This is the official code of the paper "A Unified Hyperparameter Optimization Pipeline for Transformer-Based Time Series Forecasting Models"

How to Run

0. Prepare

Datasets

Google Drive

You can check the details of dataset from Time-Series-Library

Key

Add your wandb key to the line wandb.login(key='your key') at the file ~/exp/exp_long_term_forecasting_ray_tune.py

1. Set the Environment

# clone code
git clone https://github.com/jingjing-unilu/HPO_transformer_time_series.git
# create conda enviroment
conda create -n time_hyperS_0909 python=3.8
conda activate time_hyperS_0909
cd /your/path/to/HPO_transformer_time_series

# install necessary libraries
pip install -r requirements.txt
pip install scikit-learn==1.2
pip install pandas
pip install matplotlib
pip install jupyter
pip install eco2ai
pip install "ray[tune]"
pip install optuna
pip install torchvision #for testing
pip install wandb
pip install mamba_ssm

# exit the env after finishing
conda deactivate

2. Quick Simple Test

conda activate time_hyperS_0909
bash ./scripts/long_term_forecast_hyperS/ETT_script/PatchTST_ETTh1.sh > ./log_20240909/ETT_script_PatchTST_ETTh1
conda deactivate

3. Run All Experiments

nohup ./scripts/long_term_forecast_hyperS/0_all_experiments_1.sh > log0 &

4. Some Results

4.0 Best Model & Loss

4.0.0 ETTh1 dataset

<iframe src="https://api.wandb.ai/links/jingjing_unilu-university-of-luxembourg/1vyla5qc" style="border:none;height:1024px;width:100%"></iframe>

4.0.1 Weather dataset

<iframe src="https://api.wandb.ai/links/jingjing_unilu-university-of-luxembourg/qmgok20k" style="border:none;height:1024px;width:100%"></iframe>

4.0.1 ECL dataset

<iframe src="https://api.wandb.ai/links/jingjing_unilu-university-of-luxembourg/uv9kf6yp" style="border:none;height:1024px;width:100%"></iframe>

4.1 Parallel Coordinates Chart

4.1.0 Weather dataset

<iframe src="https://api.wandb.ai/links/jingjing_unilu-university-of-luxembourg/wo74wre6" style="border:none;height:1024px;width:100%"></iframe>

4.1.1 ETTh1 dataset

<iframe src="https://api.wandb.ai/links/jingjing_unilu-university-of-luxembourg/8q7ede82" style="border:none;height:1024px;width:100%"></iframe>

Figures in Paper

Fig.1 - HPO pipeline for Transformer-based forecasting.

Fig.2 - Common Parameters.

Citation

If you find this repo useful, please cite our paper.

@misc{xu2025unifiedhyperparameteroptimizationpipeline,
      title={A Unified Hyperparameter Optimization Pipeline for Transformer-Based Time Series Forecasting Models}, 
      author={Jingjing Xu, Caesar Wu, Yuan-Fang Li, Grégoire Danoy, Pascal Bouvry},
      year={2025},
      eprint={2501.01394},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2501.01394}, 
}

Acknowledgement

This project is supported by the Luxembourg National Research Fund (Fonds National de la Recherche - FNR), Grant ID: 15748747 and Grant ID: C21/IS/16221483/CBD.

The code is constructed based on the following repos: https://github.com/thuml/Time-Series-Library/tree/main

Useful Refrences:

https://pytorch.org/tutorials/beginner/hyperparameter_tuning_tutorial.html
https://docs.ray.io/en/latest/tune/index.html
https://docs.ray.io/en/latest/tune/examples/tune-pytorch-cifar.html#tune-pytorch-cifar-ref