zlpmichelle / xgb2cpp

Converts XGBoost raw model dumps to C++ code.

xgb2cpp

xgb2cpp converts XGBoost raw text dump to C++ code that you can use in your C++ projects.

READ FIRST

This code (atm) only works on XGB models trained with 'multi:softprob' objective. 'multi:softprob' objective tells XGBoost to train trees in "1 vs. all" fashion. If you want it to work with 'binary:logistic' or 'binary:logitraw' you'll have to tweak code-generating Python script a bit or wait for me to implement it.

How to run it?

Use generate_cpp_code.py script to generate xgboost_classifier.cpp. Generated xgboost_classifier.cpp will contain xgb_classify() function that you can use to classify samples. In folder sample you can find main.cpp and xgboost_classifier.h files that form use case project.

Also, there is a pima-indian-diabetes dataset that I used to generate dummy model. File /sample/pima-indians-diabetes/train.py trains XGB model and creates dump.raw.txt which is raw text representation of trained model.

Example run:

cd sample/pima-indians-diabetes
# Train model and generate dump.raw.txt
python train.py
cd ../../
# Generate xgboost_classifier.cpp
python python generate_cpp_code.py --num_classes=2 --xgb_dump=sample/pima-indians-diabetes/dump.raw.txt
# Copy xgboost_classifier.cpp to sample project
cp xgboost_classifier.cpp sample/codeblocks_project/xgboost_classifier.cpp
# Now build and run main.cpp, xgboost_classifier.cpp and xgboost_classifier.h in sample/codeblocks_project

If you have any questions or improvements just open issue and I'll be on it. Thanks