From 28c1da4226bee99f88b2a0e9a0c16a05568d5eab Mon Sep 17 00:00:00 2001
From: Thomas Purcell <purcell@fhi-berlin.mpg.de>
Date: Fri, 12 Jun 2020 20:36:30 +0200
Subject: [PATCH] Update past feature checks to use scores
limit the number of operations done
---
src/descriptor_identifier/Model/Model.cpp | 2 +-
src/descriptor_identifier/Model/Model.hpp | 4 +-
src/descriptor_identifier/SISSORegressor.cpp | 2 +-
.../feature_space/FeatureSpace.cpp | 65 +++++++++----------
.../feature_space/FeatureSpace.hpp | 8 +--
5 files changed, 38 insertions(+), 43 deletions(-)
diff --git a/src/descriptor_identifier/Model/Model.cpp b/src/descriptor_identifier/Model/Model.cpp
index 10eb8175..b8968f25 100644
--- a/src/descriptor_identifier/Model/Model.cpp
+++ b/src/descriptor_identifier/Model/Model.cpp
@@ -1,6 +1,6 @@
#include <descriptor_identifier/Model/Model.hpp>
-Model::Model(std::vector<double> prop_train, std::vector<double> prop_test, std::vector<std::shared_ptr<FeatureNode>> feats, std::vector<int> task_sizes_train, std::vector<int> task_sizes_test) :
+Model::Model(std::vector<double> prop_train, std::vector<double> prop_test, std::vector<node_ptr> feats, std::vector<int> task_sizes_train, std::vector<int> task_sizes_test) :
_n_samp_train(feats[0]->n_samp()),
_n_samp_test(feats[0]->n_test_samp()),
_n_dim(feats.size() + 1),
diff --git a/src/descriptor_identifier/Model/Model.hpp b/src/descriptor_identifier/Model/Model.hpp
index d660b6da..f52463d4 100644
--- a/src/descriptor_identifier/Model/Model.hpp
+++ b/src/descriptor_identifier/Model/Model.hpp
@@ -19,7 +19,7 @@ class Model
int _n_samp_test; //!< The number of test samples per feature
int _n_dim; //!< Dimension of the model
- std::vector<std::shared_ptr<FeatureNode>> _feats; //!< List of features in the model
+ std::vector<node_ptr> _feats; //!< List of features in the model
std::vector<std::vector<double>> _coefs; //!< Coefficients for teh features
std::vector<double> _prop_train; //!< The property to be modeled
@@ -41,7 +41,7 @@ public:
* @param prop The property
* @param feats The features for the model
*/
- Model(std::vector<double> prop_train, std::vector<double> prop_test, std::vector<std::shared_ptr<FeatureNode>> feats, std::vector<int> task_sizes_train, std::vector<int> task_sizes_test);
+ Model(std::vector<double> prop_train, std::vector<double> prop_test, std::vector<node_ptr> feats, std::vector<int> task_sizes_train, std::vector<int> task_sizes_test);
/**
diff --git a/src/descriptor_identifier/SISSORegressor.cpp b/src/descriptor_identifier/SISSORegressor.cpp
index 77b35c2a..0799aaed 100644
--- a/src/descriptor_identifier/SISSORegressor.cpp
+++ b/src/descriptor_identifier/SISSORegressor.cpp
@@ -176,7 +176,7 @@ void SISSORegressor::l0_norm(std::vector<double>& prop, int n_dim)
inds = util_funcs::argsort(all_min_error);
- std::vector<std::shared_ptr<FeatureNode>> min_nodes(n_dim);
+ std::vector<node_ptr> min_nodes(n_dim);
std::vector<Model> models;
for(int rr = 0; rr < _n_residual; ++rr)
diff --git a/src/feature_creation/feature_space/FeatureSpace.cpp b/src/feature_creation/feature_space/FeatureSpace.cpp
index c740c4f6..e5e64bb7 100644
--- a/src/feature_creation/feature_space/FeatureSpace.cpp
+++ b/src/feature_creation/feature_space/FeatureSpace.cpp
@@ -295,8 +295,15 @@ void FeatureSpace::generate_feature_space()
_n_feat = _phi.size();
}
-void FeatureSpace::project_generated(double* prop, int size, std::vector<std::shared_ptr<FeatureNode>>& phi_sel, std::vector<double>& scores_sel, std::vector<double>& scores_comp)
+void FeatureSpace::project_generated(double* prop, int size, std::vector<node_ptr>& phi_sel, std::vector<double>& scores_sel, std::vector<double>& scores_comp)
{
+ std::vector<double> scores_prev_sel;
+ if(node_value_arrs::N_SELECTED > _n_sis_select)
+ {
+ scores_prev_sel.resize(_phi_selected.size());
+ _project(prop, scores_prev_sel.data(), _phi_selected, _task_sizes, size / _n_samp);
+ }
+
