This repository contains a reference implementation of RED-SDS, a non-linear state space model proposed in the NeurIPS 2021 paper Deep Explicit Duration Switching Models for Time Series.

• Run pip install -r requirements.txt.

• Download the trained models from this link.
• Run python reevaluate.py --ckpt <model-path>.pt.

• Generate/download datasets.
  • To generate the bouncing ball and 3 mode system datasets, use the notebooks in ./data/. Alternatively, you can download the datasets from this link.
  • To download and preprocess the dancing bees dataset, run ./data/bee.sh.
• Run python run_segmentation.py --config configs/<config>.yaml --device cuda:0 to train the RED-SDS model.
• Run tensorboard --logdir /path/to/results/dir to visualize results.

• Run python run_gts_univariate.py --config configs/<config>.yaml --device cuda:0 to train the RED-SDS model. The dataset will be downloaded automatically.
• Run tensorboard --logdir /path/to/results/dir to visualize results.

For any questions regarding the code or the paper, please email Fatir, Konstantinos, or Richard.

If you find this repository or the ideas presented in our paper useful for your research, please consider citing our paper.

@inproceedings{ansari2021deep,
  author    = {Abdul Fatir Ansari and Konstantinos Benidis and Richard Kurle and Ali Caner Turkmen and Harold Soh and Alex Smola and Bernie Wang and Tim Januschowski},
  title     = {Deep Explicit Duration Switching Models for Time Series},
  year      = {2021},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
}

This repo contains parts of code based on the following repos: