Transpile BDTs to C++ code.
Generate C++ representations of boosted decision trees
This project tries to provide a generic functionality to transpile trained BDTs into minimal, efficient C++ functions to evaluate single vectors of features.
While many frameworks exist to train, evaluate and store BDTs, its often hard to use the results in a productive manner.
So far there is only python3 support. Run
pip install bdt2cpp
to install the latest tagged version or
pip install git+https://github.com/bixel/bdt2cpp.git
for the current master version.
If you want to use bdt2cpp on CERNs lxplus machines, you need to get hold of minimum python3.6. According to CERNs Service Article KB0000730, one way to install the tool is:
# On lxplus scl enable rh-python36 # this will install bdt2cpp to your `~/.local/` directory pip install --user bdt2cpp
To generate a minimal Makefile together with the C++ code inside a
directory from a given XGBoost dump or TMVA
.xml file, simply run
You will find the corresponding files within the
build/ directory and if you
clang, you can simply
cd build make
Note for CERN Users: Currently, the Makefile uses clang as the default
compiler. You might need to adjust that in the generated file (inside the
The generated executable is essentially a very minimal placeholder, if you had 3 input features you could quickly cross-check the predictions against the original training framework:
cd build ./main 1 2 3
should give the same output as received within the training framework if a
f = (1, 2, 3) is evaluated.
To see the complete list of features with some explanations, run
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size & hash||File type||Python version||Upload date|
|bdt2cpp-0.2.0-py2.py3-none-any.whl (10.4 kB) View hashes||Wheel||py2.py3|
|bdt2cpp-0.2.0.tar.gz (7.2 kB) View hashes||Source||None|
Hashes for bdt2cpp-0.2.0-py2.py3-none-any.whl