Skip to content
G2SAT: Learning to Generate SAT Formulas
Python C++
Branch: master
Clone or download

Latest commit

Fetching latest commit…
Cannot retrieve the latest commit at this time.


Type Name Latest commit message Commit time
Failed to load latest commit information.
dataset conversion Oct 21, 2019
eval graph statistics Oct 22, 2019
.gitignore Init commit Oct 10, 2019 correct typo Jan 10, 2020 fix Oct 10, 2019 fix Oct 10, 2019 fast template implementation Oct 10, 2019 Init commit Oct 10, 2019

G2SAT: Learning to Generate SAT Formulas

This repository is the official PyTorch implementation of "G2SAT: Learning to Generate SAT Formulas".

Jiaxuan You*, Haoze Wu*, Clark Barrett, Raghuram Ramanujan, Jure Leskovec, G2SAT: Learning to Generate SAT Formulas, NeurIPS 2019.


  • Install PyTorch (tested on 1.0.0), please refer to the offical website for further details
conda install pytorch torchvision cudatoolkit=9.0 -c pytorch
  • Install PyTorch Geometric (tested on 1.1.2), please refer to the offical website for further details
pip install --verbose --no-cache-dir torch-scatter
pip install --verbose --no-cache-dir torch-sparse
pip install --verbose --no-cache-dir torch-cluster
pip install --verbose --no-cache-dir torch-spline-conv (optional)
pip install torch-geometric
  • Install networkx (tested on 2.3), make sure you are not using networkx 1.x version!
pip install networkx
  • Install tensorboardx
pip install tensorboardX

Example Run

You can try out the following 4 steps one by one.

  1. Preprocess data
python eval/ --src dataset/train_formulas/ -s dataset/train_set/
python eval/ --src dataset/test_formulas/ -s dataset/test_set/
  1. Train G2SAT
python --epoch_num 201

After this step, trained G2SAT models will be saved in model/ directory.

  1. Use G2SAT to generate Formulas
python --epoch_load 200

After this step, generated graphs will be saved to graphs/ directory. 1 graph is generated out of 1 template.

Graphs will be saved in 2 formats: a single .dat file containing all the generated graphs; a directory where each generated graph is saved as a single .dat file.

(It may take fairly long time: Runing G2SAT is fast, but updating networkx takes the majority of time in current implementation.)

We can then generate CNF formulas from the generated graphs

python --src graphs/GCN_3_32_preTrue_dropFalse_yield1_019501.120000_0.dat --store-dir formulas --action=lcg2sat
  1. Analyze results We make use of this script to compute the scale-free structure (
g++ -o eval/scale_free eval/scale_free.cpp
python eval/ -s eval/scalefree -d formulas/ 

This will print out the mean/std of the graph statistics of formulas in the formulas/ directory. You could also dump the raw statistics by adding the following flag: -o graph_statistics.csv

You are highly encouraged to tune all kinds of hyper-parameters to get better performance. We only did very limited hyper-parameter tuning.

We recommend using tensorboard to monitor the training process. To do this, you may run

tensorboard --logdir runs


If you find this work useful, please cite our paper:

  title={G2SAT: Learning to Generate SAT Formulas},
  author={You, Jiaxuan and Wu, Haoze and Barrett, Clark and Ramanujan, Raghuram and Leskovec, Jure},
You can’t perform that action at this time.