diff options
author | mcarrier <mcarrier@636b058d-ea47-450e-bf9e-a15bfbe3eedb> | 2018-06-23 04:59:39 +0000 |
---|---|---|
committer | mcarrier <mcarrier@636b058d-ea47-450e-bf9e-a15bfbe3eedb> | 2018-06-23 04:59:39 +0000 |
commit | 0741c3eabbfece1c73ac76aa44adbe2904b6124d (patch) | |
tree | 83fc9edee2be2c5b1fc73d9d3cc1ddc4a66c828b /src/Persistence_representations/include/gudhi | |
parent | 10c6f6be72a2631cd1a1d28ed61343d55bd2b759 (diff) |
git-svn-id: svn+ssh://scm.gforge.inria.fr/svnroot/gudhi/branches/kernels@3628 636b058d-ea47-450e-bf9e-a15bfbe3eedb
Former-commit-id: 12f32a1c8ca31e7e0a40e1c3502e2a3d810d5bc5
Diffstat (limited to 'src/Persistence_representations/include/gudhi')
9 files changed, 158 insertions, 634 deletions
diff --git a/src/Persistence_representations/include/gudhi/Betti_sequence.h b/src/Persistence_representations/include/gudhi/Betti_sequence.h deleted file mode 100644 index 57c52ad2..00000000 --- a/src/Persistence_representations/include/gudhi/Betti_sequence.h +++ /dev/null @@ -1,95 +0,0 @@ -/* This file is part of the Gudhi Library. The Gudhi library - * (Geometric Understanding in Higher Dimensions) is a generic C++ - * library for computational topology. - * - * Author(s): Mathieu Carriere - * - * Copyright (C) 2018 INRIA (France) - * - * This program is free software: you can redistribute it and/or modify - * it under the terms of the GNU General Public License as published by - * the Free Software Foundation, either version 3 of the License, or - * (at your option) any later version. - * - * This program is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. - * - * You should have received a copy of the GNU General Public License - * along with this program. If not, see <http://www.gnu.org/licenses/>. - */ - -#ifndef BETTI_SEQUENCE_H_ -#define BETTI_SEQUENCE_H_ - -// gudhi include -#include <gudhi/read_persistence_from_file.h> -#include <gudhi/common_persistence_representations.h> -#include <gudhi/Debug_utils.h> - -// standard include -#include <cmath> -#include <iostream> -#include <vector> -#include <limits> -#include <fstream> -#include <sstream> -#include <algorithm> -#include <string> -#include <utility> -#include <functional> - -namespace Gudhi { -namespace Persistence_representations { - -/** - * \class Betti_sequence gudhi/Betti_sequence.h - * \brief A class implementing Betti sequences - * - * \ingroup Persistence_representations - * - * \details -**/ - -class Betti_sequence { - - protected: - Persistence_diagram diagram; - int res_x, nb_cv; - double min_x, max_x; - - public: - - /** \brief Betti_sequence constructor. - * \ingroup Betti_sequence - * - * @param[in] _diagram persistence diagram. - * @param[in] _min_x minimum value of samples. - * @param[in] _max_x maximum value of samples. - * @param[in] _res_x number of samples. - * - */ - Betti_sequence(const Persistence_diagram & _diagram, double _min_x = 0.0, double _max_x = 1.0, int _res_x = 10){diagram = _diagram; min_x = _min_x; max_x = _max_x; res_x = _res_x;} - - /** \brief Computes the Betti sequences of a diagram. - * \ingroup Betti_sequence - * - */ - std::vector<int> vectorize() const { - int num_pts = diagram.size(); double step = (max_x - min_x)/(res_x - 1); - std::vector<int> bs(res_x); for(int i = 0; i < res_x; i++) bs[i] = 0; - for(int j = 0; j < num_pts; j++){ - double px = diagram[j].first; double py = diagram[j].second; - int first = std::ceil((px-min_x)/step); int last = std::ceil((py-min_x)/step); - for(int i = first; i < last; i++) bs[i] += 1; - } - - return bs; - } - -}; // class Betti_sequence -} // namespace Persistence_representations -} // namespace Gudhi - -#endif // BETTI_SEQUENCE_H_ diff --git a/src/Persistence_representations/include/gudhi/Persistence_heat_maps.h b/src/Persistence_representations/include/gudhi/Persistence_heat_maps.h index 35e51e63..63c6e239 100644 --- a/src/Persistence_representations/include/gudhi/Persistence_heat_maps.h +++ b/src/Persistence_representations/include/gudhi/Persistence_heat_maps.h @@ -245,6 +245,20 @@ class Persistence_heat_maps { unsigned dimension = std::numeric_limits<unsigned>::max()); /** + * Construction that takes as inputs (1) the diagram, (2) grid parameters (min, max and number of samples for x and y axes), and (3) a universal kernel on the plane used + * to turn the diagram into a function. + **/ + Persistence_heat_maps(const Persistence_diagram & interval, size_t number_of_x_pixels, size_t number_of_y_pixels, + double min_x = 0, double max_x = 1, double min_y = 0, double max_y = 1, const Kernel & kernel = Gaussian_kernel(1.0)); + + /** + * Construction that takes as inputs (1) the diagram and (2) a universal kernel on the plane used + * to turn the diagram into a function. Note that this construction is infinite dimensional so + * only compute_scalar_product() method is valid after calling this constructor. + **/ + Persistence_heat_maps(const Persistence_diagram & interval, const Kernel & kernel = Gaussian_kernel(1.0)); + + /** * Compute