Sherlog

Deep, generative logic programming.

Introduction

Sherlog is a logic programming language with value introductions that supports differentiable functions and probability distributions (both discrete and continuous).

For further examples, see the ./examples directory.

Architecture

SherLog is constructed as the union of three separate sub-components:

1. The externally-guarded generative logic programming core, Watson (./lib/watson),
2. A translation layer converting problem files to Watson programs and the resulting abducted obligations to generative models (./lib/sherlog),
3. And a host of inference techniques built around the produced generative models (./sherlog).

Sub-components 1. and 2. are implemented in OCaml for speed. Sub-component 3. is implemented in Python so as to stand on the shoulders of PyTorch. Communication is handled through JSON serialization sent via a local network socket.

Installation and Building

This section assumes you have the latest versions of OPAM (installation instructions here) and Python.

Option 1

Just run make.

Option 2

First, we can install the OCaml core with opam install .. This will install the executables sherlog-interpreter and sherlog-server in the default OPAM installation path (by default, at ~/.opam/default/bin); either run opam init to add the installation path to your path in your shell configuration file, or run eval $(opam env) to add the path temporarily.

Next, we can install the SherLog Python interface by running python3 -m pip install .. We recommend doing so in a virtual environment.

Usage

Open a Python interpreter and import sherlog to get started. Additionally, the OPAM build process installs sherlog-interpreter and sherlog-server: run them to see their command-line arguments.