignore-color into account
* Fix the cropping rect's aspect ratio; also slightly expand the rect
* Draw the largest findings last
* Improve masking of the current instance
* Add some perturbation to the inputs
* Simplify ground-truth reading; fix random cropping
* Read even class labels
* Tweak default minibatch size
* Learn only one class
* Really train only instances of the selected class
* Remove outdated TODO remark
* Automatically skip images with no detections
* Print to console what was found
* Fix class index problem
* Fix indentation
* Allow to choose multiple classes
* Draw rect in the color of the corresponding class
* Write detector window classes to ostream; also group detection windows by class (when ostreaming)
* Train a separate instance segmentation network for each classlabel
* Use separate synchronization file for each seg net of each class
* Allow more overlap
* Fix sorting criterion
* Fix interpolating the predicted mask
* Improve bilinear interpolation: if output type is an integer, round instead of truncating
* Add helpful comments
* Ignore large aspect ratios; refactor the code; tweak some network parameters
* Simplify the segmentation network structure; make the object detection network more complex in turn
* Problem: CUDA errors not reported properly to console
Solution: stop and join data loader threads even in case of exceptions
* Minor parameters tweaking
* Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value
* Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing
* Improve behaviour when loss actually increased after disk sync
* Revert some of the earlier change
* Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync
* Revert "Revert some of the earlier change"
This reverts commit 6c852124efe6473a5c962de0091709129d6fcde3.
* Keep enough previous loss values, until the disk sync
* Fix maintaining the dumped (now "effectively disregarded") loss values count
* Detect cats instead of aeroplanes
* Add helpful logging
* Clarify the intention and the code
* Review fixes
* Add operator== for the other pixel types as well; remove the inline
* If available, use constexpr if
* Revert "If available, use constexpr if"
This reverts commit 503d4dd3355ff8ad613116e3ffcc0fa664674f69.
* Simplify code as per review comments
* Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh
* Clarify console output
* Revert "Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh"
This reverts commit 9191ebc7762d17d81cdfc334a80ca9a667365740.
* To keep the changes to a bare minimum, revert the steps_since_last_learning_rate_shrink change after all (at least for now)
* Even empty out some of the previous test loss values
* Minor review fixes
* Can't use C++14 features here
* Do not use the struct name as a variable name
2019-11-15 11:53:16 +08:00
|
|
|
// The contents of this file are in the public domain. See LICENSE_FOR_EXAMPLE_PROGRAMS.txt
|
|
|
|
/*
|
|
|
|
This example shows how to train a instance segmentation net using the PASCAL VOC2012
|
|
|
|
dataset. For an introduction to what segmentation is, see the accompanying header file
|
|
|
|
dnn_instance_segmentation_ex.h.
|
|
|
|
|
|
|
|
Instructions how to run the example:
|
|
|
|
1. Download the PASCAL VOC2012 data, and untar it somewhere.
|
|
|
|
http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
|
|
|
|
2. Build the dnn_instance_segmentation_train_ex example program.
|
|
|
|
3. Run:
|
|
|
|
./dnn_instance_segmentation_train_ex /path/to/VOC2012
|
|
|
|
4. Wait while the network is being trained.
|
|
|
|
5. Build the dnn_instance_segmentation_ex example program.
|
|
|
|
6. Run:
|
|
|
|
./dnn_instance_segmentation_ex /path/to/VOC2012-or-other-images
|
|
|
|
|
|
|
|
It would be a good idea to become familiar with dlib's DNN tooling before reading this
|
|
|
|
example. So you should read dnn_introduction_ex.cpp, dnn_introduction2_ex.cpp,
|
|
|
|
and dnn_semantic_segmentation_train_ex.cpp before reading this example program.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "dnn_instance_segmentation_ex.h"
|
|
|
|
#include "pascal_voc_2012.h"
|
|
|
|
|
|
|
|
#include <iostream>
|
|
|
|
#include <dlib/data_io.h>
|
|
|
|
#include <dlib/image_transforms.h>
|
|
|
|
#include <dlib/dir_nav.h>
|
|
|
|
#include <iterator>
|
|
|
|
#include <thread>
|
|
|
|
#if __cplusplus >= 201703L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
|
|
|
|
#include <execution>
|
|
|
|
#endif // __cplusplus >= 201703L
|
|
|
|
|
|
|
|
using namespace std;
|
|
|
|
using namespace dlib;
|
|
|
|
|
|
|
|
// ----------------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
// A single training sample for detection. A mini-batch comprises many of these.
|
|
|
|
struct det_training_sample
|
|
|
|
{
|
|
|
|
matrix<rgb_pixel> input_image;
|
|
|
|
std::vector<dlib::mmod_rect> mmod_rects;
|
|
|
|
};
|
|
|
|
|
|
|
|
// A single training sample for segmentation. A mini-batch comprises many of these.
|
|
|
|
struct seg_training_sample
|
|
|
|
{
|
|
|
|
matrix<rgb_pixel> input_image;
|
2020-01-20 20:47:47 +08:00
|
|
|
matrix<float> label_image; // The ground-truth label of each pixel. (+1 or -1)
|
Instance segmentation (#1918)
* Add instance segmentation example - first version of training code
* Add MMOD options; get rid of the cache approach, and instead load all MMOD rects upfront
* Improve console output
* Set filter count
* Minor tweaking
* Inference - first version, at least compiles!
* Ignore overlapped boxes
* Ignore even small instances
* Set overlaps_ignore
* Add TODO remarks
* Revert "Set overlaps_ignore"
This reverts commit 65adeff1f89af62b10c691e7aa86c04fc358d03e.
* Set result size
* Set label image size
* Take ignore-color into account
* Fix the cropping rect's aspect ratio; also slightly expand the rect
* Draw the largest findings last
* Improve masking of the current instance
* Add some perturbation to the inputs
* Simplify ground-truth reading; fix random cropping
* Read even class labels
* Tweak default minibatch size
* Learn only one class
* Really train only instances of the selected class
* Remove outdated TODO remark
* Automatically skip images with no detections
* Print to console what was found
* Fix class index problem
* Fix indentation
* Allow to choose multiple classes
* Draw rect in the color of the corresponding class
* Write detector window classes to ostream; also group detection windows by class (when ostreaming)
* Train a separate instance segmentation network for each classlabel
* Use separate synchronization file for each seg net of each class
* Allow more overlap
* Fix sorting criterion
* Fix interpolating the predicted mask
* Improve bilinear interpolation: if output type is an integer, round instead of truncating
* Add helpful comments
* Ignore large aspect ratios; refactor the code; tweak some network parameters
* Simplify the segmentation network structure; make the object detection network more complex in turn
* Problem: CUDA errors not reported properly to console
Solution: stop and join data loader threads even in case of exceptions
* Minor parameters tweaking
* Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value
* Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing
* Improve behaviour when loss actually increased after disk sync
* Revert some of the earlier change
* Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync
* Revert "Revert some of the earlier change"
This reverts commit 6c852124efe6473a5c962de0091709129d6fcde3.
* Keep enough previous loss values, until the disk sync
* Fix maintaining the dumped (now "effectively disregarded") loss values count
* Detect cats instead of aeroplanes
* Add helpful logging
* Clarify the intention and the code
* Review fixes
* Add operator== for the other pixel types as well; remove the inline
* If available, use constexpr if
* Revert "If available, use constexpr if"
This reverts commit 503d4dd3355ff8ad613116e3ffcc0fa664674f69.
* Simplify code as per review comments
* Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh
* Clarify console output
* Revert "Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh"
This reverts commit 9191ebc7762d17d81cdfc334a80ca9a667365740.