a mean value of a collection of heat maps and store it in the current object. Note that all the persistence *maps send in a vector to this procedure need to have the same parameters. * If this is not the case, the program will throw an exception. @@ -512,15 +526,27 @@ class Persistence_heat_maps { size_t number_of_functions_for_projections_to_reals; void construct(const std::vector<std::pair<double, double> >& intervals_, std::vector<std::vector<double> > filter = create_Gaussian_filter(5, 1), - bool erase_below_diagonal = false, size_t number_of_pixels = 1000, double min_ = std::numeric_limits<double>::max(), double max_ = std::numeric_limits<double>::max()); + void construct_image_from_exact_universal_kernel(const Persistence_diagram & interval, + size_t number_of_x_pixels = 10, size_t number_of_y_pixels = 10, + double min_x = 0, double max_x = 1, double min_y = 0, double max_y = 1, const Kernel & kernel = Gaussian_kernel(1.0)); + void construct_kernel_from_exact_universal_kernel(const Persistence_diagram & interval, const Kernel & kernel = Gaussian_kernel(1.0)); + void set_up_parameters_for_basic_classes() { this->number_of_functions_for_vectorization = 1; this->number_of_functions_for_projections_to_reals = 1; } + // Boolean indicating if we are computing persistence image (true) or persistence weighted gaussian kernel (false) + bool discrete = true; + + // PWGK + Kernel k; + Persistence_diagram d; + std::vector<double> weights; + // data Scalling_of_kernels f; bool erase_below_diagonal; @@ -529,6 +555,59 @@ class Persistence_heat_maps { std::vector<std::vector<double> > heat_map; }; +template <typename Scalling_of_kernels> +void Persistence_heat_maps<Scalling_of_kernels>::construct_image_from_exact_universal_kernel(const Persistence_diagram & diagram, + size_t number_of_x_pixels, size_t number_of_y_pixels, + double min_x, double max_x, + double min_y, double max_y, const Kernel & kernel) { + + this->discrete = true; Scalling_of_kernels f; this->f = f; this->min_ = min_x; this->max_ = max_x; + for(size_t i = 0; i < number_of_y_pixels; i++) this->heat_map.emplace_back(); + double step_x = (max_x - min_x)/(number_of_x_pixels - 1); double step_y = (max_y - min_y)/(number_of_y_pixels - 1); + + int num_pts = diagram.size(); + + for(size_t i = 0; i < number_of_y_pixels; i++){ + double y = min_y + i*step_y; + for(size_t j = 0; j < number_of_x_pixels; j++){ + double x = min_x + j*step_x; + + std::pair<double, double> grid_point(x,y); double pixel_value = 0; + for(int k = 0; k < num_pts; k++){ + double px = diagram[k].first; double py = diagram[k].second; std::pair<double, double> diagram_point(px,py); + pixel_value += this->f(diagram_point) * kernel(diagram_point, grid_point); + } + this->heat_map[i].push_back(pixel_value); + + } + } + +} + + +template <typename Scalling_of_kernels> +Persistence_heat_maps<Scalling_of_kernels>::Persistence_heat_maps(const Persistence_diagram & diagram, + size_t number_of_x_pixels, size_t number_of_y_pixels, + double min_x, double max_x, + double min_y, double max_y, const Kernel & kernel) { + this->construct_image_from_exact_universal_kernel(diagram, number_of_x_pixels, number_of_y_pixels, min_x, max_x, min_y, max_y, kernel); + this->set_up_parameters_for_basic_classes(); +} + +template <typename Scalling_of_kernels> +void Persistence_heat_maps<Scalling_of_kernels>::construct_kernel_from_exact_universal_kernel(const Persistence_diagram & diagram, const Kernel & kernel){ + this->discrete = false; Scalling_of_kernels f; this->f = f; this->k = kernel; this->d = diagram; + int num_pts = this->d.size(); + for (int i = 0; i < num_pts; i++) this->weights.push_back(this->f(this->d[i])); +} + + +template <typename Scalling_of_kernels> +Persistence_heat_maps<Scalling_of_kernels>::Persistence_heat_maps(const Persistence_diagram& diagram, const Kernel & kernel) { + this->construct_kernel_from_exact_universal_kernel(diagram, kernel); + this->set_up_parameters_for_basic_classes(); +} + // if min_ == max_, then the program is requested to set up the values itself based on persistence intervals template <typename Scalling_of_kernels> void Persistence_heat_maps<Scalling_of_kernels>::construct(const std::vector<std::pair<double, double> >& intervals_, @@ -826,13 +905,16 @@ void Persistence_heat_maps<Scalling_of_kernels>::load_from_file(const char* file // Concretizations of virtual methods: template <typename Scalling_of_kernels> std::vector<double> Persistence_heat_maps<Scalling_of_kernels>::vectorize(int number_of_function) const { + + std::vector<double> result; + if(!discrete){std::cout << "No vectorize method in case of infinite dimensional vectorization" << std::endl; return result;} + // convert this->heat_map into one large vector: size_t size_of_result = 0; for (size_t i = 0; i != this->heat_map.size(); ++i) { size_of_result += this->heat_map[i].size(); } - std::vector<double> result; result.reserve(size_of_result); for (size_t i = 0; i != this->heat_map.size(); ++i) { @@ -846,34 +928,39 @@ std::vector<double> Persistence_heat_maps<Scalling_of_kernels>::vectorize(int nu template <typename Scalling_of_kernels> double Persistence_heat_maps<Scalling_of_kernels>::distance(const Persistence_heat_maps& second, double power) const { - // first we need to check if (*this) and second are defined on the same domain and have the same dimensions: - if (!this->check_if_the_same(second)) { - std::cerr << "The persistence images are of non compatible sizes. We cannot therefore compute distance between " - "them. The program will now terminate"; - throw "The persistence images are of non compatible sizes. The program will now terminate"; - } + if(this->discrete){ + // first we need to check if (*this) and second are defined on the same domain and have the same dimensions: + if (!this->check_if_the_same(second)) { + std::cerr << "The persistence images are of non compatible sizes. We cannot therefore compute distance between " + "them. The program will now terminate"; + throw "The persistence images are of non compatible sizes. The program will now terminate"; + } - // if we are here, we know that the two persistence images are defined on the same domain, so we can start computing - // their distances: + // if we are here, we know that the two persistence images are defined on the same domain, so we can start computing their distances: - double distance = 0; - if (power < std::numeric_limits<double>::max()) { - for (size_t i = 0; i != this->heat_map.size(); ++i) { - for (size_t j = 0; j != this->heat_map[i].size(); ++j) { - distance += pow(fabs(this->heat_map[i][j] - second.heat_map[i][j]), power); + double distance = 0; + if (power < std::numeric_limits<double>::max()) { + for (size_t i = 0; i != this->heat_map.size(); ++i) { + for (size_t j = 0; j != this->heat_map[i].size(); ++j) { + distance += pow(fabs(this->heat_map[i][j] - second.heat_map[i][j]), power); + } } - } - } else { - // in this case, we compute max norm distance - for (size_t i = 0; i != this->heat_map.size(); ++i) { - for (size_t j = 0; j != this->heat_map[i].size(); ++j) { - if (distance < fabs(this->heat_map[i][j] - second.heat_map[i][j])) { - distance = fabs(this->heat_map[i][j] - second.heat_map[i][j]); + } else { + // in this case, we compute max norm distance + for (size_t i = 0; i != this->heat_map.size(); ++i) { + for (size_t j = 0; j != this->heat_map[i].size(); ++j) { + if (distance < fabs(this->heat_map[i][j] - second.heat_map[i][j])) { + distance = fabs(this->heat_map[i][j] - second.heat_map[i][j]); + } } } } + return distance; + } else { + + return std::sqrt(this->compute_scalar_product(*this) + second.compute_scalar_product(second) -2 * this->compute_scalar_product(second)); + } - return distance; } template <typename Scalling_of_kernels> @@ -895,22 +982,37 @@ void Persistence_heat_maps<Scalling_of_kernels>::compute_average( template <typename Scalling_of_kernels> double Persistence_heat_maps<Scalling_of_kernels>::compute_scalar_product(const Persistence_heat_maps& second) const { - // first we need to check if (*this) and second are defined on the same domain and have the same dimensions: - if (!this->check_if_the_same(second)) { - std::cerr << "The persistence images are of non compatible sizes. We cannot therefore compute distance between " - "them. The program will now terminate"; - throw "The persistence images are of non compatible sizes. The program will now terminate"; - } - // if we are here, we know that the two persistence images are defined on the same domain, so we can start computing - // their scalar product: - double scalar_prod = 0; - for (size_t i = 0; i != this->heat_map.size(); ++i) { - for (size_t j = 0; j != this->heat_map[i].size(); ++j) { - scalar_prod += this->heat_map[i][j] * second.heat_map[i][j]; + if(discrete){ + // first we need to check if (*this) and second are defined on the same domain and have the same dimensions: + if (!this->check_if_the_same(second)) { + std::cerr << "The persistence images are of non compatible sizes. We cannot therefore compute distance between " + "them. The program will now terminate"; + throw "The persistence images are of non compatible sizes. The program will now terminate"; } + + // if we are here, we know that the two persistence images are defined on the same domain, so we can start computing + // their scalar product: + double scalar_prod = 0; + for (size_t i = 0; i != this->heat_map.size(); ++i) { + for (size_t j = 0; j != this->heat_map[i].size(); ++j) { + scalar_prod += this->heat_map[i][j] * second.heat_map[i][j]; + } + } + return scalar_prod; } - return scalar_prod; + + else{ + GUDHI_CHECK(this->approx != second.approx || this->f != second.f, std::invalid_argument("Error: different values for representations")); + + int num_pts1 = this->d.size(); int num_pts2 = second.d.size(); double kernel_val = 0; + for(int i = 0; i < num_pts1; i++) + for(int j = 0; j < num_pts2; j++) + kernel_val += this->weights[i] * second.weights[j] * this->k(this->d[i], second.d[j]); + return