GitLab

* Fix LightGBM models locale sensitivity and improve R/W performance.

When Java is used, the default C++ locale is broken. This is true for
Java providers that use the C API or even Python models that require JEP.

This patch solves that issue making the model reads/writes insensitive
to such settings.
To achieve it, within the model read/write codebase:
 - C++ streams are imbued with the classic locale
 - Calls to functions that are dependent on the locale are replaced
 - The default locale is not changed!

This approach means:
 - The user's locale is never tampered with, avoiding issues such as
    https://github.com/microsoft/LightGBM/issues/2979 with the previous
    approach https://github.com/microsoft/LightGBM/pull/2891
 - Datasets can still be read according the user's locale
 - The model file has a single format independent of locale

Changes:
 - Add CommonC namespace which provides faster locale-independent versions of Common's methods
 - Model code makes conversions through CommonC
 - Cleanup unused Common methods
 - Performance improvements. Use fast libraries for locale-agnostic conversion:
   - value->string: https://github.com/fmtlib/fmt
   - string->double: https://github.com/lemire/fast_double_parser (10x
      faster double parsing according to their benchmark)

Bugfixes:
 - https://github.com/microsoft/LightGBM/issues/2500
 - https://github.com/microsoft/LightGBM/issues/2890
 - https://github.com/ninia/jep/issues/205

 (as it is related to LGBM as well)

* Align CommonC namespace

* Add new external_libs/ to python setup

* Try fast_double_parser fix #1

Testing commit e09e5aad828bcb16bea7ed0ed8322e019112fdbe

If it works it should fix more LGBM builds

* CMake: Attempt to link fmt without explicit PUBLIC tag

* Exclude external_libs from linting

* Add exernal_libs to MANIFEST.in

* Set dynamic linking option for fmt.

* linting issues

* Try to fix lint includes

* Try to pass fPIC with static fmt lib

* Try CMake P_I_C option with fmt library

* [R-package] Add CMake support for R and CRAN

* Cleanup CMakeLists

* Try fmt hack to remove stdout

* Switch to header-only mode

* Add PRIVATE argument to target_link_libraries

* use fmt in header-only mode

* Remove CMakeLists comment

* Change OpenMP to PUBLIC linking in Mac

* Update fmt submodule to 7.1.2

* Use fmt in header-only-mode

* Remove fmt from CMakeLists.txt

* Upgrade fast_double_parser to v0.2.0

* Revert "Add PRIVATE argument to target_link_libraries"

This reverts commit 3dd45dde7b92531b2530ab54522bb843c56227a7.

* Address James Lamb's comments

* Update R-package/.Rbuildignore
Co-authored-by: James Lamb <jaylamb20@gmail.com>

* Upgrade to fast_double_parser v0.3.0 - Solaris support

* Use legacy code only in Solaris

* Fix lint issues

* Fix comment

* Address StrikerRUS's comments (solaris ifdef).

* Change header guards
Co-authored-by: James Lamb <jaylamb20@gmail.com>

* add dummy gpu solver code

* initial GPU code

* fix crash bug

* first working version

* use asynchronous copy

* use a better kernel for root

* parallel read histogram

* sparse features now works, but no acceleration, compute on CPU

* compute sparse feature on CPU simultaneously

* fix big bug; add gpu selection; add kernel selection

* better debugging

* clean up

* add feature scatter

* Add sparse_threshold control

* fix a bug in feature scatter

* clean up debug

* temporarily add OpenCL kernels for k=64,256

* fix up CMakeList and definition USE_GPU

* add OpenCL kernels as string literals

* Add boost.compute as a submodule

* add boost dependency into CMakeList

* fix opencl pragma

* use pinned memory for histogram

* use pinned buffer for gradients and hessians

* better debugging message

* add double precision support on GPU

* fix boost version in CMakeList

* Add a README

* reconstruct GPU initialization code for ResetTrainingData

* move data to GPU in parallel

* fix a bug during feature copy

* update gpu kernels

* update gpu code

* initial port to LightGBM v2

* speedup GPU data loading process

* Add 4-bit bin support to GPU

* re-add sparse_threshold parameter

* remove kMaxNumWorkgroups and allows an unlimited number of features

* add feature mask support for skipping unused features

* enable kernel cache

* use GPU kernels withoug feature masks when all features are used

* REAdme.

* REAdme.

* update README

* fix typos (#349)

* change compile to gcc on Apple as default

* clean vscode related file

* refine api of constructing from sampling data.

* fix bug in the last commit.

* more efficient algorithm to sample k from n.

* fix bug in filter bin

* change to boost from average output.

* fix tests.

* only stop training when all classes are finshed in multi-class.

* limit the max tree output. change hessian in multi-class objective.

* robust tree model loading.

* fix test.

* convert the probabilities to raw score in boost_from_average of classification.

* fix the average label for binary classification.

* Add boost_from_average to docs (#354)

* don't use "ConvertToRawScore" for self-defined objective function.

* boost_from_average seems doesn't work well in binary classification. remove it.

* For a better jump link (#355)

* Update Python-API.md

* for a better jump in page

A space is needed between `#` and the headers content according to Github's markdown format [guideline](https://guides.github.com/features/mastering-markdown/)

After adding the spaces, we can jump to the exact position in page by click the link.

* fixed something mentioned by @wxchan

* Update Python-API.md

* add FitByExistingTree.

* adapt GPU tree learner for FitByExistingTree

* avoid NaN output.

* update boost.compute

* fix typos (#361)

* fix broken links (#359)

* update README

* disable GPU acceleration by default

* fix image url

* cleanup debug macro

* remove old README

* do not save sparse_threshold_ in FeatureGroup

* add details for new GPU settings

* ignore submodule when doing pep8 check

* allocate workspace for at least one thread during builing Feature4

* move sparse_threshold to class Dataset

* remove duplicated code in GPUTreeLearner::Split

* Remove duplicated code in FindBestThresholds and BeforeFindBestSplit

* do not rebuild ordered gradients and hessians for sparse features

* support feature groups in GPUTreeLearner

* Initial parallel learners with GPU support

* add option device, cleanup code

* clean up FindBestThresholds; add some omp parallel

* constant hessian optimization for GPU

* Fix GPUTreeLearner crash when there is zero feature

* use np.testing.assert_almost_equal() to compare lists of floats in tests

* travis for GPU