* To keep the changes to a bare minimum, revert the steps_since_last_learning_rate_shrink change after all (at least for now)
* Even empty out some of the previous test loss values
* Minor review fixes
* Can't use C++14 features here
* Do not use the struct name as a variable name
2019-11-15 11:53:16 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
// ----------------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
bool is_instance_pixel(const dlib::rgb_pixel& rgb_label)
|
|
|
|
{
|
|
|
|
if (rgb_label == dlib::rgb_pixel(0, 0, 0))
|
|
|
|
return false; // Background
|
|
|
|
if (rgb_label == dlib::rgb_pixel(224, 224, 192))
|
|
|
|
return false; // The cream-colored `void' label is used in border regions and to mask difficult objects
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Provide hash function for dlib::rgb_pixel
|
|
|
|
namespace std {
|
|
|
|
template <>
|
|
|
|
struct hash<dlib::rgb_pixel>
|
|
|
|
{
|
|
|
|
std::size_t operator()(const dlib::rgb_pixel& p) const
|
|
|
|
{
|
|
|
|
return (static_cast<uint32_t>(p.red) << 16)
|
|
|
|
| (static_cast<uint32_t>(p.green) << 8)
|
|
|
|
| (static_cast<uint32_t>(p.blue));
|
|
|
|
}
|
|
|
|
};
|
|
|
|
}
|
|
|
|
|
|
|
|
struct truth_instance
|
|
|
|
{
|
|
|
|
dlib::rgb_pixel rgb_label;
|
|
|
|
dlib::mmod_rect mmod_rect;
|
|
|
|
};
|
|
|
|
|
|
|
|
std::vector<truth_instance> rgb_label_images_to_truth_instances(
|
|
|
|
const dlib::matrix<dlib::rgb_pixel>& instance_label_image,
|
|
|
|
const dlib::matrix<dlib::rgb_pixel>& class_label_image
|
|
|
|
)
|
|
|
|
{
|
|
|
|
std::unordered_map<dlib::rgb_pixel, mmod_rect> result_map;
|
|
|
|
|
|
|
|
DLIB_CASSERT(instance_label_image.nr() == class_label_image.nr());
|
|
|
|
DLIB_CASSERT(instance_label_image.nc() == class_label_image.nc());
|
|
|
|
|
|
|
|
const auto nr = instance_label_image.nr();
|
|
|
|
const auto nc = instance_label_image.nc();
|
|
|
|
|
|
|
|
for (int r = 0; r < nr; ++r)
|
|
|
|
{
|
|
|
|
for (int c = 0; c < nc; ++c)
|
|
|
|
{
|
|
|
|
const auto rgb_instance_label = instance_label_image(r, c);
|
|
|
|
|
|
|
|
if (!is_instance_pixel(rgb_instance_label))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
const auto rgb_class_label = class_label_image(r, c);
|
|
|
|
const Voc2012class& voc2012_class = find_voc2012_class(rgb_class_label);
|
|
|
|
|
|
|
|
const auto i = result_map.find(rgb_instance_label);
|
|
|
|
if (i == result_map.end())
|
|
|
|
{
|
|
|
|
// Encountered a new instance
|
|
|
|
result_map[rgb_instance_label] = rectangle(c, r, c, r);
|
|
|
|
result_map[rgb_instance_label].label = voc2012_class.classlabel;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
// Not the first occurrence - update the rect
|
|
|
|
auto& rect = i->second.rect;
|
|
|
|
|
|
|
|
if (c < rect.left())
|
|
|
|
rect.set_left(c);
|
|
|
|
else if (c > rect.right())
|
|
|
|
rect.set_right(c);
|
|
|
|
|
|
|
|
if (r > rect.bottom())
|
|
|
|
rect.set_bottom(r);
|
|
|
|
|
|
|
|
DLIB_CASSERT(i->second.label == voc2012_class.classlabel);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<truth_instance> flat_result;
|
|
|
|
flat_result.reserve(result_map.size());
|
|
|
|
|
|
|
|
for (const auto& i : result_map) {
|
|
|
|
flat_result.push_back(truth_instance{
|
|
|
|
i.first, i.second
|
|
|
|
});
|
|
|
|
}
|
|
|
|
|
|
|
|
return flat_result;
|
|
|
|
}
|
|
|
|
|
|
|
|
// ----------------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
struct truth_image
|
|
|
|
{
|
|
|
|
image_info info;
|
|
|
|
std::vector<truth_instance> truth_instances;
|
|
|
|
};
|
|
|
|
|
|
|
|
std::vector<mmod_rect> extract_mmod_rects(
|
|
|
|
const std::vector<truth_instance>& truth_instances
|
|
|
|
)
|
|
|
|
{
|
|
|
|
std::vector<mmod_rect> mmod_rects(truth_instances.size());
|
|
|
|
|
|
|
|
std::transform(
|
|
|
|
truth_instances.begin(),
|
|
|
|
truth_instances.end(),
|
|
|
|
mmod_rects.begin(),
|
|
|
|
[](const truth_instance& truth) { return truth.mmod_rect; }
|
|
|
|
);
|
|
|
|
|
|
|
|
return mmod_rects;
|
2020-04-01 07:35:23 +08:00
|
|
|
}
|
Instance segmentation (#1918)
* Add instance segmentation example - first version of training code
* Add MMOD options; get rid of the cache approach, and instead load all MMOD rects upfront
* Improve console output
* Set filter count
* Minor tweaking
* Inference - first version, at least compiles!
* Ignore overlapped boxes
* Ignore even small instances
* Set overlaps_ignore
* Add TODO remarks
* Revert "Set overlaps_ignore"
This reverts commit 65adeff1f89af62b10c691e7aa86c04fc358d03e.
* Set result size
* Set label image size
* Take ignore-color into account
* Fix the cropping rect's aspect ratio; also slightly expand the rect
* Draw the largest findings last
* Improve masking of the current instance
* Add some perturbation to the inputs
* Simplify ground-truth reading; fix random cropping
* Read even class labels
* Tweak default minibatch size
* Learn only one class
* Really train only instances of the selected class
* Remove outdated TODO remark
* Automatically skip images with no detections
* Print to console what was found
* Fix class index problem
* Fix indentation
* Allow to choose multiple classes
* Draw rect in the color of the corresponding class
* Write detector window classes to ostream; also group detection windows by class (when ostreaming)
* Train a separate instance segmentation network for each classlabel
* Use separate synchronization file for each seg net of each class
* Allow more overlap
* Fix sorting criterion
* Fix interpolating the predicted mask
* Improve bilinear interpolation: if output type is an integer, round instead of truncating
* Add helpful comments
* Ignore large aspect ratios; refactor the code; tweak some network parameters
* Simplify the segmentation network structure; make the object detection network more complex in turn
* Problem: CUDA errors not reported properly to console
Solution: stop and join data loader threads even in case of exceptions
* Minor parameters tweaking
* Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value
* Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing
* Improve behaviour when loss actually increased after disk sync
* Revert some of the earlier change
* Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync
* Revert "Revert some of the earlier change"
This reverts commit 6c852124efe6473a5c962de0091709129d6fcde3.
* Keep enough previous loss values, until the disk sync
* Fix maintaining the dumped (now "effectively disregarded") loss values count
* Detect cats instead of aeroplanes
* Add helpful logging
* Clarify the intention and the code
* Review fixes
* Add operator== for the other pixel types as well; remove the inline
* If available, use constexpr if
* Revert "If available, use constexpr if"
This reverts commit 503d4dd3355ff8ad613116e3ffcc0fa664674f69.
* Simplify code as per review comments
* Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh
* Clarify console output
* Revert "Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh"
This reverts commit 9191ebc7762d17d81cdfc334a80ca9a667365740.
