Skip to content

Commit 4dc2de3

Browse files
mathias-von-ottenbreitmathias-von-ottenbreit
committed
fixed small bug
1 parent ec8ae9a commit 4dc2de3

File tree

5 files changed

+82
-61
lines changed

5 files changed

+82
-61
lines changed

‎.github/workflows/build_wheels_linux.yml‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -11,13 +11,13 @@ jobs:
1111
os: [ubuntu-20.04]
1212

1313
steps:
14-
- uses: actions/checkout@v2
14+
- uses: actions/checkout@v3
1515

1616
- name: Build wheels
17-
uses: pypa/cibuildwheel@v2.5.0
17+
uses: pypa/cibuildwheel@v2.11.3
1818
env:
1919
CIBW_SKIP: "*musllinux*"
2020

21-
- uses: actions/upload-artifact@v2
21+
- uses: actions/upload-artifact@v3
2222
with:
2323
path: ./wheelhouse/*.whl

‎.github/workflows/build_wheels_macos.yml‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -11,13 +11,13 @@ jobs:
1111
os: [macos-latest]
1212

1313
steps:
14-
- uses: actions/checkout@v2
14+
- uses: actions/checkout@v3
1515

1616
- name: Build wheels
17-
uses: pypa/cibuildwheel@v2.5.0
17+
uses: pypa/cibuildwheel@v2.11.3
1818
env:
1919
MACOSX_DEPLOYMENT_TARGET: "10.14"
2020

21-
- uses: actions/upload-artifact@v2
21+
- uses: actions/upload-artifact@v3
2222
with:
2323
path: ./wheelhouse/*.whl

‎.github/workflows/build_wheels_windows.yml‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -11,11 +11,11 @@ jobs:
1111
os: [windows-2019]
1212

1313
steps:
14-
- uses: actions/checkout@v2
14+
- uses: actions/checkout@v3
1515

1616
- name: Build wheels
17-
run: pipx run cibuildwheel==2.5.0
17+
run: pipx run cibuildwheel==2.11.3
1818

19-
- uses: actions/upload-artifact@v2
19+
- uses: actions/upload-artifact@v3
2020
with:
2121
path: ./wheelhouse/*.whl

‎cpp/APLRRegressor.h‎

Lines changed: 72 additions & 51 deletions
Original file line numberDiff line numberDiff line change
@@ -64,8 +64,10 @@ class APLRRegressor
6464
void add_promising_interactions_and_select_the_best_one();
6565
void update_intercept(size_t boosting_step);
6666
void select_the_best_term_and_update_errors(size_t boosting_step);
67+
void update_terms(size_t boosting_step);
6768
void update_gradient_and_errors();
6869
void add_new_term(size_t boosting_step);
70+
void calculate_and_validate_validation_error(size_t boosting_step);
6971
void update_term_eligibility();
7072
void print_summary_after_boosting_step(size_t boosting_step);
7173
void update_coefficients_for_all_steps();
@@ -505,9 +507,12 @@ void APLRRegressor::execute_boosting_steps()
505507
void APLRRegressor::execute_boosting_step(size_t boosting_step)
506508
{
507509
update_intercept(boosting_step);
508-
find_best_split_for_each_eligible_term();
509-
consider_interactions();
510-
select_the_best_term_and_update_errors(boosting_step);
510+
if(!abort_boosting)
511+
{
512+
find_best_split_for_each_eligible_term();
513+
consider_interactions();
514+
select_the_best_term_and_update_errors(boosting_step);
515+
}
511516
if(abort_boosting) return;
512517
update_term_eligibility();
513518
print_summary_after_boosting_step(boosting_step);
@@ -522,10 +527,29 @@ void APLRRegressor::update_intercept(size_t boosting_step)
522527
intercept_update=v*(neg_gradient_current.array()*sample_weight_train.array()).sum()/sample_weight_train.array().sum();
523528
linear_predictor_update=VectorXd::Constant(neg_gradient_current.size(),intercept_update);
524529
linear_predictor_update_validation=VectorXd::Constant(y_validation.size(),intercept_update);
525-
intercept+=intercept_update;
526-
intercept_steps[boosting_step]=intercept;
527530
update_linear_predictor_and_predictors();
528531
update_gradient_and_errors();
532+
calculate_and_validate_validation_error(boosting_step);
533+
if(!abort_boosting)
534+
{
535+
intercept+=intercept_update;
536+
intercept_steps[boosting_step]=intercept;
537+
}
538+
}
539+
540+
void APLRRegressor::update_linear_predictor_and_predictors()
541+
{
542+
linear_predictor_current+=linear_predictor_update;
543+
linear_predictor_current_validation+=linear_predictor_update_validation;
544+
predictions_current=transform_linear_predictor_to_predictions(linear_predictor_current,link_function,tweedie_power);
545+
predictions_current_validation=transform_linear_predictor_to_predictions(linear_predictor_current_validation,link_function,tweedie_power);
546+
}
547+
548+
void APLRRegressor::update_gradient_and_errors()
549+
{
550+
neg_gradient_current=calculate_neg_gradient_current();
551+
neg_gradient_nullmodel_errors=calculate_errors(neg_gradient_current,linear_predictor_null_model,sample_weight_train);
552+
neg_gradient_nullmodel_errors_sum=calculate_sum_error(neg_gradient_nullmodel_errors);
529553
}
530554

