GitLab

* refine categorical split

* a bug fix

* fix a bug

* add network apis.

* support parallel loading dataset in c api.

* fix bug

* fix bug

* thread local for the network interface.

* fix bug.

* Bounds checking of used_row_indices (#959)

* add Common::check_int32_elements_interval_closed function.

* insert a bounds checking code in LGBM_DatasetGetSubset function of C API.

* Delete a needless function

* delete the function check_int32_elements_interval_closed

* Fix typos

* Bounds checking of used_row_indices (#959)

* add Common::check_int32_elements_interval_closed function.

* insert a bounds checking code in LGBM_DatasetGetSubset function of C API.

* Delete a needless function

* delete the function check_int32_elements_interval_closed

* Fix coding style (#969)

Function names must be in the "Pascal Case" style.

* check_elements_interval_closed to CheckElementsIntervalClosed

* obtain_min_max_sum to ObtainMinMaxSum

* Change functions in common.h into template functions (#969)

* CheckElementsIntervalClosed

* ObtainMinMaxSum

These two functions were changed into template functions.

* Remove an unpreferable overload

* remove an overload of the function ObtainMinMaxSum

* Use stringstream to format T type

Function names must be in the "Pascal Case" style.

* check_elements_interval_closed to CheckElementsIntervalClosed

* obtain_min_max_sum to ObtainMinMaxSum

* add get params

clean codes

* check duplicate params

* Revert "add get params"

This reverts commit ec1d8dd17aa2c83dd0d5f7716c59b6a6fb94102c.

* set priority by length & check duplicate

* rename function

* Update dataset.cpp

* Update dataset.cpp

* refine categorical split

* add test

* better log for "Check max_depth and num_leaves"

The default log said "[LightGBM] [Warning] Accuarcy may be bad since you didn't set num_leaves." but the case max_depth²>num_leaves could also trigger this log.
Correction of a typo too : Accuarcy > Accuracy

* Update config.cpp

While in parallel training, when one worker have 0 data, then it will not execute ConstructHistogram.
if so, ptr_smaller_leaf_hist_data will not be 0 but the old data. that will get the wrong Histogram, and the wrong split info.

* Fix feature attributions for regression models and add Python bindings

* Address pylint issue

* Lazy fix missing tree depth info

* Attempt to address issue #807 Poisson regression objective.

* Fix coding style.

* Added label non-negative safety check.

* expose feature importance to c_api

* support type=gain

* remove dump model from examples and tests temporarily because it's unstable

* use double instead of float

* output warning when only set `max_depth`

* Remove check of max_depth

max_depth can take both positive and negative value

* Explain individual predictions using SHAP value feature attributions

* Address code review