github.com/Xtra-Computing/thundergbm

Documentations | Installation | Parameters | Python (scikit-learn) interface

What's new?

ThunderGBM won 2019 Best Paper Award from IEEE Transactions on Parallel and Distributed Systems by the IEEE Computer Society Publications Board (1 out of 987 submissions, for the work "Zeyi Wen^, Jiashuai Shi*, Bingsheng He, Jian Chen, Kotagiri Ramamohanarao, and Qinbin Li*, Exploiting GPUs for Efficient Gradient Boosting Decision Tree Training , IEEE Transactions on Parallel and Distributed Systems, vol. 30, no. 12, 2019, pp. 2706-2717."). see more details: Best Paper Award Winners from IEEE, News from NUS School of Computing

Overview

The mission of ThunderGBM is to help users easily and efficiently apply GBDTs and Random Forests to solve problems. ThunderGBM exploits GPUs to achieve high efficiency. Key features of ThunderGBM are as follows.

• Often by 10x times over other libraries.
• Support Python (scikit-learn) interfaces.
• Supported Operating System(s): Linux and Windows.
• Support classification, regression and ranking.

Why accelerate GBDT and Random Forests: A survey conducted by Kaggle in 2017 shows that 50%, 46% and 24% of the data mining and machine learning practitioners are users of Decision Trees, Random Forests and GBMs, respectively.

GBDTs and Random Forests are often used for creating state-of-the-art data science solutions. We've listed three winning solutions using GBDTs below. Please check out the XGBoost website for more winning solutions and use cases. Here are some example successes of GDBTs and Random Forests:

• Halla Yang, 2nd place, Recruit Coupon Purchase Prediction Challenge, Kaggle interview.
• Owen Zhang, 1st place, Avito Context Ad Clicks competition, Kaggle interview.
• Keiichi Kuroyanagi, 2nd place, Airbnb New User Bookings, Kaggle interview.

Getting Started

Prerequisites

• cmake 2.8 or above
  • gcc 4.8 or above for Linux | CUDA 9 or above
  • Visual C++ for Windows | CUDA 10

Quick Install

• For Linux with CUDA 9.0
  • pip install thundergbm
• For Windows (64bit)

  • Download the Python wheel file (for Python3 or above)

    • CUDA 10.0 - Win64

  • Install the Python wheel file

    • pip install thundergbm-0.3.4-py3-none-win_amd64.whl
• Currently only support python3
• After you have installed thundergbm, you can import and use the classifier (similarly for regressor) by:

from thundergbm import TGBMClassifier
clf = TGBMClassifier()
clf.fit(x, y)

Build from source

git clone https://github.com/zeyiwen/thundergbm.git
cd thundergbm
#under the directory of thundergbm
git submodule init cub && git submodule update

Build on Linux (build instructions for Windows)

#under the directory of thundergbm
mkdir build && cd build && cmake .. && make -j

Quick Start

./bin/thundergbm-train ../dataset/machine.conf
./bin/thundergbm-predict ../dataset/machine.conf

You will see RMSE = 0.489562 after successful running.

MacOS is not supported, as Apple has suspended support for some NVIDIA GPUs. We will consider supporting MacOS based on our user community feedbacks. Please stay tuned.

How to cite ThunderGBM

If you use ThunderGBM in your paper, please cite our work (TPDS and JMLR).

@ARTICLE{8727750,
  author={Z. {Wen} and J. {Shi} and B. {He} and J. {Chen} and K. {Ramamohanarao} and Q. {Li}},
  journal={IEEE Transactions on Parallel and Distributed Systems}, 
  title={Exploiting GPUs for Efficient Gradient Boosting Decision Tree Training}, 
  year={2019},
  volume={30},
  number={12},
  pages={2706-2717},
  }

@article{wenthundergbm19,
 author = {Wen, Zeyi and Shi, Jiashuai and He, Bingsheng and Li, Qinbin and Chen, Jian},
 title = {{ThunderGBM}: Fast {GBDTs} and Random Forests on {GPUs}},
 journal = {Journal of Machine Learning Research},
 volume={21},
 year = {2020}
}

Related papers

• Zeyi Wen, Jiashuai Shi, Bingsheng He, Jian Chen, Kotagiri Ramamohanarao and Qinbin Li. Exploiting GPUs for Efficient Gradient Boosting Decision Tree Training. IEEE Transactions on Parallel and Distributed Systems (TPDS), accepted in May 2019. pdf

• Zeyi Wen, Hanfeng Liu, Jiashuai Shi, Qinbin Li, Bingsheng He, Jian Chen. ThunderGBM: Fast GBDTs and Random Forests on GPUs. Featured at JMLR MLOSS (Machine Learning Open Source Software). Year: 2020, Volume: 21, Issue: 108, Pages: 1−5. pdf

• Zeyi Wen, Bingsheng He, Kotagiri Ramamohanarao, Shengliang Lu, and Jiashuai Shi. Efficient Gradient Boosted Decision Tree Training on GPUs. The 32nd IEEE Intern ational Parallel and Distributed Processing Symposium (IPDPS), pages 234-243, 2018. pdf

Key members of ThunderGBM

• Zeyi Wen, NUS (now at The University of Western Australia)
• Hanfeng Liu, GDUFS (a visting student at NUS)
• Jiashuai Shi, SCUT (a visiting student at NUS)
• Qinbin Li, NUS
• Advisor: Bingsheng He, NUS
• Collaborators: Jian Chen (SCUT)

Other information

• This work is supported by a MoE AcRF Tier 2 grant (MOE2017-T2-1-122) and an NUS startup grant in Singapore.

Related libraries

• ThunderSVM, which is another Thunder serier software tool developed by our group.
• XGBoost | LightGBM | CatBoost | cuML