531555
void APLRRegressor::find_best_split_for_each_eligible_term()
@@ -745,61 +769,46 @@ void APLRRegressor::select_the_best_term_and_update_errors(size_t boosting_step)
745769
if(no_improvement)
746770
{
747771
abort_boosting=true;
748-
return;
749772
}
750773
else
751774
{
752775
update_linear_predictor_and_predictors();
753776
update_gradient_and_errors();
754-
755-
//Has the term been entered into the model before?
756-
if(terms.size()==0) //If nothing is in the model add the term
757-
add_new_term(boosting_step);
758-
else //If at least one term was added before
759-
{
760-
//Searching in existing terms
761-
bool found{false};
762-
for (size_t j = 0; j < terms.size(); ++j)
763-
{
764-
if(terms[j]==terms_eligible_current[best_term]) //if term was found, update coefficient and coefficient_steps
765-
{
766-
terms[j].coefficient+=terms_eligible_current[best_term].coefficient;
767-
terms[j].coefficient_steps[boosting_step]=terms[j].coefficient;
768-
found=true;
769-
break;
770-
}
771-
}
772-
//term was not in the model and is added to the model
773-
if(!found)
774-
{
775-
add_new_term(boosting_step);
776-
}
777-
}
778-
}
779-
780-
validation_error_steps[boosting_step]=calculate_mean_error(calculate_errors(y_validation,predictions_current_validation,sample_weight_validation,family,tweedie_power),sample_weight_validation);
781-
bool validation_error_is_invalid{std::isinf(validation_error_steps[boosting_step])};
782-
if(validation_error_is_invalid)
783-
{
784-
abort_boosting=true;
785-
std::string warning_message{"Warning: Encountered numerical problems when calculating prediction errors in the previous boosting step. Not continuing with further boosting steps. One potential reason is if the combination of family and link_function is invalid."};
786-
std::cout<<warning_message<<"\n";
777+
double backup_of_validation_error{validation_error_steps[boosting_step]};
778+
calculate_and_validate_validation_error(boosting_step);
779+
if(abort_boosting)
780+
validation_error_steps[boosting_step]=backup_of_validation_error;
781+
else
782+
update_terms(boosting_step);
787783
}
788784
}
789785

790-
void APLRRegressor::update_linear_predictor_and_predictors()
786+
void APLRRegressor::update_terms(size_t boosting_step)
791787
{
792-
linear_predictor_current+=linear_predictor_update;
793-
linear_predictor_current_validation+=linear_predictor_update_validation;
794-
predictions_current=transform_linear_predictor_to_predictions(linear_predictor_current,link_function,tweedie_power);
795-
predictions_current_validation=transform_linear_predictor_to_predictions(linear_predictor_current_validation,link_function,tweedie_power);
796-
}
797-
798-
void APLRRegressor::update_gradient_and_errors()
799-
{
800-
neg_gradient_current=calculate_neg_gradient_current();
801-
neg_gradient_nullmodel_errors=calculate_errors(neg_gradient_current,linear_predictor_null_model,sample_weight_train);
802-
neg_gradient_nullmodel_errors_sum=calculate_sum_error(neg_gradient_nullmodel_errors);
788+
bool no_term_is_in_model{terms.size()==0};
789+
if(no_term_is_in_model)
790+
add_new_term(boosting_step);
791+
else
792+
{
793+
//Searching in existing terms
794+
bool found{false};
795+
for (size_t j = 0; j < terms.size(); ++j)
796+
{
797+
bool term_is_already_in_model{terms[j]==terms_eligible_current[best_term]};
798+
if(term_is_already_in_model)
799+
{
800+
terms[j].coefficient+=terms_eligible_current[best_term].coefficient;
801+
terms[j].coefficient_steps[boosting_step]=terms[j].coefficient;
802+
found=true;
803+
break;
804+
}
805+
}
806+
//term was not in the model and is added to the model
807+
if(!found)
808+
{
809+
add_new_term(boosting_step);
810+
}
811+
}
803812
}
804813

805814
void APLRRegressor::add_new_term(size_t boosting_step)
@@ -812,6 +821,18 @@ void APLRRegressor::add_new_term(size_t boosting_step)
812821
terms.push_back(Term(terms_eligible_current[best_term]));
813822
}
814823

824+
void APLRRegressor::calculate_and_validate_validation_error(size_t boosting_step)
825+
{
826+
validation_error_steps[boosting_step]=calculate_mean_error(calculate_errors(y_validation,predictions_current_validation,sample_weight_validation,family,tweedie_power),sample_weight_validation);
827+
bool validation_error_is_invalid{std::isinf(validation_error_steps[boosting_step])};
828+
if(validation_error_is_invalid)
829+
{
830+
abort_boosting=true;
831+
std::string warning_message{"Warning: Encountered numerical problems when calculating prediction errors in the previous boosting step. Not continuing with further boosting steps. One potential reason is if the combination of family and link_function is invalid."};
832+
std::cout<<warning_message<<"\n";
833+
}
834+
}
835+
815836
void APLRRegressor::update_term_eligibility()
816837
{
817838
number_of_eligible_terms=terms_eligible_current.size();

‎setup.py‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -15,7 +15,7 @@
1515

1616
setuptools.setup(
1717
name='aplr',
18-
version='1.6.0',
18+
version='1.6.1',
1919
description='Automatic Piecewise Linear Regression',
2020
ext_modules=[sfc_module],
2121
author="Mathias von Ottenbreit",

0 commit comments

Comments
 (0)