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.

Installation

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.

lxplus

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

Usage

To generate a minimal Makefile together with the C++ code inside a build/ directory from a given XGBoost dump or TMVA .xml file, simply run

bdt2cpp my-bdt-dump.xgb

You will find the corresponding files within the build/ directory and if you have installed 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 build/ directory)

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 feature vector f = (1, 2, 3) is evaluated.

To see the complete list of features with some explanations, run

bdt2cpp -h