* To keep the changes to a bare minimum, revert the steps_since_last_learning_rate_shrink change after all (at least for now)
* Even empty out some of the previous test loss values
* Minor review fixes
* Can't use C++14 features here
* Do not use the struct name as a variable name
2019-11-15 11:53:16 +08:00
|
|
|
|
|
|
|
std::vector<std::vector<mmod_rect>> extract_mmod_rect_vectors(
|
|
|
|
const std::vector<truth_image>& truth_images
|
|
|
|
)
|
|
|
|
{
|
|
|
|
std::vector<std::vector<mmod_rect>> mmod_rects(truth_images.size());
|
|
|
|
|
|
|
|
const auto extract_mmod_rects_from_truth_image = [](const truth_image& truth_image)
|
|
|
|
{
|
|
|
|
return extract_mmod_rects(truth_image.truth_instances);
|
|
|
|
};
|
|
|
|
|
|
|
|
std::transform(
|
|
|
|
truth_images.begin(),
|
|
|
|
truth_images.end(),
|
|
|
|
mmod_rects.begin(),
|
|
|
|
extract_mmod_rects_from_truth_image
|
|
|
|
);
|
|
|
|
|
|
|
|
return mmod_rects;
|
|
|
|
}
|
|
|
|
|
|
|
|
det_bnet_type train_detection_network(
|
|
|
|
const std::vector<truth_image>& truth_images,
|
|
|
|
unsigned int det_minibatch_size
|
|
|
|
)
|
|
|
|
{
|
|
|
|
const double initial_learning_rate = 0.1;
|
|
|
|
const double weight_decay = 0.0001;
|
|
|
|
const double momentum = 0.9;
|
|
|
|
const double min_detector_window_overlap_iou = 0.65;
|
|
|
|
|
|
|
|
const int target_size = 70;
|
|
|
|
const int min_target_size = 30;
|
|
|
|
|
|
|
|
mmod_options options(
|
|
|
|
extract_mmod_rect_vectors(truth_images),
|
|
|
|
target_size, min_target_size,
|
|
|
|
min_detector_window_overlap_iou
|
|
|
|
);
|
|
|
|
|
|
|
|
options.overlaps_ignore = test_box_overlap(0.5, 0.9);
|
|
|
|
|
|
|
|
det_bnet_type det_net(options);
|
|
|
|
|
|
|
|
det_net.subnet().layer_details().set_num_filters(options.detector_windows.size());
|
|
|
|
|
|
|
|
dlib::pipe<det_training_sample> data(200);
|
|
|
|
auto f = [&data, &truth_images, target_size, min_target_size](time_t seed)
|
|
|
|
{
|
|
|
|
dlib::rand rnd(time(0) + seed);
|
|
|
|
matrix<rgb_pixel> input_image;
|
|
|
|
|
|
|
|
random_cropper cropper;
|
|
|
|
cropper.set_seed(time(0));
|
|
|
|
cropper.set_chip_dims(350, 350);
|
|
|
|
|
|
|
|
// Usually you want to give the cropper whatever min sizes you passed to the
|
|
|
|
// mmod_options constructor, or very slightly smaller sizes, which is what we do here.
|
|
|
|
cropper.set_min_object_size(target_size - 2, min_target_size - 2);
|
|
|
|
cropper.set_max_rotation_degrees(2);
|
|
|
|
|
|
|
|
det_training_sample temp;
|
|
|
|
|
|
|
|
while (data.is_enabled())
|
|
|
|
{
|
|
|
|
// Pick a random input image.
|
|
|
|
const auto random_index = rnd.get_random_32bit_number() % truth_images.size();
|
|
|
|
const auto& truth_image = truth_images[random_index];
|
|
|
|
|
|
|
|
// Load the input image.
|
|
|
|
load_image(input_image, truth_image.info.image_filename);
|
|
|
|
|
|
|
|
// Get a random crop of the input.
|
|
|
|
const auto mmod_rects = extract_mmod_rects(truth_image.truth_instances);
|
|
|
|
cropper(input_image, mmod_rects, temp.input_image, temp.mmod_rects);
|
|
|
|
|
|
|
|
disturb_colors(temp.input_image, rnd);
|
|
|
|
|
|
|
|
// Push the result to be used by the trainer.
|
|
|
|
data.enqueue(temp);
|
|
|
|
}
|
|
|
|
};
|
|
|
|
std::thread data_loader1([f]() { f(1); });
|
|
|
|
std::thread data_loader2([f]() { f(2); });
|
|
|
|
std::thread data_loader3([f]() { f(3); });
|
|
|
|
std::thread data_loader4([f]() { f(4); });
|
|
|
|
|
|
|
|
const auto stop_data_loaders = [&]()
|
|
|
|
{
|
|
|
|
data.disable();
|
|
|
|
data_loader1.join();
|
|
|
|
data_loader2.join();
|
|
|
|
data_loader3.join();
|
|
|
|
data_loader4.join();
|
|
|
|
};
|
|
|
|
|
|
|
|
dnn_trainer<det_bnet_type> det_trainer(det_net, sgd(weight_decay, momentum));
|
|
|
|
|
|
|
|
try
|
|
|
|
{
|
|
|
|
det_trainer.be_verbose();
|
|
|
|
det_trainer.set_learning_rate(initial_learning_rate);
|
|
|
|
det_trainer.set_synchronization_file("pascal_voc2012_det_trainer_state_file.dat", std::chrono::minutes(10));
|
|
|
|
det_trainer.set_iterations_without_progress_threshold(5000);
|
|
|
|
|
|
|
|
// Output training parameters.
|
|
|
|
cout << det_trainer << endl;
|
|
|
|
|
|
|
|
std::vector<matrix<rgb_pixel>> samples;
|
|
|
|
std::vector<std::vector<mmod_rect>> labels;
|
|
|
|
|
|
|
|
// The main training loop. Keep making mini-batches and giving them to the trainer.
|
|
|
|
// We will run until the learning rate becomes small enough.
|
|
|
|
while (det_trainer.get_learning_rate() >= 1e-4)
|
|
|
|
{
|
|
|
|
samples.clear();
|
|
|
|
labels.clear();
|
|
|
|
|
|
|
|
// make a mini-batch
|
|
|
|
det_training_sample temp;
|
|
|
|
while (samples.size() < det_minibatch_size)
|
|
|
|
{
|
|
|
|
data.dequeue(temp);
|
|
|
|
|
|
|
|
samples.push_back(std::move(temp.input_image));
|
|
|
|
labels.push_back(std::move(temp.mmod_rects));
|
|
|
|
}
|
|
|
|
|
|
|
|
det_trainer.train_one_step(samples, labels);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
catch (std::exception&)
|
|
|
|
{
|
|
|
|
stop_data_loaders();
|
|
|
|
throw;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Training done, tell threads to stop and make sure to wait for them to finish before
|
|
|
|
// moving on.
|
|
|
|
stop_data_loaders();
|
|
|
|
|
|
|
|
// also wait for threaded processing to stop in the trainer.
|
|
|
|
det_trainer.get_net();
|
|
|
|
|
|
|
|
det_net.clean();
|
|
|
|
|
|
|
|
return det_net;
|
|
|
|
}
|
|
|
|
|
|
|
|
// ----------------------------------------------------------------------------------------
|
|
|
|
|
2020-01-20 20:47:47 +08:00
|
|
|
matrix<float> keep_only_current_instance(const matrix<rgb_pixel>& rgb_label_image, const rgb_pixel rgb_label)
|
Instance segmentation (#1918)
* Add instance segmentation example - first version of training code
* Add MMOD options; get rid of the cache approach, and instead load all MMOD rects upfront
* Improve console output
* Set filter count
* Minor tweaking
* Inference - first version, at least compiles!
