PatchTST Model Implementation: Reconstructed Benchmarks (ICLR 2023)

This is an un-offical implementation of PatchTST: A Time Series is Worth 64 Words: Long-term Forecasting with Transformers.

The original authors provide a video with a concise overview of the paper for individuals seeking further comprehension: https://www.youtube.com/watch?v=Z3-NrohddJw

Key Designs

🌟 Patching: segmentation of time series into subseries-level patches which are used as input tokens to the Transformer.

🌟 Channel-independence: each channel is processed independently by the Transformer to predict the forecast. The prediction for multivariate time series is a concatenation of the channel predictions, as shown in the figure below.

Acknowledgement

The original authors acknowledgement these repos, and so in turn we appreciate the following github repo very much for the valuable code base and datasets:

https://github.com/cure-lab/LTSF-Linear

https://github.com/zhouhaoyi/Informer2020

https://github.com/thuml/Autoformer

https://github.com/MAZiqing/FEDformer

https://github.com/alipay/Pyraformer

https://github.com/ts-kim/RevIN

https://github.com/timeseriesAI/tsai

Get Started

Download Datasets

To download the datasets, please use the following link to access our shared Google Drive folder. Once downloaded, place the data folder in root of the github directory

https://drive.google.com/drive/folders/14VMQ5msUCNvZkEJqpEfvF2Ul0iBpzWTz?usp=sharing

Install the Package

Before using the code, please install patchtstblind as a package to allow imports using the following:

pip install -e .

Train Model

To run the experiments, navigate to the patchtstblind/jobs/local directory:

cd patchtstblind/jobs/local

and locate the submit.py file. Execute the file using the following command:

python submit.py --exp_name "<exp_name>"

All experiments are contained in patchtstblind/jobs/exp with experimental parameters stored in the args.yaml files. For example, if you wish to run the ETTh1 experiment with sequence length $L = 512$ and prediction length $T = 96$:

python submit.py --exp_name "etth1/etth1_512_96"

Citation

@inproceedings{Yuqietal-2023-PatchTST,
  title     = {A Time Series is Worth 64 Words: Long-term Forecasting with Transformers},
  author    = {Nie, Yuqi and
               H. Nguyen, Nam and
               Sinthong, Phanwadee and 
               Kalagnanam, Jayant},
  booktitle = {International Conference on Learning Representations},
  year      = {2023}
}