This is the source code repository for our paper "Learning Task-General Representations with Generative Neuro-Symbolic Modeling" (Feinman & Lake, 2020).

This code repository requires Python 3 with PyTorch. We recommend using PyTorch >= 1.5.0, as this version was used for development and testing.

In order to use the provided sources code modules, make sure to add the GNS-Modeling root folder to your python path (Unix machines):

export PYTHONPATH="/path/to/GNS-Modeling:$PYTHONPATH"

In order to use this code, you will need to pre-install the pyBPL python library, a separate Python library that we also maintain. Make sure to download the repository and add its root folder to your python path, e.g.:

export PYTHONPATH="/path/to/pyBPL:$PYTHONPATH"

At the moment, pyBPL uses the bottom-up parser from the BPL matlab code during inference. Therefore to use our inference algorithms you will need to download the BPL matlab repository and install the MATLAB Engine API for Python to call its functions from python. Make sure you have created an environment variable BPL_PATH that points to the BPL base directory.

Note: we are working on re-implementing the bottom-up parser in pyBPL. Once complete, pyBPL will be a fully self-contained python package.

After completing the setup steps above, navigate to the experiments/ directory to find scripts for running our experiments with the GNS model. Each experiment has its own unique sub-directory and a README with directions for running the experiment scripts. Some experiments require a cluster to complete in reasonable time, and we've included python scripts to execute cluster jobs using the submitit package for Slurm cluster scheduling.

We are still working to clean up some of the experiment code and appologize if some pieces are missing or unclear.

Please cite our paper:

Feinman, R. and Lake, B. M. (2020). Learning task-general representations with Generative Neuro-Symbolic Modeling. arXiv preprint arXiv:2006.14448.