* Ignore overlapped boxes
* Ignore even small instances
* Set overlaps_ignore
* Add TODO remarks
* Revert "Set overlaps_ignore"
This reverts commit 65adeff1f89af62b10c691e7aa86c04fc358d03e.
* Set result size
* Set label image size
* Take ignore-color into account
* Fix the cropping rect's aspect ratio; also slightly expand the rect
* Draw the largest findings last
* Improve masking of the current instance
* Add some perturbation to the inputs
* Simplify ground-truth reading; fix random cropping
* Read even class labels
* Tweak default minibatch size
* Learn only one class
* Really train only instances of the selected class
* Remove outdated TODO remark
* Automatically skip images with no detections
* Print to console what was found
* Fix class index problem
* Fix indentation
* Allow to choose multiple classes
* Draw rect in the color of the corresponding class
* Write detector window classes to ostream; also group detection windows by class (when ostreaming)
* Train a separate instance segmentation network for each classlabel
* Use separate synchronization file for each seg net of each class
* Allow more overlap
* Fix sorting criterion
* Fix interpolating the predicted mask
* Improve bilinear interpolation: if output type is an integer, round instead of truncating
* Add helpful comments
* Ignore large aspect ratios; refactor the code; tweak some network parameters
* Simplify the segmentation network structure; make the object detection network more complex in turn
* Problem: CUDA errors not reported properly to console
Solution: stop and join data loader threads even in case of exceptions
* Minor parameters tweaking
* Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value
* Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing
* Improve behaviour when loss actually increased after disk sync
* Revert some of the earlier change
* Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync
* Revert "Revert some of the earlier change"
This reverts commit 6c852124efe6473a5c962de0091709129d6fcde3.
* Keep enough previous loss values, until the disk sync
* Fix maintaining the dumped (now "effectively disregarded") loss values count
* Detect cats instead of aeroplanes
* Add helpful logging
* Clarify the intention and the code
* Review fixes
* Add operator== for the other pixel types as well; remove the inline
* If available, use constexpr if
* Revert "If available, use constexpr if"
This reverts commit 503d4dd3355ff8ad613116e3ffcc0fa664674f69.
* Simplify code as per review comments
* Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh
* Clarify console output
* Revert "Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh"
This reverts commit 9191ebc7762d17d81cdfc334a80ca9a667365740.
* To keep the changes to a bare minimum, revert the steps_since_last_learning_rate_shrink change after all (at least for now)
* Even empty out some of the previous test loss values
* Minor review fixes
* Can't use C++14 features here
* Do not use the struct name as a variable name
2019-11-15 11:53:16 +08:00
|
|
|
{
|
|
|
|
const auto nr = rgb_label_image.nr();
|
|
|
|
const auto nc = rgb_label_image.nc();
|
|
|
|
|
2020-01-20 20:47:47 +08:00
|
|
|
matrix<float> result(nr, nc);
|
Instance segmentation (#1918)
* Add instance segmentation example - first version of training code
* Add MMOD options; get rid of the cache approach, and instead load all MMOD rects upfront
* Improve console output
* Set filter count
* Minor tweaking
* Inference - first version, at least compiles!
* Ignore overlapped boxes
* Ignore even small instances
* Set overlaps_ignore
* Add TODO remarks
* Revert "Set overlaps_ignore"
This reverts commit 65adeff1f89af62b10c691e7aa86c04fc358d03e.
* Set result size
* Set label image size
* Take ignore-color into account
* Fix the cropping rect's aspect ratio; also slightly expand the rect
* Draw the largest findings last
* Improve masking of the current instance
* Add some perturbation to the inputs
* Simplify ground-truth reading; fix random cropping
* Read even class labels
* Tweak default minibatch size
* Learn only one class
* Really train only instances of the selected class
* Remove outdated TODO remark
* Automatically skip images with no detections
* Print to console what was found
* Fix class index problem
* Fix indentation
* Allow to choose multiple classes
* Draw rect in the color of the corresponding class
* Write detector window classes to ostream; also group detection windows by class (when ostreaming)
* Train a separate instance segmentation network for each classlabel
* Use separate synchronization file for each seg net of each class
* Allow more overlap
* Fix sorting criterion
* Fix interpolating the predicted mask
* Improve bilinear interpolation: if output type is an integer, round instead of truncating
* Add helpful comments
* Ignore large aspect ratios; refactor the code; tweak some network parameters
* Simplify the segmentation network structure; make the object detection network more complex in turn
* Problem: CUDA errors not reported properly to console
Solution: stop and join data loader threads even in case of exceptions
* Minor parameters tweaking
* Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value
* Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing
* Improve behaviour when loss actually increased after disk sync
* Revert some of the earlier change
* Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync
* Revert "Revert some of the earlier change"
This reverts commit 6c852124efe6473a5c962de0091709129d6fcde3.
* Keep enough previous loss values, until the disk sync
* Fix maintaining the dumped (now "effectively disregarded") loss values count
* Detect cats instead of aeroplanes
* Add helpful logging
* Clarify the intention and the code
* Review fixes
* Add operator== for the other pixel types as well; remove the inline
* If available, use constexpr if
* Revert "If available, use constexpr if"
This reverts commit 503d4dd3355ff8ad613116e3ffcc0fa664674f69.
* Simplify code as per review comments
* Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh
* Clarify console output
* Revert "Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh"
This reverts commit 9191ebc7762d17d81cdfc334a80ca9a667365740.
* To keep the changes to a bare minimum, revert the steps_since_last_learning_rate_shrink change after all (at least for now)
* Even empty out some of the previous test loss values
* Minor review fixes
* Can't use C++14 features here
* Do not use the struct name as a variable name
2019-11-15 11:53:16 +08:00
|
|
|
|
|
|
|
for (long r = 0; r < nr; ++r)
|
|
|
|
{
|
|
|
|
for (long c = 0; c < nc; ++c)
|
|
|
|
{
|
|
|
|
const auto& index = rgb_label_image(r, c);
|
|
|
|
if (index == rgb_label)
|
2020-01-20 20:47:47 +08:00
|
|
|
result(r, c) = +1;
|
Instance segmentation (#1918)
* Add instance segmentation example - first version of training code
* Add MMOD options; get rid of the cache approach, and instead load all MMOD rects upfront
* Improve console output
* Set filter count
* Minor tweaking
* Inference - first version, at least compiles!
* Ignore overlapped boxes
* Ignore even small instances
* Set overlaps_ignore
* Add TODO remarks
* Revert "Set overlaps_ignore"
This reverts commit 65adeff1f89af62b10c691e7aa86c04fc358d03e.
* Set result size
* Set label image size
* Take ignore-color into account
* Fix the cropping rect's aspect ratio; also slightly expand the rect
* Draw the largest findings last
* Improve masking of the current instance
* Add some perturbation to the inputs
* Simplify ground-truth reading; fix random cropping
* Read even class labels
* Tweak default minibatch size
* Learn only one class
* Really train only instances of the selected class
* Remove outdated TODO remark
* Automatically skip images with no detections
* Print to console what was found
* Fix class index problem
* Fix indentation
* Allow to choose multiple classes
* Draw rect in the color of the corresponding class
* Write detector window classes to ostream; also group detection windows by class (when ostreaming)
* Train a separate instance segmentation network for each classlabel
* Use separate synchronization file for each seg net of each class
* Allow more overlap
* Fix sorting criterion
* Fix interpolating the predicted mask
* Improve bilinear interpolation: if output type is an integer, round instead of truncating
* Add helpful comments
* Ignore large aspect ratios; refactor the code; tweak some network parameters
* Simplify the segmentation network structure; make the object detection network more complex in turn
* Problem: CUDA errors not reported properly to console
Solution: stop and join data loader threads even in case of exceptions
* Minor parameters tweaking
* Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value
* Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing
* Improve behaviour when loss actually increased after disk sync
* Revert some of the earlier change
* Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync
* Revert "Revert some of the earlier change"
This reverts commit 6c852124efe6473a5c962de0091709129d6fcde3.