for(auto feat = _phi.begin() + _start_gen.back() + _mpi_comm->rank(); feat < _phi.end(); feat += _mpi_comm->size())
{
std::fill_n(node_value_arrs::TEMP_STORAGE_REG.data(), node_value_arrs::TEMP_STORAGE_REG.size(), -1);
@@ -322,11 +329,11 @@ void FeatureSpace::project_generated(double* prop, int size, std::vector<std::sh
bool is_valid = valid_score_against_current(end_check, generated_phi[inds[ii]]->value_ptr(), scores[inds[ii]], scores_sel, scores_comp);
// Check the feature against those selected from previous SIS iterations
if((node_value_arrs::N_SELECTED > _n_sis_select) && is_valid)
- is_valid = valid_score_against_past(generated_phi[inds[ii]]->value_ptr(), scores_comp);
+ is_valid = valid_score_against_past(_phi[inds[ii]]->value_ptr(), _scores[inds[ii]], scores_prev_sel, scores_comp);
if(is_valid)
{
- std::shared_ptr<FeatureNode> new_feat = std::make_shared<FeatureNode>(node_value_arrs::N_SELECTED - _n_sis_select + end_check, generated_phi[inds[ii]]->expr(), generated_phi[inds[ii]]->value(), generated_phi[inds[ii]]->test_value(), generated_phi[inds[ii]]->unit(), true);
+ node_ptr new_feat = std::make_shared<FeatureNode>(node_value_arrs::N_SELECTED - _n_sis_select + end_check, generated_phi[inds[ii]]->expr(), generated_phi[inds[ii]]->value(), generated_phi[inds[ii]]->test_value(), generated_phi[inds[ii]]->unit(), true);
phi_sel.insert(phi_sel.begin() + end_check, new_feat);
scores_sel.insert(scores_sel.begin() + end_check, cur_score);
for(int jj = end_check + 1; jj < _n_sis_select; ++jj)
@@ -345,19 +352,17 @@ void FeatureSpace::project_generated(double* prop, int size, std::vector<std::sh
}
}
-bool FeatureSpace::valid_score_against_past(double* val_ptr, std::vector<double>& scores_comp)
+bool FeatureSpace::valid_score_against_past(double* val_ptr, double cur_score, std::vector<double> scores_past, std::vector<double>& scores_comp)
{
- double cur_feat_mean = util_funcs::mean(val_ptr, _n_samp);
- double cur_feat_std = util_funcs::stand_dev(val_ptr, _n_samp);
-
- std::transform(val_ptr, val_ptr + _n_samp, val_ptr, [&cur_feat_mean, &cur_feat_std](double val){return (val - cur_feat_mean) / cur_feat_std;});
-
- dgemv_('T', _n_samp, scores_comp.size(), 1.0 / static_cast<double>(_n_samp), node_value_arrs::D_MATRIX.data(), _n_samp, val_ptr, 1, 0.0, scores_comp.data(), 1);
+ std::transform(scores_past.begin(), scores_past.end(), scores_comp.begin(), [&cur_score](double score){return cur_score - score;});
- if(1.0 - util_funcs::max_abs_val<double>(scores_comp.data(), scores_comp.size()) < 1e-13)
- return false;
-
- std::transform(val_ptr, val_ptr + _n_samp, val_ptr, [&cur_feat_mean, &cur_feat_std](double val){return val * cur_feat_std + cur_feat_mean;});
+ // If two scores are the same then they are possibly the same feature, if not then they can't be
+ if(*std::min_element(scores_comp.begin(), scores_comp.end()) < 1e-10)
+ {
+ int dd = std::min_element(scores_comp.begin(), scores_comp.end()) - scores_comp.begin();
+ if(1.0 - std::abs(util_funcs::r(node_value_arrs::get_d_matrix_ptr(dd), val_ptr, _n_samp)) < 1e-13)
+ return false;
+ }
return true;
}
@@ -377,26 +382,14 @@ bool FeatureSpace::valid_score_against_current(int end_check, double* val_ptr, d
void FeatureSpace::sis(std::vector<double>& prop)
{
- std::vector<double> means(node_value_arrs::N_SELECTED);
- std::vector<double> stand_devs(node_value_arrs::N_SELECTED);
std::vector<double> scores_comp(std::max(node_value_arrs::N_SELECTED, _n_sis_select), 1.0);
-
std::vector<double> scores_sel(_n_sis_select, 0.0);
- std::vector<std::shared_ptr<FeatureNode>> phi_sel;
+
+ std::vector<node_ptr> phi_sel;
phi_sel.reserve(_n_sis_select);
int cur_feat = node_value_arrs::N_SELECTED;
- // Standardize the description matrix
- if(cur_feat > 0)
- {
- for(int dd = 0; dd < cur_feat; ++dd)
- {
- means[dd] = util_funcs::mean(node_value_arrs::get_d_matrix_ptr(dd), _n_samp);
- stand_devs[dd] = util_funcs::stand_dev(node_value_arrs::get_d_matrix_ptr(dd), _n_samp);
