Reimagine TSFM

Code to reproduce results in our paper: Reimagining Time Series Foundation Models: Metadata and State-Space Model Perspectives

Setup

1. Set up the environment in the SpaceTime paper.
2. Install dependency via

pip install -r requirements.txt

Quick start

Use the inference_tutorial.ipynb for a quick start on how to use our models. The default model is TS-SSM.

Training

Here we uploaded a toy dataset for our users to start train our models. We plan to upload the full dataset to Huggingface to enable users to fully reproduce our results.

Upload our dataset to Huggingface

• DOTOs

SSM

accelerate launch --config_file ./acc_config.yaml --main_process_port 29509 main.py --dataset customized_s --config_file ./ssm.yaml

TS-SSM

accelerate launch --config_file ./acc_config.yaml --main_process_port 29509 main.py --dataset customized_s --config_file ./ssm_ts_small_v2.yaml

Transformer

accelerate launch --config_file ./acc_config.yaml --main_process_port 29509 main.py --dataset customized_s --config_file ./attn.yaml

TS-Transformer

accelerate launch --config_file ./acc_config.yaml --main_process_port 29509 main.py --dataset customized_s --config_file ./attn_ts_small_v2.yaml

Acknowledgment

This research was sponsored in part by the AFOSR award #FA95502210193, the DEVCOM ARL award #W911NF1720196, the NSF award #CNS-23091241, the NIH award #1P41EB028242, and the Pennsylvania Infrastructure Technology Alliance.

We'd like to thank the authors of SpaceTime paper, from which this repo is adapted.

Citation

If you use our code or found our work valuable, please cite:

@inproceedings{quanreimagining,
  title={Reimagining Time Series Foundation Models: Metadata and State-Space Model Perspectives},
  author={Quan, Pengrui and Mulayim, Ozan Baris and Han, Liying and Hong, Dezhi and Berges, Mario and Srivastava, Mani},
  booktitle={NeurIPS Workshop on Time Series in the Age of Large Models}
}