* Keep enough previous loss values, until the disk sync
* Fix maintaining the dumped (now "effectively disregarded") loss values count
* Detect cats instead of aeroplanes
* Add helpful logging
* Clarify the intention and the code
* Review fixes
* Add operator== for the other pixel types as well; remove the inline
* If available, use constexpr if
* Revert "If available, use constexpr if"
This reverts commit 503d4dd3355ff8ad613116e3ffcc0fa664674f69.
* Simplify code as per review comments
* Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh
* Clarify console output
* Revert "Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh"
This reverts commit 9191ebc7762d17d81cdfc334a80ca9a667365740.
* To keep the changes to a bare minimum, revert the steps_since_last_learning_rate_shrink change after all (at least for now)
* Even empty out some of the previous test loss values
* Minor review fixes
* Can't use C++14 features here
* Do not use the struct name as a variable name
2019-11-15 11:53:16 +08:00
|
|
|
else if (index == dlib::rgb_pixel(224, 224, 192))
|
|
|
|
result(r, c) = 0;
|
2020-01-20 20:47:47 +08:00
|
|
|
else
|
|
|
|
result(r, c) = -1;
|
Instance segmentation (#1918)
* Add instance segmentation example - first version of training code
* Add MMOD options; get rid of the cache approach, and instead load all MMOD rects upfront
* Improve console output
* Set filter count
* Minor tweaking
* Inference - first version, at least compiles!
* Ignore overlapped boxes
* Ignore even small instances
* Set overlaps_ignore
* Add TODO remarks
* Revert "Set overlaps_ignore"
This reverts commit 65adeff1f89af62b10c691e7aa86c04fc358d03e.
* Set result size
* Set label image size
* Take ignore-color into account
* Fix the cropping rect's aspect ratio; also slightly expand the rect
* Draw the largest findings last
* Improve masking of the current instance
* Add some perturbation to the inputs
* Simplify ground-truth reading; fix random cropping
* Read even class labels
* Tweak default minibatch size
* Learn only one class
* Really train only instances of the selected class
* Remove outdated TODO remark
* Automatically skip images with no detections
* Print to console what was found
* Fix class index problem
* Fix indentation
* Allow to choose multiple classes
* Draw rect in the color of the corresponding class
* Write detector window classes to ostream; also group detection windows by class (when ostreaming)
* Train a separate instance segmentation network for each classlabel
* Use separate synchronization file for each seg net of each class
* Allow more overlap
* Fix sorting criterion
* Fix interpolating the predicted mask
* Improve bilinear interpolation: if output type is an integer, round instead of truncating
* Add helpful comments
* Ignore large aspect ratios; refactor the code; tweak some network parameters
* Simplify the segmentation network structure; make the object detection network more complex in turn
* Problem: CUDA errors not reported properly to console
Solution: stop and join data loader threads even in case of exceptions
* Minor parameters tweaking
* Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value
* Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing
* Improve behaviour when loss actually increased after disk sync
* Revert some of the earlier change
* Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync
* Revert "Revert some of the earlier change"
This reverts commit 6c852124efe6473a5c962de0091709129d6fcde3.
* Keep enough previous loss values, until the disk sync
* Fix maintaining the dumped (now "effectively disregarded") loss values count
* Detect cats instead of aeroplanes
* Add helpful logging
* Clarify the intention and the code
* Review fixes
* Add operator== for the other pixel types as well; remove the inline
* If available, use constexpr if
* Revert "If available, use constexpr if"
This reverts commit 503d4dd3355ff8ad613116e3ffcc0fa664674f69.
* Simplify code as per review comments
* Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh
* Clarify console output
* Revert "Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh"
This reverts commit 9191ebc7762d17d81cdfc334a80ca9a667365740.
* To keep the changes to a bare minimum, revert the steps_since_last_learning_rate_shrink change after all (at least for now)
* Even empty out some of the previous test loss values
* Minor review fixes
* Can't use C++14 features here
* Do not use the struct name as a variable name
2019-11-15 11:53:16 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
seg_bnet_type train_segmentation_network(
|
|
|
|
const std::vector<truth_image>& truth_images,
|
|
|
|
unsigned int seg_minibatch_size,
|
|
|
|
const std::string& classlabel
|
|
|
|
)
|
|
|
|
{
|
|
|
|
seg_bnet_type seg_net;
|
|
|
|
|
|
|
|
const double initial_learning_rate = 0.1;
|
|
|
|
const double weight_decay = 0.0001;
|
|
|
|
const double momentum = 0.9;
|
|
|
|
|
|
|
|
const std::string synchronization_file_name
|
|
|
|
= "pascal_voc2012_seg_trainer_state_file"
|
|
|
|
+ (classlabel.empty() ? "" : ("_" + classlabel))
|
|
|
|
+ ".dat";
|
|
|
|
|
|
|
|
dnn_trainer<seg_bnet_type> seg_trainer(seg_net, sgd(weight_decay, momentum));
|
|
|
|
seg_trainer.be_verbose();
|
|
|
|
seg_trainer.set_learning_rate(initial_learning_rate);
|
|
|
|
seg_trainer.set_synchronization_file(synchronization_file_name, std::chrono::minutes(10));
|
|
|
|
seg_trainer.set_iterations_without_progress_threshold(2000);
|
|
|
|
set_all_bn_running_stats_window_sizes(seg_net, 1000);
|
|
|
|
|
|
|
|
// Output training parameters.
|
|
|
|
cout << seg_trainer << endl;
|
|
|
|
|
|
|
|
std::vector<matrix<rgb_pixel>> samples;
|
2020-01-20 20:47:47 +08:00
|
|
|
std::vector<matrix<float>> labels;
|
Instance segmentation (#1918)
* Add instance segmentation example - first version of training code
* Add MMOD options; get rid of the cache approach, and instead load all MMOD rects upfront
* Improve console output
* Set filter count
* Minor tweaking
* Inference - first version, at least compiles!
* Ignore overlapped boxes
* Ignore even small instances
* Set overlaps_ignore
* Add TODO remarks
* Revert "Set overlaps_ignore"
This reverts commit 65adeff1f89af62b10c691e7aa86c04fc358d03e.
* Set result size
* Set label image size
* Take ignore-color into account
* Fix the cropping rect's aspect ratio; also slightly expand the rect
* Draw the largest findings last
* Improve masking of the current instance
* Add some perturbation to the inputs
* Simplify ground-truth reading; fix random cropping
* Read even class labels
* Tweak default minibatch size
* Learn only one class
* Really train only instances of the selected class
* Remove outdated TODO remark
* Automatically skip images with no detections
* Print to console what was found
* Fix class index problem
* Fix indentation
* Allow to choose multiple classes
* Draw rect in the color of the corresponding class
* Write detector window classes to ostream; also group detection windows by class (when ostreaming)
* Train a separate instance segmentation network for each classlabel
* Use separate synchronization file for each seg net of each class
* Allow more overlap
* Fix sorting criterion
* Fix interpolating the predicted mask
* Improve bilinear interpolation: if output type is an integer, round instead of truncating
* Add helpful comments
* Ignore large aspect ratios; refactor the code; tweak some network parameters
* Simplify the segmentation network structure; make the object detection network more complex in turn
* Problem: CUDA errors not reported properly to console
Solution: stop and join data loader threads even in case of exceptions
* Minor parameters tweaking
* Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value
* Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing
* Improve behaviour when loss actually increased after disk sync
* Revert some of the earlier change
* Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync
* Revert "Revert some of the earlier change"
This reverts commit 6c852124efe6473a5c962de0091709129d6fcde3.
