Neural Guided Constraint Logic Programming for Program Synthesis

This repository contains code that goes with the paper: Neural Guided Constraint Logic Programming for Program Synthesis (TODO: add link)

This repository contains an implementation of miniKanren where the constraint trees are represented transparently. We add scaffolding code to show how to drive miniKanren using an external agent in Python. We provide Recurrent Neural Network (RNN) and Graph Neural Network (GNN) agents as examples. The implementations of the RNN and GNN are consistent with the models described in the paper. (TODO: add link)

Dependencies

Chez Scheme: https://github.com/cisco/ChezScheme
Python3+
PyTorch 0.3+: https://pytorch.org/

Transprent miniKanren

The following files contain an implementation of miniKanren where the constraint trees are represented transparently, and a python interface for interacting with a miniKanren process.

File	Description
mk.scm	transparent implemenation of minikanren
evalo.scm	definition of evalo, a relational interpreter
query.scm	build queries annotated with ground truth
query-outputs.scm	compute ground truth outputs for queries
interact.scm	interaction process for python to talk to
interact.py	python interface for scheme interaction
lisp.py	helper for parsing lisp in python

Neural Network Model

The following files contain neural network agents that can drive miniKanren's search.

File	Description
helper.py	helper for working with constraints
rnn_grammar.py	parsing constraints for RNN model
rnn.py	forward pass for RNN model
gnn_grammar.py	parsing constraints for GNN model
gnn.py	forward pass for GNN model

Data Files

The following files contain the test problems used in the paper.

File	Description
data/test_problems.txt	held out tree manipulation problems
data/repeat_test.txt	generalization: repeat(N)
data/drop_last_test.txt	generalization: dropLast(N)
data/bring_to_front_test.txt	generalization: bringToFront(N)

References

[0] (TODO: link to paper)

[1] ICLR 2018 Workshop Paper: with slightly less detail (4 pages)

[2] Lisa Zhang's Master's Thesis: with slightly more detail (23 pages)

[3] More information about miniKanren