PySBRL

Python interface of Scalable Bayesian Rule Lists described in the Paper: Yang, Hongyu, Cynthia Rudin, and Margo Seltzer. "Scalable Bayesian rule lists." arXiv preprint arXiv:1602.08610 (2016).

Original C Implementation: https://github.com/Hongyuy/sbrlmod by Hongyu Yang et al.

Note: This package is not fully tested.

Install

Now pysbrl is available through PyPi (with wheels for macOS and linux), you can install it via pip.

pip install pysbrl

Usage

import pysbrl

rule_ids, outputs, rule_strings = pysbrl.train_sbrl("data/ttt_train.out", "data/ttt_train.label", 20.0, eta=2.0, max_iters=2000, nchain=10, alphas=[1,1])

rule_ids is the list of ids of rules outputs is the outputs matrix (prob distribution as a vector per rule) rule_strings is the whole list of rules in the format of strings like u'{c2=x,c4=o,c5=b}'.

You can retrieve the selected rules by

rules = [rule_strings[i] for i in rule_ids]

Also note that the last rule is always "default".

Development

Prerequisite

A few C libraries are required to compile this package

1. GNU GSL: You can download from here: https://www.gnu.org/software/gsl/

2. GMP: https://gmplib.org/

On macOS, you can easily install them via Homebrew

brew install gsl gmp

To properly build this module, you need also install swig to generate interface code for wrapping.

Install

git clone https://github.com/myaooo/pysbrl
cd pysbrl
pip install -e .

Dev

The major code for this project is train.c, rulelib.c, and pysbrl.c. pysbrl.c defines the api that exposes to python.

The interface between C and python is defined in pysbrl.i.

After modifying the code, run swig -Wall -python pysbrl.i to generate pysbrl_wrap.c and pysbrl.py.

Build

To build all wheels for linux, you can use docker.

First pull the manylinux docker image:

docker pull quay.io/pypa/manylinux1_x86_64:latest

Then run the scripts/run_docker_build.sh to build the wheels.