TOP-N: Equivariant set and graph generation without exchangeability

The repository contains code for the paper, accepted at ICLR 2022.

To run the SetMNISt or CLEVR experiments, download the corresponding repositories (https://github.com/Cyanogenoid/dspn and https://github.com/LukeBolly/tf-tspn/blob/master/models/set_prior.py) and replace the corresponding file (dspn.py or tspn.py) by the one present in the dspn_tspn folder.

To run set generation,

• generate the dataset: python3 utils/generate_synthetic_dataset.py
• run main.py.

To run graph generation, run molecule_generation/graph_main.py.

To cite the paper, you can use the following reference:

@inproceedings{
vignac2022topn,
title={Top-N: Equivariant Set and Graph Generation without Exchangeability},
author={Clement Vignac and Pascal Frossard},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=-Gk_IPJWvk}
}