Compiler for LightGBM gradient-boosted trees, based on LLVM. Speeds up prediction by ≥10x.
Full Changelog: https://github.com/siboehm/lleaves/compare/1.2.0...1.2.1
Full Changelog: https://github.com/siboehm/lleaves/compare/1.0.0...1.1.0
FIRST MAJOR VERSION RELEASE. There are actually no big changes in this version, but lleaves has been so stable over the past 2 years that not calling it v1 was just me chickening out 🐓
Minor changes:
llvm::fptosi
by @siboehm in https://github.com/siboehm/lleaves/pull/23. This broke categorical predictions when NaNs occurred, but only on ARM arch.Full Changelog: https://github.com/siboehm/lleaves/compare/0.2.6...0.2.7
Minor new feature: Allow specification of the root function's name in the compiled binary. This enables linking against multiple lleaves-compiled trees. Thanks @fuyw!
Full Changelog: https://github.com/siboehm/lleaves/compare/0.2.5...0.2.6
pandas_categorical:null
. Thanks @chenglin!Bugfix release: Removes an accidentally introduced dependency on pandas
.
fblocksize
, finline
, fcodemodel
.raw_score
, equivalent to the raw_score
parameter of LightGBM's Booster.predict()
.Add support for multiclass prediction.