kernel_val; + } + + } } // namespace Persistence_representations diff --git a/src/Persistence_representations/include/gudhi/Persistence_heat_maps_exact.h b/src/Persistence_representations/include/gudhi/Persistence_heat_maps_exact.h deleted file mode 100644 index 7c5b2fdc..00000000 --- a/src/Persistence_representations/include/gudhi/Persistence_heat_maps_exact.h +++ /dev/null @@ -1,125 +0,0 @@ -/* This file is part of the Gudhi Library. The Gudhi library - * (Geometric Understanding in Higher Dimensions) is a generic C++ - * library for computational topology. - * - * Author(s): Mathieu Carriere - * - * Copyright (C) 2018 INRIA (France) - * - * This program is free software: you can redistribute it and/or modify - * it under the terms of the GNU General Public License as published by - * the Free Software Foundation, either version 3 of the License, or - * (at your option) any later version. - * - * This program is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. - * - * You should have received a copy of the GNU General Public License - * along with this program. If not, see <http://www.gnu.org/licenses/>. - */ - -#ifndef PERSISTENCE_HEAT_MAPS_EXACT_H_ -#define PERSISTENCE_HEAT_MAPS_EXACT_H_ - -// gudhi include -#include <gudhi/read_persistence_from_file.h> -#include <gudhi/common_persistence_representations.h> -#include <gudhi/Weight_functions.h> -#include <gudhi/Debug_utils.h> - -// standard include -#include <cmath> -#include <iostream> -#include <vector> -#include <limits> -#include <fstream> -#include <sstream> -#include <algorithm> -#include <string> -#include <utility> -#include <functional> - -namespace Gudhi { -namespace Persistence_representations { - -/** - * \class Persistence_heat_maps_exact gudhi/Persistence_heat_maps_exact.h - * \brief A class implementing exact persistence heat maps. - * - * \ingroup Persistence_representations - * - * \details - * - * In this class, we propose a way to approximate persistence heat maps, or persistence surfaces, by centering weighted Gaussians on each point of the persistence diagram, and evaluating these (exact) weighted Gaussian functions - * on the pixels of a 2D grid. Note that this scheme is different from the one proposed in Persistence_heat_maps, which first maps the points of the diagram to a 2D grid, and then evaluates the (approximate) weighted Gaussian functions. - * Hence, the difference is that we do not modify the diagram in this implementation, but the code can be slower to run. -**/ - -class Persistence_heat_maps_exact { - - protected: - Persistence_diagram diagram; - int res_x, res_y; - double min_x, max_x, min_y, max_y; - Weight weight; - double sigma; - - public: - - /** \brief Persistence_heat_maps_exact constructor. - * \ingroup Persistence_heat_maps_exact - * - * @param[in] _diagram persistence diagram. - * @param[in] _min_x minimum value of pixel abscissa. - * @param[in] _max_x maximum value of pixel abscissa. - * @param[in] _res_x number of pixels for the x-direction. - * @param[in] _min_y minimum value of pixel ordinate. - * @param[in] _max_y maximum value of pixel ordinate. - * @param[in] _res_y number of pixels for the y-direction. - * @param[in] _weight weight function for the Gaussians. - * @param[in] _sigma bandwidth parameter for the Gaussians. - * - */ - Persistence_heat_maps_exact(const Persistence_diagram & _diagram, double _min_x = 0.0, double _max_x = 1.0, int _res_x = 10, double _min_y = 0.0, double _max_y = 1.0, int _res_y = 10, const Weight & _weight = arctan_weight(1,1), double _sigma = 1.0){ - diagram = _diagram; min_x = _min_x; max_x = _max_x; res_x = _res_x; min_y = _min_y; max_y = _max_y; res_y = _res_y, weight = _weight; sigma = _sigma; - } - - /** \brief Computes the persistence image of a diagram. - * \ingroup Persistence_heat_maps_exact - * - */ - std::vector<std::vector<double> > vectorize() const { - std::vector<std::vector<double> > im; for(int i = 0; i < res_y; i++) im.emplace_back(); - double step_x = (max_x - min_x)/(res_x - 1); double step_y = (max_y - min_y)/(res_y - 1); - - int num_pts = diagram.size(); - - for(int i = 0; i < res_y; i++){ - double y = min_y + i*step_y; - for(int j = 0; j < res_x; j++){ - double x = min_x + j*step_x; - - double pixel_value = 0; - for(int k = 0; k < num_pts; k++){ - double px = diagram[k].first; double py = diagram[k].second; - pixel_value += weight(std::pair<double,double>(px,py)) * std::exp( -((x-px)*(x-px) + (y-(py-px))*(y-(py-px))) / (2*sigma*sigma) ) / (sigma*std::sqrt(2*pi)); - } - im[i].push_back(pixel_value); - - } - } - - return im; - - } - - - - -}; // class Persistence_heat_maps_exact -} // namespace Persistence_representations -} // namespace Gudhi - -#endif // PERSISTENCE_HEAT_MAPS_EXACT_H_ diff --git a/src/Persistence_representations/include/gudhi/Persistence_landscape_on_grid.h b/src/Persistence_representations/include/gudhi/Persistence_landscape_on_grid.h index fd8a181c..db0e362a 100644 --- a/src/Persistence_representations/include/gudhi/Persistence_landscape_on_grid.h +++ b/src/Persistence_representations/include/gudhi/Persistence_landscape_on_grid.h @@ -986,7 +986,7 @@ void Persistence_landscape_on_grid::set_up_values_of_landscapes(const std::vecto for (size_t int_no = 0; int_no != p.size(); ++int_no) { size_t grid_interval_begin = (p[int_no].first - grid_min_) / dx; size_t grid_interval_end = (p[int_no].second - grid_min_) / dx; - size_t grid_interval_midpoint = (size_t)(0.5 * (grid_interval_begin + grid_interval_end)); + size_t grid_interval_midpoint = (size_t)(0.5 * (p[int_no].first + p[int_no].second) - grid_min + 1); if (dbg) { std::cerr << "Considering an interval : " << p[int_no].first << "," << p[int_no].second << std::endl; @@ -996,7 +996,7 @@ void Persistence_landscape_on_grid::set_up_values_of_landscapes(const std::vecto std::cerr << "grid_interval_midpoint : " << grid_interval_midpoint << std::endl; } - double landscape_value = dx; + double landscape_value = grid_min + dx * (grid_interval_begin + 1) - p[int_no].first; for (size_t i = grid_interval_begin + 1; i < grid_interval_midpoint; ++i) { if (dbg) { std::cerr << "Adding landscape value (going up) for a point : " << i << " equal : " << landscape_value @@ -1030,6 +1030,8 @@ void Persistence_landscape_on_grid::set_up_values_of_landscapes(const std::vecto } landscape_value += dx; } + + landscape_value = p[int_no].second - grid_min - dx * grid_interval_midpoint; for (size_t i = grid_interval_midpoint; i <= grid_interval_end; ++i) { if (landscape_value > 0) { if (number_of_levels != std::numeric_limits<unsigned>::max()) { diff --git a/src/Persistence_representations/include/gudhi/Persistence_landscape_on_grid_exact.h b/src/Persistence_representations/include/gudhi/Persistence_landscape_on_grid_exact.h deleted file mode 100644 index 52f24195..00000000 --- a/src/Persistence_representations/include/gudhi/Persistence_landscape_on_grid_exact.h +++ /dev/null @@ -1,108 +0,0 @@ -/* This file is part of the Gudhi Library. The Gudhi library - * (Geometric Understanding in Higher Dimensions) is a generic C++ - * library for computational topology. - * - * Author(s): Mathieu Carriere - * - * Copyright (C) 2018 INRIA (France) - * - * This program is free software: you can redistribute it and/or modify - * it under the terms of the GNU General Public License as published by - * the Free Software Foundation, either version 3 of the License, or - * (at your option) any later version. - * - * This program is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. - * - * You should have received a copy of the GNU General Public License - * along with this program. If not, see <http://www.gnu.org/licenses/>. - */ - -#ifndef LANDSCAPE_H_ -#define LANDSCAPE_H_ - -// gudhi include -#include <gudhi/read_persistence_from_file.h> -#include <gudhi/common_persistence_representations.h> -#include <gudhi/Debug_utils.h> - -// standard include -#include <cmath> -#include <iostream> -#include <vector> -#include <limits> -#include <fstream> -#include <sstream> -#include <algorithm> -#include <string> -#include <utility> -#include <functional> - -namespace Gudhi { -namespace Persistence_representations { - -/** - * \class Persistence_landscape_on_grid_exact gudhi/Persistence_landscape_on_grid_exact.h - * \brief A class implementing exact persistence landscapes by approximating them on a collection of grid points - * - * \ingroup Persistence_representations - * - * \details - * In this class, we propose a way to approximate landscapes by sampling the x-axis of the persistence diagram and evaluating the (exact) landscape functions on the sample projections onto the diagonal. Note that this is a different approximation scheme - * from the one proposed in Persistence_landscape_on_grid, which puts a grid on the diagonal, maps the persistence intervals on this grid and computes the (approximate) landscape functions on the samples. - * Hence, the difference is that we do not modify the diagram in this implementation, but the code can be slower to run. -**/ - -class Persistence_landscape_on_grid_exact { - - protected: - Persistence_diagram diagram; - int res_x, nb_ls; - double min_x, max_x; - - public: - - /** \brief Persistence_landscape_on_grid_exact constructor. - * \ingroup Persistence_landscape_on_grid_exact - * - * @param[in] _diagram persistence diagram. - * @param[in] _nb_ls number of landscape functions. - * @param[in] _min_x minimum value of samples. - * @param[in] _max_x maximum value of samples. - * @param[in] _res_x number of samples. - * - */ - Persistence_landscape_on_grid_exact(const Persistence_diagram & _diagram, int _nb_ls = 5, double _min_x = 0.0, double _max_x = 1.0, int _res_x = 10){diagram = _diagram; nb_ls = _nb_ls; min_x = _min_x; max_x = _max_x; res_x = _res_x;} - - /** \brief Computes the landscape approximation of a diagram. - * \ingroup Persistence_landscape_on_grid_exact - * - */ - std::vector<std::vector<double> > vectorize() const { - std::vector<std::vector<double> > ls; for(int i = 0; i < nb_ls; i++) ls.emplace_back(); - int num_pts = diagram.size(); double step = (max_x - min_x)/(res_x - 1); - - std::vector<std::vector<double> > ls_t; for(int i = 0; i < res_x; i++) ls_t.emplace_back(); - for(int j = 0; j < num_pts; j++){ - double px = diagram[j].first; double py = diagram[j].second; double mid = (px+py)/2; - int first = std::ceil((px-min_x)/step); int middle = std::ceil((mid-min_x)/step); int last = std::ceil((py-min_x)/step); double x = min_x + first*step; - for(int i = first; i < middle; i++){ double value = std::sqrt(2)*(x-px); ls_t[i].push_back(value); x += step; } - for(int i = middle; i < last; i++){ double value = std::sqrt(2)*(py-x); ls_t[i].push_back(value); x += step; } - } - - for(int i = 0; i < res_x; i++){ - std::sort(ls_t[i].begin(), ls_t[i].end(), [](const double & a, const double & b){return a > b;}); - int nb_events_i = ls_t[i].size(); - for (int j = 0; j < nb_ls; j++){ if(j < nb_events_i) ls[j].push_back(ls_t[i][j]); else ls[j].push_back(0); } - } - - return ls; - } - -}; // class Persistence_landscape_on_grid_exact -} // namespace Persistence_representations -} // namespace Gudhi - -#endif // LANDSCAPE_H_ diff --git a/src/Persistence_representations/include/gudhi/Persistence_weighted_gaussian.h b/src/Persistence_representations/include/gudhi/Persistence_weighted_gaussian.h deleted file mode 100644 index 9ef47bf1..00000000 --- a/src/Persistence_representations/include/gudhi/Persistence_weighted_gaussian.h +++ /dev/null @@ -1,182 +0,0 @@ -/* This file is part of the Gudhi Library. The Gudhi library - * (Geometric Understanding in Higher Dimensions) is a generic C++ - * library for computational topology. - * - * Author(s): Mathieu Carriere - * - * Copyright (C) 2018 INRIA (France) - * - * This program is free software: you can redistribute it and/or modify - * it under the terms of the GNU General Public License as published by - * the Free Software Foundation, either version 3 of the License, or - * (at your option) any later version. - * - * This program is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. - * - * You should have received a copy of the GNU General Public License - * along with this program. If not, see <http://www.gnu.org/licenses/>. - */ - -#ifndef PERSISTENCE_WEIGHTED_GAUSSIAN_H_ -#define PERSISTENCE_WEIGHTED_GAUSSIAN_H_ - -// gudhi include -#include <gudhi/read_persistence_from_file.h> -#include <gudhi/common_persistence_representations.h> -#include <gudhi/Weight_functions.h> - -// standard include -#include <cmath> -#include <iostream> -#include <vector> -#include <limits> -#include <fstream> -#include <sstream> -#include <algorithm> -#include <string> -#include <utility> -#include <functional> -#include <random> - -namespace Gudhi { -namespace Persistence_representations { -/** - * \class Persistence_weighted_gaussian gudhi/Persistence_weighted_gaussian.h - * \brief A class implementing the Persistence Weighted Gaussian kernel and a specific case thereof called the Persistence Scale Space kernel. - * - * \ingroup Persistence_representations - * - * \details - * The Persistence Weighted Gaussian kernel is built with Gaussian Kernel Mean Embedding, meaning that each persistence diagram is first - * sent to the Hilbert space of a Gaussian kernel with bandwidth parameter \f$\sigma >0\f$ using a weighted mean embedding \f$\Phi\f$: - * - * \f$ \Phi\,:\,D\,\rightarrow\,\sum_{p\in D}\,w(p)\,{\rm exp}\left(-\frac{\|p-\cdot\|_2^2}{2\sigma^2}\right) \f$, - * - * Usually, the weight function is chosen to be an arctan function of the distance of the point to the diagonal: - * \f$w(p) = {\rm arctan}(C\,|y-x|^\alpha)\f$, for some parameters \f$C,\alpha >0\f$. - * Then, their scalar product in this space is computed: - * - * \f$ k(D_1,D_2)=\langle\Phi(D_1),\Phi(D_2)\rangle - * \,=\,\sum_{p\in D_1}\,\sum_{q\in D_2}\,w(p)\,w(q)\,{\rm exp}\left(-\frac{\|p-q\|_2^2}{2\sigma^2}\right).\f$ - * - * Note that one may apply a second Gaussian kernel to their distance in this space and still get a kernel. - * - * It follows that the computation time is \f$O(n^2)\f$ where \f$n\f$ is the number of points - * in the diagrams. This time can be improved by computing approximations of the kernel - * with \f$m\f$ Fourier features \cite Rahimi07randomfeatures. In that case, the computation time becomes \f$O(mn)\f$. - * - * The Persistence Scale Space kernel is a Persistence Weighted Gaussian kernel between modified diagrams: - * the symmetric of each point with respect to the diagonal is first added in each diagram, and then the weight function - * is set to be +1 if the point is above the diagonal and -1 otherwise. - * - * For more details, please see \cite Kusano_Fukumizu_Hiraoka_PWGK - * and \cite Reininghaus_Huber_ALL_PSSK . - * -**/ -class Persistence_weighted_gaussian{ - - protected: - Persistence_diagram diagram; - Weight weight; - double sigma; - int approx; - - public: - - /** \brief Persistence Weighted