LGBM inference using C++ API (Python/C++ Interface)

This repository provides sample scripts which demonstrate the use of the C++ API provided by the LGBM package in order to perform inference without using the Python package or its 3rd party dependencies (numpy, scipy, etc...).

This approach has minimal dependencies and is ideal for environments with limited resources such as mobile and IoT devices.

For reference both Python and C++ interface options are provided. One can access the C++ API both from Python and C++ scripts, depending on the environment of choice.

Python Code Explanation

The code in this project is inspired from the official repository

The linking between C++ and Python is presented in the c_interface.py file. This file contains the needed functions to use the C++ library through Python.

This example uses a model trained on the Iris dataset on a normal python environment. A sample script to train the model is provided in train.py. After training is complete you can save the model by calling:

# Assuming the model instance is called model
model.save_model('saved_model.txt')

We can now perform inference by pointing to the saved model through the C++ API. In order to use the library we have to convert the python variables into the corresponding C++ variables. The functions c_array and c_str take care of that. The run_booster wrapper function integrates the logic so that we can feed python data directly without having to deal with the conversions running under the hood.

Predicting

Executing the predict.py will read the data from a test_data.csv file, and using the saved model will perform predictions which will be saved in the output.csv file.

Testing

In order to test the predictions between the python package and the C++ API execute the test.py which will load python predictions from a test_predictions.csv file and then compare them with predictions inferred from the test_data.csv

C++ Code Explanation

This tutorial provides the steps to:

• build the LGBM Package from source
• Build an executable that performs inference using a pre-trained model

Instructions for building the LGBM Package from source

Follow the steps as provided in the Official Instalation Guide

git clone --recursive https://github.com/microsoft/LightGBM ; cd LightGBM
mkdir build ; cd build
cmake ..
make -j4

The above code will build the LGBM Package on a local folder called LightGBM.

Instructions for building executable for inference (prediction)

In order to build the executable we have to link the Include folder as well as the main library. These are provided with the -I,-L and -l arguments:

• -I{path_to_LightGBM}/include
• -L{path_to_LightGBM}
• -l_lightgbm

For example if the user is on the cpp directory of this repo you can use the following command to build an executable with the name lgbm_predict

g++ main_c.cpp -I${PWD}/LightGBM/include -L${PWD}/LightGBM -l_lightgbm -o lgbm_predict

Then in order to run the executable first export the library path to your env:

export LD_LIBRARY_PATH=${PWD}/LightGBM/

Then run the executable:

./lgbm_predict

Exploration of the code

The main_c.cpp file is used to load the model (the model in this repo is saved_model.txt) and perform predictions. The main function for predictions called predict is responsible for:

• Reading the data from a csv file (here test_data.csv)
• Loading the pre-trained model
• Performing the prediction and saving the result to a csv file (here c_test_predictions.csv)