* Keep enough previous loss values, until the disk sync
* Fix maintaining the dumped (now "effectively disregarded") loss values count
* Detect cats instead of aeroplanes
* Add helpful logging
* Clarify the intention and the code
* Review fixes
* Add operator== for the other pixel types as well; remove the inline
* If available, use constexpr if
* Revert "If available, use constexpr if"
This reverts commit 503d4dd3355ff8ad613116e3ffcc0fa664674f69.
* Simplify code as per review comments
* Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh
* Clarify console output
* Revert "Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh"
This reverts commit 9191ebc7762d17d81cdfc334a80ca9a667365740.
* To keep the changes to a bare minimum, revert the steps_since_last_learning_rate_shrink change after all (at least for now)
* Even empty out some of the previous test loss values
* Minor review fixes
* Can't use C++14 features here
* Do not use the struct name as a variable name
2019-11-15 11:53:16 +08:00
|
|
|
|
|
|
|
// Start a bunch of threads that read images from disk and pull out random crops. It's
|
|
|
|
// important to be sure to feed the GPU fast enough to keep it busy. Using multiple
|
|
|
|
// thread for this kind of data preparation helps us do that. Each thread puts the
|
|
|
|
// crops into the data queue.
|
|
|
|
dlib::pipe<seg_training_sample> data(200);
|
|
|
|
auto f = [&data, &truth_images](time_t seed)
|
|
|
|
{
|
|
|
|
dlib::rand rnd(time(0) + seed);
|
|
|
|
matrix<rgb_pixel> input_image;
|
|
|
|
matrix<rgb_pixel> rgb_label_image;
|
|
|
|
matrix<rgb_pixel> rgb_label_chip;
|
|
|
|
seg_training_sample temp;
|
|
|
|
while (data.is_enabled())
|
|
|
|
{
|
|
|
|
// Pick a random input image.
|
|
|
|
const auto random_index = rnd.get_random_32bit_number() % truth_images.size();
|
|
|
|
const auto& truth_image = truth_images[random_index];
|
|
|
|
const auto image_truths = truth_image.truth_instances;
|
|
|
|
|
|
|
|
if (!image_truths.empty())
|
|
|
|
{
|
|
|
|
const image_info& info = truth_image.info;
|
|
|
|
|
|
|
|
// Load the input image.
|
|
|
|
load_image(input_image, info.image_filename);
|
|
|
|
|
|
|
|
// Load the ground-truth (RGB) instance labels.
|
|
|
|
load_image(rgb_label_image, info.instance_label_filename);
|
|
|
|
|
|
|
|
// Pick a random training instance.
|
|
|
|
const auto& truth_instance = image_truths[rnd.get_random_32bit_number() % image_truths.size()];
|
|
|
|
const auto& truth_rect = truth_instance.mmod_rect.rect;
|
|
|
|
const auto cropping_rect = get_cropping_rect(truth_rect);
|
|
|
|
|
|
|
|
// Pick a random crop around the instance.
|
|
|
|
const auto max_x_translate_amount = static_cast<long>(truth_rect.width() / 10.0);
|
|
|
|
const auto max_y_translate_amount = static_cast<long>(truth_rect.height() / 10.0);
|
|
|
|
|
|
|
|
const auto random_translate = point(
|
|
|
|
rnd.get_integer_in_range(-max_x_translate_amount, max_x_translate_amount + 1),
|
|
|
|
rnd.get_integer_in_range(-max_y_translate_amount, max_y_translate_amount + 1)
|
|
|
|
);
|
|
|
|
|
|
|
|
const rectangle random_rect(
|
|
|
|
cropping_rect.left() + random_translate.x(),
|
|
|
|
cropping_rect.top() + random_translate.y(),
|
|
|
|
cropping_rect.right() + random_translate.x(),
|
|
|
|
cropping_rect.bottom() + random_translate.y()
|
|
|
|
);
|
|
|
|
|
|
|
|
const chip_details chip_details(random_rect, chip_dims(seg_dim, seg_dim));
|
|
|
|
|
|
|
|
// Crop the input image.
|
|
|
|
extract_image_chip(input_image, chip_details, temp.input_image, interpolate_bilinear());
|
|
|
|
|
|
|
|
disturb_colors(temp.input_image, rnd);
|
|
|
|
|
|
|
|
// Crop the labels correspondingly. However, note that here bilinear
|
|
|
|
// interpolation would make absolutely no sense - you wouldn't say that
|
|
|
|
// a bicycle is half-way between an aeroplane and a bird, would you?
|
|
|
|
extract_image_chip(rgb_label_image, chip_details, rgb_label_chip, interpolate_nearest_neighbor());
|
|
|
|
|
|
|
|
// Clear pixels not related to the current instance.
|
|
|
|
temp.label_image = keep_only_current_instance(rgb_label_chip, truth_instance.rgb_label);
|
|
|
|
|
|
|
|
// Push the result to be used by the trainer.
|
|
|
|
data.enqueue(temp);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
// TODO: use background samples as well
|
|
|
|
}
|
|
|
|
}
|
|
|
|
};
|
|
|
|
std::thread data_loader1([f]() { f(1); });
|
|
|
|
std::thread data_loader2([f]() { f(2); });
|
|
|
|
std::thread data_loader3([f]() { f(3); });
|
|
|
|
std::thread data_loader4([f]() { f(4); });
|
|
|
|
|
|
|
|
const auto stop_data_loaders = [&]()
|
|
|
|
{
|
|
|
|
data.disable();
|
|
|
|
data_loader1.join();
|
|
|
|
data_loader2.join();
|
|
|
|
data_loader3.join();
|
|
|
|
data_loader4.join();
|
|
|
|
};
|
|
|
|
|
|
|
|
try
|
|
|
|
{
|
|
|
|
// The main training loop. Keep making mini-batches and giving them to the trainer.
|
|
|
|
// We will run until the learning rate has dropped by a factor of 1e-4.
|
|
|
|
while (seg_trainer.get_learning_rate() >= 1e-4)
|
|
|
|
{
|
|
|
|
samples.clear();
|
|
|
|
labels.clear();
|
|
|
|
|
|
|
|
// make a mini-batch
|
|
|
|
seg_training_sample temp;
|
|
|
|
while (samples.size() < seg_minibatch_size)
|
|
|
|
{
|
|
|
|
data.dequeue(temp);
|
|
|
|
|
|
|
|
samples.push_back(std::move(temp.input_image));
|
|
|
|
labels.push_back(std::move(temp.label_image));
|
|
|
|
}
|
|
|
|
|
|
|
|
seg_trainer.train_one_step(samples, labels);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
catch (std::exception&)
|
|
|
|
{
|
|
|
|
stop_data_loaders();
|
|
|
|
throw;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Training done, tell threads to stop and make sure to wait for them to finish before
|
|
|
|
// moving on.
|
|
|
|
stop_data_loaders();
|
|
|
|
|
|
|
|
// also wait for threaded processing to stop in the trainer.
|
|
|
|
seg_trainer.get_net();
|
|
|
|
|
|
|
|
seg_net.clean();
|
|
|
|
|
|
|
|
return seg_net;
|
|
|
|
}
|
|
|
|
|
|
|
|
// ----------------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
int ignore_overlapped_boxes(
|
|
|
|
std::vector<truth_instance>& truth_instances,
|
|
|
|
const test_box_overlap& overlaps
|
|
|
|
)
|
|
|
|
/*!