Gaussian kernel constructor. - * \ingroup Persistence_weighted_gaussian - * - * @param[in] _diagram persistence diagram. - * @param[in] _sigma bandwidth parameter of the Gaussian kernel used for the Kernel Mean Embedding of the diagrams. - * @param[in] _approx number of random Fourier features in case of approximate computation, set to -1 for exact computation. - * @param[in] _weight weight function for the points in the diagrams. - * - */ - Persistence_weighted_gaussian(const Persistence_diagram & _diagram, double _sigma = 1.0, int _approx = 1000, const Weight & _weight = arctan_weight(1,1)){diagram = _diagram; sigma = _sigma; approx = _approx; weight = _weight;} - - - // ********************************** - // Utils. - // ********************************** - - std::vector<std::pair<double,double> > Fourier_feat(const Persistence_diagram & diag, const std::vector<std::pair<double,double> > & z, const Weight & weight = arctan_weight(1,1)) const { - int md = diag.size(); std::vector<std::pair<double,double> > b; int mz = z.size(); - for(int i = 0; i < mz; i++){ - double d1 = 0; double d2 = 0; double zx = z[i].first; double zy = z[i].second; - for(int j = 0; j < md; j++){ - double x = diag[j].first; double y = diag[j].second; - d1 += weight(diag[j])*cos(x*zx + y*zy); - d2 += weight(diag[j])*sin(x*zx + y*zy); - } - b.emplace_back(d1,d2); - } - return b; - } - - std::vector<std::pair<double,double> > random_Fourier(double sigma, int m = 1000) const { - std::normal_distribution<double> distrib(0,1); std::vector<std::pair<double,double> > z; std::random_device rd; - for(int i = 0; i < m; i++){ - std::mt19937 e1(rd()); std::mt19937 e2(rd()); - double zx = distrib(e1); double zy = distrib(e2); - z.emplace_back(zx/sigma,zy/sigma); - } - return z; - } - - - - // ********************************** - // Scalar product + distance. - // ********************************** - - /** \brief Evaluation of the kernel on a pair of diagrams. - * \ingroup Persistence_weighted_gaussian - * - * @pre sigma, approx and weight attributes need to be the same for both instances. - * @param[in] second other instance of class Persistence_weighted_gaussian. - * - */ - double compute_scalar_product(const Persistence_weighted_gaussian & second) const { - - GUDHI_CHECK(this->sigma != second.sigma || this->approx != second.approx || this->weight != second.weight, std::invalid_argument("Error: different values for representations")); - Persistence_diagram diagram1 = this->diagram; Persistence_diagram diagram2 = second.diagram; - - if(this->approx == -1){ - int num_pts1 = diagram1.size(); int num_pts2 = diagram2.size(); double k = 0; - for(int i = 0; i < num_pts1; i++) - for(int j = 0; j < num_pts2; j++) - k += this->weight(diagram1[i])*this->weight(diagram2[j])*exp(-((diagram1[i].first - diagram2[j].first) * (diagram1[i].first - diagram2[j].first) + - (diagram1[i].second - diagram2[j].second) * (diagram1[i].second - diagram2[j].second)) - /(2*this->sigma*this->sigma)); - return k; - } - else{ - std::vector<std::pair<double,double> > z = random_Fourier(this->sigma, this->approx); - std::vector<std::pair<double,double> > b1 = Fourier_feat(diagram1,z,this->weight); - std::vector<std::pair<double,double> > b2 = Fourier_feat(diagram2,z,this->weight); - double d = 0; for(int i = 0; i < this->approx; i++) d += b1[i].first*b2[i].first + b1[i].second*b2[i].second; - return d/this->approx; - } - } - - /** \brief Evaluation of the distance between images of diagrams in the Hilbert space of the kernel. - * \ingroup Persistence_weighted_gaussian - * - * @pre sigma, approx and weight attributes need to be the same for both instances. - * @param[in] second other instance of class Persistence_weighted_gaussian. - * - */ - double distance(const Persistence_weighted_gaussian & second) const { - GUDHI_CHECK(this->sigma != second.sigma || this->approx != second.approx || this->weight != second.weight, std::invalid_argument("Error: different values for representations")); - return std::pow(this->compute_scalar_product(*this) + second.compute_scalar_product(second)-2*this->compute_scalar_product(second), 0.5); - } - - -}; // class Persistence_weighted_gaussian -} // namespace Persistence_representations -} // namespace Gudhi - -#endif // PERSISTENCE_WEIGHTED_GAUSSIAN_H_ diff --git a/src/Persistence_representations/include/gudhi/Sliced_Wasserstein.h b/src/Persistence_representations/include/gudhi/Sliced_Wasserstein.h index d8ed0d98..8c92ab54 100644 --- a/src/Persistence_representations/include/gudhi/Sliced_Wasserstein.h +++ b/src/Persistence_representations/include/gudhi/Sliced_Wasserstein.h @@ -62,7 +62,7 @@ namespace Persistence_representations { * in the diagrams, or approximated by sampling \f$N\f$ lines in the circle in \f$O(Nn{\rm log}(n))\f$ time. The Sliced Wasserstein Kernel is then computed as: * * \f$ k(D_1,D_2) = {\rm exp}\left(-\frac{SW(D_1,D_2)}{2\sigma^2}\right).