weifantt / depts

DEPTS

Source code for the paper, "DEPTS: Deep Expansion Learning for Periodic Time Series Forecasting", in ICLR22 Spotlight.

Overview

DEPTS is a customized deep neural network architecture for periodic time series forecasting, which aims to solve the following two challenges:

• To capture diversified periodic compositions
• To model complicated periodic dependencies

Dataset

You can download the five benchmarks from Google Drive. All the datasets are well pre-processed. More details of datasets can be found in the paper. After downloading the zip file, please unzip it to the root dir of DEPTS for experiments.

Usage

Setup

Please use Python 3(.6) as well as the following packages:

torch >= 1.6.0
dataclasses
dtaidistance
pandas
numpy
tqdm

Reproduce

To reproduce the results, you can see more details in command.sh and directly run:

sh command.sh

Note that all the results reported in the paper are ensembled results of 30 models in order to get a robust evaluation and compare with N-BEATS. You can also try to run the single model for evaluation if you find it challenging to run all the models.

Evaluation

To get the evaluation results, run

python evaluation.py

Citation

If you find our work interesting, you can cite the paper as

@inproceedings{
fan2022depts,
title={{DEPTS}: Deep Expansion Learning for Periodic Time Series Forecasting},
author={Wei Fan and Shun Zheng and Xiaohan Yi and Wei Cao and Yanjie Fu and Jiang Bian and Tie-Yan Liu},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=AJAR-JgNw__}
}

depts's Issues

A question about hyper-parameter "fftwarmlen"

Hi, very lucky to read this wonderful paper on periodical modeling and saw your publicly available code.

However, I am confused about some parameters in your code but it is not mentioned in your paper. That is "fftwarmlen" used in the "function warm_PM_parameters_perK" (utils.py - line 100). It is set in different constants varying with datasets, and then determines the length of the FFT handled. Could you tell me how to set its value?

query

in utils.py, line: 122
why use add to check if they are equivalent?