slda

This repository contains Cython implementations of Gibbs sampling for latent Dirichlet allocation and various supervised LDAs:

supervised LDA (linear regression)
binary logistic supervised LDA (logistic regression)
binary logistic hierarchical supervised LDA (trees)
generalized relational topic models (graphs)

Installation

The easy way

Use the conda-forge version here.

The hard way...

(Kept for posterity's sake.)

Dependencies

GNU Scientific Library

This module depends on GSL, please install it. For macosx users using homebrew, this is as simple as

$ brew install gsl

pypolyagamma-3 and gcc

This package depends on pypolyagamma-3, which is a bit of a pain because pypolyagamma-3 requies a C/C++ compiler with OpenMP support. Unfortunately for macosx users, Apple's native compiler, clang, does not ship with that support, so you need to install and use one that does. For macosx users using homebrew, this is as simple as:

$ brew install gcc --without-multilib

This will install a version of gcc with OpenMP support. However, Apple makes things worse by aliasing gcc to point to clang! So you need to explicitly tell the shell which gcc compiler to use. As of the writing of this README, brew installs major version 6 of gcc, and as a result will create a binary called gcc-6 in your path. So export the following to your shell

$ export CC=gcc-6 CXX=g++-6

or you can prefix the commands below with CC=gcc-6 CXX=g++-6.

As a result of this export, it may turn out that your shell cannot find the libraries associated with gcc. If this is the case, specify the path to your gcc library in the environment variable DYLD_LIBRARY_PATH. For example, if you used brew to install gcc as above, then this is probably the right thing to do:

$ export DYLD_LIBRARY_PATH=/usr/local/Cellar/gcc/6.1.0/lib/gcc/6/

Instructions

Conda environment

First create the conda environment by running

$ conda env create

This will install a conda environment called slda, defined in environment.yml, that contains all the dependencies. Activate it by running

$ source activate slda

Next we need to compile the C code in this repository. To do this, run

$ python setup.py build_ext --inplace

pip install slda

If you want slda installed in your environment, run:

$ pip install .

Tests

To run the tests, run

$ py.test slda

This may take as long as 15 minutes, so be patient.

License

This code is open source under the MIT license.

Many thanks to Allen Riddell and his LDA library for inspiration (and code :)

Savvysherpa/slda