Code to accompany the NeurIPS paper "Structured Denoising Diffusion Models in Discrete State-Spaces".
@inproceedings{austin2021structured,
author = {Jacob Austin and
Daniel D. Johnson and
Jonathan Ho and
Daniel Tarlow and
Rianne van den Berg},
title = {Structured Denoising Diffusion Models in Discrete State-Spaces},
booktitle = {Advances in Neural Information Processing Systems},
year = {2021}
}
The images subdirectory contains the code for D3PMs in image space, which can
be used to to reproduce our experiments from Section 6.
It contains code for evaluating bits-per-dimension metrics, but
the FID and IS metric evaluation code is not currently available.
The text subdirectory contains our code for D3PMs in text space, corresponding
to our experiments in Section 5.
The insertdelete subdirectory contains code for D3PMs augmented with insertion
and deletion operations, as described in the follow-up paper "Beyond In-Place
Corruption: Insertion and Deletion In Denoising Probabilistic Models".
@inproceedings{johnson2021beyond,
author = {Daniel D. Johnson and
Jacob Austin and
Rianne van den Berg and
Daniel Tarlow},
title = {Beyond In-Place Corruption: Insertion and Deletion In Denoising Probabilistic Models},
booktitle = {ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models},
year = {2021}
}