|
|
|
|
ensures
|
|
|
|
- Whenever two rectangles in boxes overlap, according to overlaps(), we set the
|
|
|
|
smallest box to ignore.
|
|
|
|
- returns the number of newly ignored boxes.
|
|
|
|
!*/
|
|
|
|
{
|
|
|
|
int num_ignored = 0;
|
|
|
|
for (size_t i = 0, end = truth_instances.size(); i < end; ++i)
|
|
|
|
{
|
|
|
|
auto& box_i = truth_instances[i].mmod_rect;
|
|
|
|
if (box_i.ignore)
|
|
|
|
continue;
|
|
|
|
for (size_t j = i+1; j < end; ++j)
|
|
|
|
{
|
|
|
|
auto& box_j = truth_instances[j].mmod_rect;
|
|
|
|
if (box_j.ignore)
|
|
|
|
continue;
|
|
|
|
if (overlaps(box_i, box_j))
|
|
|
|
{
|
|
|
|
++num_ignored;
|
|
|
|
if(box_i.rect.area() < box_j.rect.area())
|
|
|
|
box_i.ignore = true;
|
|
|
|
else
|
|
|
|
box_j.ignore = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return num_ignored;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<truth_instance> load_truth_instances(const image_info& info)
|
|
|
|
{
|
|
|
|
matrix<rgb_pixel> instance_label_image;
|
|
|
|
matrix<rgb_pixel> class_label_image;
|
|
|
|
|
|
|
|
load_image(instance_label_image, info.instance_label_filename);
|
|
|
|
load_image(class_label_image, info.class_label_filename);
|
|
|
|
|
|
|
|
return rgb_label_images_to_truth_instances(instance_label_image, class_label_image);
|
2020-04-01 07:35:23 +08:00
|
|
|
}
|
Instance segmentation (#1918)
* Add instance segmentation example - first version of training code
* Add MMOD options; get rid of the cache approach, and instead load all MMOD rects upfront
* Improve console output
* Set filter count
* Minor tweaking
* Inference - first version, at least compiles!
* Ignore overlapped boxes
* Ignore even small instances
* Set overlaps_ignore
* Add TODO remarks
* Revert "Set overlaps_ignore"
This reverts commit 65adeff1f89af62b10c691e7aa86c04fc358d03e.
* Set result size
* Set label image size
* Take ignore-color into account
* Fix the cropping rect's aspect ratio; also slightly expand the rect
* Draw the largest findings last
* Improve masking of the current instance
* Add some perturbation to the inputs
* Simplify ground-truth reading; fix random cropping
* Read even class labels
* Tweak default minibatch size
* Learn only one class
* Really train only instances of the selected class
* Remove outdated TODO remark
* Automatically skip images with no detections
* Print to console what was found
* Fix class index problem
* Fix indentation
* Allow to choose multiple classes
* Draw rect in the color of the corresponding class
* Write detector window classes to ostream; also group detection windows by class (when ostreaming)
* Train a separate instance segmentation network for each classlabel
* Use separate synchronization file for each seg net of each class
* Allow more overlap
* Fix sorting criterion
* Fix interpolating the predicted mask
* Improve bilinear interpolation: if output type is an integer, round instead of truncating
* Add helpful comments
* Ignore large aspect ratios; refactor the code; tweak some network parameters
* Simplify the segmentation network structure; make the object detection network more complex in turn
* Problem: CUDA errors not reported properly to console
Solution: stop and join data loader threads even in case of exceptions
* Minor parameters tweaking
* Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value
* Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing
* Improve behaviour when loss actually increased after disk sync
* Revert some of the earlier change
* Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync
* Revert "Revert some of the earlier change"
This reverts commit 6c852124efe6473a5c962de0091709129d6fcde3.
* Keep enough previous loss values, until the disk sync
* Fix maintaining the dumped (now "effectively disregarded") loss values count
* Detect cats instead of aeroplanes
* Add helpful logging
* Clarify the intention and the code
* Review fixes
* Add operator== for the other pixel types as well; remove the inline
* If available, use constexpr if
* Revert "If available, use constexpr if"
This reverts commit 503d4dd3355ff8ad613116e3ffcc0fa664674f69.
* Simplify code as per review comments
* Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh
* Clarify console output
* Revert "Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh"
This reverts commit 9191ebc7762d17d81cdfc334a80ca9a667365740.
* To keep the changes to a bare minimum, revert the steps_since_last_learning_rate_shrink change after all (at least for now)
* Even empty out some of the previous test loss values
* Minor review fixes
* Can't use C++14 features here
* Do not use the struct name as a variable name
2019-11-15 11:53:16 +08:00
|
|
|
|
|
|
|
std::vector<std::vector<truth_instance>> load_all_truth_instances(const std::vector<image_info>& listing)
|
|
|
|
{
|
|
|
|
std::vector<std::vector<truth_instance>> truth_instances(listing.size());
|
|
|
|
|
|
|
|
std::transform(
|
|
|
|
#if __cplusplus >= 201703L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
|
|
|
|
std::execution::par,
|
|
|
|
#endif // __cplusplus >= 201703L
|
|
|
|
listing.begin(),
|
|
|
|
listing.end(),
|
|
|
|
truth_instances.begin(),
|
|
|
|
load_truth_instances
|
|
|
|
);
|
|
|
|
|
|
|
|
return truth_instances;
|
|
|
|
}
|
|
|
|
|
|
|
|
// ----------------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
std::vector<truth_image> filter_based_on_classlabel(
|
|
|
|
const std::vector<truth_image>& truth_images,
|
|
|
|
const std::vector<std::string>& desired_classlabels
|
|
|
|
)
|
|
|
|
{
|
|
|
|
std::vector<truth_image> result;
|
|
|
|
|
|
|
|
const auto represents_desired_class = [&desired_classlabels](const truth_instance& truth_instance) {
|
|
|
|
return std::find(
|
|
|
|
desired_classlabels.begin(),
|
|
|
|
desired_classlabels.end(),
|
|
|
|
truth_instance.mmod_rect.label
|
|
|
|
) != desired_classlabels.end();
|
|
|
|
};
|
|
|
|
|
|
|
|
for (const auto& input : truth_images)
|
|
|
|
{
|
|
|
|
const auto has_desired_class = std::any_of(
|
|
|
|
input.truth_instances.begin(),
|
|
|
|
input.truth_instances.end(),
|
|
|
|
represents_desired_class
|
|
|
|
);
|
|
|
|
|
|
|
|
if (has_desired_class) {
|
|
|
|
|
|
|
|
// NB: This keeps only MMOD rects belonging to any of the desired classes.
|
|
|
|
// A reasonable alternative could be to keep all rects, but mark those
|
|
|
|
// belonging in other classes to be ignored during training.
|
|
|
|
std::vector<truth_instance> temp;
|
|
|
|
std::copy_if(
|
|
|
|
input.truth_instances.begin(),
|
|
|
|
input.truth_instances.end(),
|
|
|
|
std::back_inserter(temp),
|
|
|
|
represents_desired_class
|
|
|
|
);
|
|
|
|
|
|
|
|
result.push_back(truth_image{ input.info, temp });
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Ignore truth boxes that overlap too much, are too small, or have a large aspect ratio.
