README.md

Deep Latent State Space Models for Time-Series Generation

Implementation of Deep Latent State Space Models for Time-Series Generation

Linqi Zhou, Michael Poli, Winnie Xu, Stefano Massaroli, Stefano Ermon

Requirements:

Install requirements via

$ pip install -r requirements.txt

Install extensions by

$ cd extensions/cauchy
$ python setup.py install

Data

For generation, we use Monash Forecasting Repository, which can be downloaded here.

We experiment with NN5 Daily, FRED MD, Solar Weekly, and Temperature Rain.

After download, put each .tsf file into your folder of choice.

Training and evaluation:

For training on generation for Monash,

$ python train_monash.py --config CONFIG

Please refer to configs/monash/ folder for configs of each dataset.

For training on interpolation/extrapolation,

$ python train_interp_extrap.py --config CONFIG

Please refer to configs/interpolation/ and configs/extrapolation/ for configs of each setting.

The models will be evaluated periodically during training and produce relevant metrics in the paper.

Reference

@inproceedings{zhou2023deep,
  title={Deep latent state space models for time-series generation},
  author={Zhou, Linqi and Poli, Michael and Xu, Winnie and Massaroli, Stefano and Ermon, Stefano},
  booktitle={International Conference on Machine Learning},
  pages={42625--42643},
  year={2023},
  organization={PMLR}
}

Acknowledgement

For any questions related to codes and experiment setting, please contact Linqi Zhou.