Code for the TMLR paper A Variational Perspective on Generative Flow Networks by Heiko Zimmermann, Fredrik Lindsten, Jan-Willem van de Meent, and Christian A Naesseth.

Requirements: All required packages are listed in req.pip and can be installed by running pip install -r req.pip.

Running experiments: Clone repository and run git submodule init followed by git submodule update. The submodule is needed for the synthetic density experiments and is a fork of the code of the ICML 2022 paper Generative Flow Networks for Discrete Probabilistic Modeling by Dinghuai Zhang, Nikolay Malkin, Zhen Liu, Alexandra Volokhova, Aaron Courville, and Yoshua Bengio.

• Execute run_densities2d.sh to run the synthetic density experiments with the parameters specified in the script.
• Execute run_ising.sh to run the Ising model experiments with the parameters specified in the script.

Inspecting results: Results (model parameters, evaluation metrics, and sample plots of final model) are saved in ./multiruns/<hostname>/<experiment_name>/<experiment_id_string>/ in ./params, ./eval, and ./plots respectively.