\f$ - * + * * For more details, please see \cite pmlr-v70-carriere17a . * **/ @@ -80,7 +80,7 @@ class Sliced_Wasserstein { void build_rep(){ if(approx > 0){ - + double step = pi/this->approx; int n = diagram.size(); @@ -188,7 +188,7 @@ class Sliced_Wasserstein { * \ingroup Sliced_Wasserstein * * @pre approx attribute needs to be the same for both instances. - * @param[in] second other instance of class Sliced_Wasserstein. + * @param[in] second other instance of class Sliced_Wasserstein. * * */ diff --git a/src/Persistence_representations/include/gudhi/Weight_functions.h b/src/Persistence_representations/include/gudhi/Weight_functions.h deleted file mode 100644 index 78de406d..00000000 --- a/src/Persistence_representations/include/gudhi/Weight_functions.h +++ /dev/null @@ -1,81 +0,0 @@ -/* This file is part of the Gudhi Library. The Gudhi library - * (Geometric Understanding in Higher Dimensions) is a generic C++ - * library for computational topology. - * - * Author(s): Mathieu Carriere - * - * Copyright (C) 2018 INRIA (France) - * - * This program is free software: you can redistribute it and/or modify - * it under the terms of the GNU General Public License as published by - * the Free Software Foundation, either version 3 of the License, or - * (at your option) any later version. - * - * This program is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. - * - * You should have received a copy of the GNU General Public License - * along with this program. If not, see <http://www.gnu.org/licenses/>. - */ - -#ifndef WEIGHT_FUNCTIONS_H_ -#define WEIGHT_FUNCTIONS_H_ - -// gudhi include -#include <gudhi/read_persistence_from_file.h> -#include <gudhi/common_persistence_representations.h> - -// standard include -#include <cmath> -#include <iostream> -#include <vector> -#include <limits> -#include <fstream> -#include <sstream> -#include <algorithm> -#include <string> -#include <utility> -#include <functional> - -namespace Gudhi { -namespace Persistence_representations { - -/** \fn static double pss_weight(std::pair<double,double> p) - * \brief Persistence Scale Space kernel weight function. - * \ingroup Persistence_representations - * - * @param[in] p point in 2D. - */ -static double pss_weight(std::pair<double,double> p) {if(p.second > p.first) return 1; else return -1;} - -/** \fn static double linear_weight(std::pair<double,double> p) - * \brief Linear weight function. - * \ingroup Persistence_representations - * - * @param[in] p point in 2D. - */ -static double linear_weight(std::pair<double,double> p) {return std::abs(p.second - p.first);} - -/** \fn static double const_weight(std::pair<double,double> p) - * \brief Constant weight function. - * \ingroup Persistence_representations - * - * @param[in] p point in 2D. - */ -static double const_weight(std::pair<double,double> p) {return 1;} - -/** \fn static std::function<double (std::pair<double,double>) > arctan_weight(double C, double alpha) - * \brief Returns the arctan weight function with parameters C and alpha. - * \ingroup Persistence_representations - * - * @param[in] C positive constant. - * @param[in] alpha positive power. - */ -static std::function<double (std::pair<double,double>) > arctan_weight(double C, double alpha) {return [=](std::pair<double,double> p){return C * atan(std::pow(std::abs(p.second - p.first), alpha));};} - -} // namespace Persistence_representations -} // namespace Gudhi - -#endif // WEIGHT_FUNCTIONS_H_ diff --git a/src/Persistence_representations/include/gudhi/common_persistence_representations.h b/src/Persistence_representations/include/gudhi/common_persistence_representations.h index 539eee60..024c99ec 100644 --- a/src/Persistence_representations/include/gudhi/common_persistence_representations.h +++ b/src/Persistence_representations/include/gudhi/common_persistence_representations.h @@ -40,12 +40,23 @@ static constexpr double pi = boost::math::constants::pi<double>(); /** * In this module, we use the name Persistence_diagram for the representation of a diagram in a vector of pairs of two double. */ -using Persistence_diagram = std::vector<std::pair<double,double> >; +using Persistence_diagram = std::vector<std::pair<double, double> >; /** * In this module, we use the name Weight for the representation of a function taking a pair of two double and returning a double. */ -using Weight = std::function<double (std::pair<double,double>) >; +using Weight = std::function<double (std::pair<double, double>) >; +using Kernel = std::function<double (std::pair<double, double>, std::pair<double, double> )>; + +Kernel Gaussian_kernel(double sigma){ + return [=](std::pair<double, double> p, std::pair<double, double> q){return std::exp( -((p.first-q.first)*(p.first-q.first) + (p.second-q.second)*(p.second-q.second)) / (sigma*sigma) );}; +} + +Kernel polynomial_kernel(double c, double d){ + return [=](std::pair<double, double> p, std::pair<double, double> q){return std::pow( p.first*q.first + p.second*q.second + c, d);}; +} + + // double epsi = std::numeric_limits<double>::epsilon(); double epsi = 0.000005; |