(Real and generated data from the turbulent flows dataset)
This is the official repository for the AAAI 2022 paper SPATE-GAN: Improved Generative Modeling of Dynamic Spatio-Temporal Patterns with an Autoregressive Embedding Loss (Konstantin Klemmer*, Tianlin Xu*, Beatrice Acciaio, Daniel B. Neill).
* These authors contributed equally.
The source code for SPATE-GAN (using PyTorch) can be found in the src folder. It builds on the code base for COT-GAN (NeurIPS 2020), accessible here: [Tensorflow,PyTorch]
We also provide an interactive example notebook to test SPATE-GAN via Google Colab
(The different approaches for obtaining the spatio-temporal expectations needed to compute SPATE)
Contained within the src folder, the spatial_utils.py file contains all needed functions to compute the SPATE embedding in its different configurations.
Beyond our new SPATE metric, spatial_utils.py also includes the (to our knowledge) first PyTorch implementation of the original local Moran's I metric, along with the capacity to compute it for batches of spatial patterns / images.
(Differences between Moran's I and SPATE in its different configurations)
If you want to cite our work, you can use the following reference:
@article{klemmer2022spategan,
title={SPATE-GAN: Improved Generative Modeling of Dynamic Spatio-Temporal Patterns with an Autoregressive Embedding Loss},
volume={36},
url={https://ojs.aaai.org/index.php/AAAI/article/view/20375},
DOI={10.1609/aaai.v36i4.20375},
number={4},
journal={Proceedings of the AAAI Conference on Artificial Intelligence},
author={Klemmer, Konstantin and Xu, Tianlin and Acciaio, Beatrice and Neill, Daniel B.},
year={2022},
month={Jun.},
pages={4523-4531}
}