- std::transform(node_value_arrs::get_d_matrix_ptr(dd), node_value_arrs::get_d_matrix_ptr(dd) + _n_samp, node_value_arrs::get_d_matrix_ptr(dd), [&means, &stand_devs, &dd](double val){return (val - means[dd]) / stand_devs[dd];});
- }
- }
node_value_arrs::resize_d_matrix_arr(_n_sis_select);
_phi_selected.reserve(_phi_selected.size() + _n_sis_select);
@@ -408,12 +401,19 @@ void FeatureSpace::sis(std::vector<double>& prop)
int cur_feat_local = 0;
double cur_score = 0.0;
+ std::vector<double> scores_prev_sel;
+ if(node_value_arrs::N_SELECTED > _n_sis_select)
+ {
+ scores_prev_sel.resize(_phi_selected.size());
+ _project(prop.data(), scores_prev_sel.data(), _phi_selected, _task_sizes, prop.size() / _n_samp);
+ }
+
while((cur_feat_local != _n_sis_select) && (ii < _scores.size()))
{
bool is_valid = valid_score_against_current(cur_feat_local, _phi[inds[ii]]->value_ptr(), _scores[inds[ii]], scores_sel, scores_comp);
// Check the feature against those selected from previous SIS iterations
if(cur_feat > 0 && is_valid)
- is_valid = valid_score_against_past(_phi[inds[ii]]->value_ptr(), scores_comp);
+ is_valid = valid_score_against_past(_phi[inds[ii]]->value_ptr(), _scores[inds[ii]], scores_prev_sel, scores_comp);
if(is_valid)
{
@@ -431,11 +431,6 @@ void FeatureSpace::sis(std::vector<double>& prop)
project_generated(prop.data(), prop.size(), phi_sel, scores_sel, scores_comp);
}
- // Unstandardize the description matrix
- if(cur_feat > 0)
- for(int dd = 0; dd < cur_feat; ++dd)
- std::transform(node_value_arrs::get_d_matrix_ptr(dd), node_value_arrs::get_d_matrix_ptr(dd) + _n_samp, node_value_arrs::get_d_matrix_ptr(dd), [&means, &stand_devs, &dd](double val){return val * stand_devs[dd] + means[dd];});
-
phi_sel.resize(_n_sis_select);
scores_sel.resize(_n_sis_select);
@@ -489,7 +484,7 @@ void FeatureSpace::sis(std::vector<double>& prop)
if(_mpi_comm->rank() == 0)
{
std::vector<double> sent_scores(_n_sis_select * _mpi_comm->size(), 0.0);
- std::vector<std::shared_ptr<FeatureNode>> sent_phi(_n_sis_select * _mpi_comm->size());
+ std::vector<node_ptr> sent_phi(_n_sis_select * _mpi_comm->size());
std::copy_n(scores_sel.begin(), _n_sis_select, sent_scores.begin());
std::copy_n(phi_sel.begin(), _n_sis_select, sent_phi.begin());
diff --git a/src/feature_creation/feature_space/FeatureSpace.hpp b/src/feature_creation/feature_space/FeatureSpace.hpp
index 50a4d1e9..1299afcf 100644
--- a/src/feature_creation/feature_space/FeatureSpace.hpp
+++ b/src/feature_creation/feature_space/FeatureSpace.hpp
@@ -20,7 +20,7 @@
*/
class FeatureSpace
{
- std::vector<std::shared_ptr<FeatureNode>> _phi_selected; //!< selected features
+ std::vector<node_ptr> _phi_selected; //!< selected features
std::vector<node_ptr> _phi; //!< all features
std::vector<node_ptr> _phi_0; //!< initial feature space
@@ -82,7 +82,7 @@ public:
/**
* @brief Accessor function for _phi_selected
*/
- inline std::vector<std::shared_ptr<FeatureNode>> phi_selected(){return _phi_selected;};
+ inline std::vector<node_ptr> phi_selected(){return _phi_selected;};
/**
* @brief Accessor function for _phi
@@ -108,9 +108,9 @@ public:
void generate_new_feats(std::vector<node_ptr>::iterator& feat, std::vector<node_ptr>& feat_set, int& feat_ind, double l_bound=1e-50, double u_bound=1e50);
- void project_generated(double* prop, int size, std::vector<std::shared_ptr<FeatureNode>>& phi_selected, std::vector<double>& scores_selected, std::vector<double>& scores_comp);
+ void project_generated(double* prop, int size, std::vector<node_ptr>& phi_selected, std::vector<double>& scores_selected, std::vector<double>& scores_comp);
- bool valid_score_against_past(double* val_ptr, std::vector<double>& scores_comp);
+ bool valid_score_against_past(double* val_ptr, double cur_score, std::vector<double> scores_past, std::vector<double>& scores_comp);
bool valid_score_against_current(int end_check, double* val_ptr, double cur_score, std::vector<double>& scores_selected, std::vector<double>& scores_comp);
/**
--
GitLab