DAG-GFlowNet

Paper - Installation - Example - Citation

This repository contains the official implementation in JAX of DAG-GFlowNet (Deleu et al., 2022), a Bayesian structure learning algorithm based on Generative Flow Networks (GFlowNets; Bengio et al., 2021). This contains the environment to sample sequentially a graph one edge at a time, written as a Gym environment.

Tristan Deleu, António Góis, Chris Emezue, Mansi Rankawat, Simon Lacoste-Julien, Stefan Bauer, Yoshua Bengio, Bayesian Structure Learning with Generative Flow Networks, 2022

Installation

To avoid any conflict with your existing Python setup, we suggest to work in a virtual environment:

python -m venv venv
source venv/bin/activate

Follow these instructions to install the version of JAX corresponding to your versions of CUDA and CuDNN.

git clone https://github.com/tristandeleu/jax-dag-gflownet.git
cd jax-dag-gflownet
pip install -r requirements.txt

Example

You can train DAG-GFlowNet on a randomly generated dataset of 100 observations from an Erdos-Renyi graph over 5 nodes using the following command:

python train.py --batch_size 256 erdos_renyi_lingauss --num_variables 5 --num_edges 5 --num_samples 100

Citation

If you want to cite DAG-GFlowNet, use the following Bibtex entry:

@article{deleu2022daggflownet,
    title={{Bayesian Structure Learning with Generative Flow Networks}},
    author={Deleu, Tristan and G{\'o}is, Ant{\'o}nio and Emezue, Chris and Rankawat, Mansi and Lacoste-Julien, Simon and Bauer, Stefan and Bengio, Yoshua},
    journal={arXiv preprint},
    year={2022}
}