|
|
|
|
void ignore_some_truth_boxes(std::vector<truth_image>& truth_images)
|
|
|
|
{
|
|
|
|
for (auto& i : truth_images)
|
|
|
|
{
|
|
|
|
auto& truth_instances = i.truth_instances;
|
|
|
|
|
|
|
|
ignore_overlapped_boxes(truth_instances, test_box_overlap(0.90, 0.95));
|
|
|
|
|
|
|
|
for (auto& truth : truth_instances)
|
|
|
|
{
|
|
|
|
if (truth.mmod_rect.ignore)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
const auto& rect = truth.mmod_rect.rect;
|
|
|
|
|
|
|
|
constexpr unsigned long min_width = 35;
|
|
|
|
constexpr unsigned long min_height = 35;
|
|
|
|
if (rect.width() < min_width && rect.height() < min_height)
|
|
|
|
{
|
|
|
|
truth.mmod_rect.ignore = true;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
constexpr double max_aspect_ratio_width_to_height = 3.0;
|
|
|
|
constexpr double max_aspect_ratio_height_to_width = 1.5;
|
|
|
|
const double aspect_ratio_width_to_height = rect.width() / static_cast<double>(rect.height());
|
|
|
|
const double aspect_ratio_height_to_width = 1.0 / aspect_ratio_width_to_height;
|
|
|
|
const bool is_aspect_ratio_too_large
|
|
|
|
= aspect_ratio_width_to_height > max_aspect_ratio_width_to_height
|
|
|
|
|| aspect_ratio_height_to_width > max_aspect_ratio_height_to_width;
|
|
|
|
|
|
|
|
if (is_aspect_ratio_too_large)
|
|
|
|
truth.mmod_rect.ignore = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Filter images that have no (non-ignored) truth
|
|
|
|
std::vector<truth_image> filter_images_with_no_truth(const std::vector<truth_image>& truth_images)
|
|
|
|
{
|
|
|
|
std::vector<truth_image> result;
|
|
|
|
|
|
|
|
for (const auto& truth_image : truth_images)
|
|
|
|
{
|
|
|
|
const auto ignored = [](const truth_instance& truth) { return truth.mmod_rect.ignore; };
|
|
|
|
const auto& truth_instances = truth_image.truth_instances;
|
|
|
|
if (!std::all_of(truth_instances.begin(), truth_instances.end(), ignored))
|
|
|
|
result.push_back(truth_image);
|
|
|
|
}
|
|
|
|
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
int main(int argc, char** argv) try
|
|
|
|
{
|
|
|
|
if (argc < 2)
|
|
|
|
{
|
|
|
|
cout << "To run this program you need a copy of the PASCAL VOC2012 dataset." << endl;
|
|
|
|
cout << endl;
|
|
|
|
cout << "You call this program like this: " << endl;
|
|
|
|
cout << "./dnn_instance_segmentation_train_ex /path/to/VOC2012 [det-minibatch-size] [seg-minibatch-size] [class-1] [class-2] [class-3] ..." << endl;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
cout << "\nSCANNING PASCAL VOC2012 DATASET\n" << endl;
|
|
|
|
|
|
|
|
const auto listing = get_pascal_voc2012_train_listing(argv[1]);
|
|
|
|
cout << "images in entire dataset: " << listing.size() << endl;
|
|
|
|
if (listing.size() == 0)
|
|
|
|
{
|
|
|
|
cout << "Didn't find the VOC2012 dataset. " << endl;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
// mini-batches smaller than the default can be used with GPUs having less memory
|
|
|
|
const unsigned int det_minibatch_size = argc >= 3 ? std::stoi(argv[2]) : 35;
|
|
|
|
const unsigned int seg_minibatch_size = argc >= 4 ? std::stoi(argv[3]) : 100;
|
|
|
|
cout << "det mini-batch size: " << det_minibatch_size << endl;
|
|
|
|
cout << "seg mini-batch size: " << seg_minibatch_size << endl;
|
|
|
|
|
|
|
|
std::vector<std::string> desired_classlabels;
|
|
|
|
|
|
|
|
for (int arg = 4; arg < argc; ++arg)
|
|
|
|
desired_classlabels.push_back(argv[arg]);
|
|
|
|
|
|
|
|
if (desired_classlabels.empty())
|
|
|
|
{
|
|
|
|
desired_classlabels.push_back("bicycle");
|
|
|
|
desired_classlabels.push_back("car");
|
|
|
|
desired_classlabels.push_back("cat");
|
|
|
|
}
|
|
|
|
|
|
|
|
cout << "desired classlabels:";
|
|
|
|
for (const auto& desired_classlabel : desired_classlabels)
|
|
|
|
cout << " " << desired_classlabel;
|
|
|
|
cout << endl;
|
|
|
|
|
|
|
|
// extract the MMOD rects
|
|
|
|
cout << endl << "Extracting all truth instances...";
|
|
|
|
const auto truth_instances = load_all_truth_instances(listing);
|
|
|
|
cout << " Done!" << endl << endl;
|
|
|
|
|
|
|
|
DLIB_CASSERT(listing.size() == truth_instances.size());
|
|
|
|
|
|
|
|
std::vector<truth_image> original_truth_images;
|
|
|
|
for (size_t i = 0, end = listing.size(); i < end; ++i)
|
|
|
|
{
|
|
|
|
original_truth_images.push_back(truth_image{
|
|
|
|
listing[i], truth_instances[i]
|
|
|
|
});
|
|
|
|
}
|
|
|
|
|
|
|
|
auto truth_images_filtered_by_class = filter_based_on_classlabel(original_truth_images, desired_classlabels);
|
|
|
|
|
|
|
|
cout << "images in dataset filtered by class: " << truth_images_filtered_by_class.size() << endl;
|
|
|
|
|
|
|
|
ignore_some_truth_boxes(truth_images_filtered_by_class);
|
|
|
|
const auto truth_images = filter_images_with_no_truth(truth_images_filtered_by_class);
|
|
|
|
|
|
|
|
cout << "images in dataset after ignoring some truth boxes: " << truth_images.size() << endl;
|
|
|
|
|
|
|
|
// First train an object detector network (loss_mmod).
|
|
|
|
cout << endl << "Training detector network:" << endl;
|
|
|
|
const auto det_net = train_detection_network(truth_images, det_minibatch_size);
|
|
|
|
|
|
|
|
// Then train mask predictors (segmentation).
|
|
|
|
std::map<std::string, seg_bnet_type> seg_nets_by_class;
|
|
|
|
|
|
|
|
// This flag controls if a separate mask predictor is trained for each class.
|
|
|
|
// Note that it would also be possible to train a separate mask predictor for
|
|
|
|
// class groups, each containing somehow similar classes -- for example, one
|
|
|
|
// mask predictor for cars and buses, another for cats and dogs, and so on.
|
|
|
|
constexpr bool separate_seg_net_for_each_class = true;
|
|
|
|
|
|
|
|
if (separate_seg_net_for_each_class)
|
|
|
|
{
|
|
|
|
for (const auto& classlabel : desired_classlabels)
|
|
|
|
{
|
|
|
|
// Consider only the truth images belonging to this class.
|
|
|
|
const auto class_images = filter_based_on_classlabel(truth_images, { classlabel });
|
|
|
|
|
|
|
|
cout << endl << "Training segmentation network for class " << classlabel << ":" << endl;
|
|
|
|
seg_nets_by_class[classlabel] = train_segmentation_network(class_images, seg_minibatch_size, classlabel);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
cout << "Training a single segmentation network:" << endl;
|
|
|
|
seg_nets_by_class[""] = train_segmentation_network(truth_images, seg_minibatch_size, "");
|
|
|
|
}
|
|
|
|
|
|
|
|
cout << "Saving networks" << endl;
|
|
|
|
serialize(instance_segmentation_net_filename) << det_net << seg_nets_by_class;
|
|
|
|
}
|
|
|
|
|
|
|
|
catch(std::exception& e)
|
|
|
|
{
|
|
|
|
cout << e.what() << endl;
|
|
|
|
}
|
|
|
|
|