jlu-phyComputer / T2B-PE

The repo is the official implementation for the paper: Intriguing Properties of Positional Encoding in Time Series Forecasting.

T2B-PE

🥇 This repository contains the pytorch code for the IEEE TKDE paper: "Intriguing Properties of Positional Encoding in Time Series Forecasting".

Model

Create Environment

Install Pytorch and other necessary dependencies.

pip install -r requirements.txt

Data Availability

The datasets can be obtained from the official “itransformer” repository or directly from Google Drive or Tsinghua Cloud.

Run

Example: (Train, evaluate, and test on ECL dataset, lookback length:96, prediction length:196):

python -u run.py \
  --is_training 1 \
  --root_path ./dataset/electricity/ \
  --data_path electricity.csv \
  --model_id ECL_96_192 \
  --enc_in 321 \
  --dec_in 321 \
  --c_out 321 \
  --des 'Exp' \
  --batch_size 16 \
  --learning_rate 0.00108 \
  --weight_decay 9e-06\
  --use_weight_dec\
  --pred_len 192
  --itr 1

The training command can be modified according to the above statement :)

Cite

If you find our work and codes useful, please consider citing our paper and star our repository, thanks a lot.

@misc{zhang2024intriguing,
      title={Intriguing Properties of Positional Encoding in Time Series Forecasting}, 
      author={Jianqi Zhang and Jingyao Wang and Wenwen Qiang and Fanjiang Xu and Changwen Zheng and Fuchun Sun and Hui Xiong},
      year={2024},
      eprint={2404.10337},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}

(arXiv version, the final version will be updated after the paper is published.)