This repo is no longer active.
This project is intended to build BPL into highly reusable Python modules, for the purpose of expiermentation and eventual use in production systems.
The idea is to encourage discussion and experimentation around BPL and BPL-inspired variants, and to explore this class of models in production settings.
Please see the wiki for details.
What PyBPL is: A framework for developing BPL variants from a generalized form of the BPL algortithm. See this wiki entry for more info.
What PyBPL is not: A faithful implementation of the BPL as applied to the handwriting use case. To run the Matlab tests for handwriting, get the original repository and run the code as described there.
numpynose
Do:
pip install -r requirements.txt
The following system packages are are also required (the command below should work on Ubuntu/Linux):
sudo apt-get install g++ python-dev liblapack-dev gfortran
Do:
nosetests
The original BPL Matlab repo is here: https://github.com/brendenlake/BPL
The original BPL paper can be found on Science:
Lake, B. M., Salakhutdinov, R., and Tenenbaum, J. B. (2015). Human-level concept learning through probabilistic program induction. Science, 350(6266), 1332-1338.
-Base BPL class, with tests
-Base parser class (for learning primitives), with tests
-Tests for base class
-Data formatters (text, time series, etc)
-Helper methods and utility classes as needed