This is the official Pytorch implementation of the paper: CirT: Global Subseaonal-to-Seasonal Forecasting with Geometry-inspired Transformer in ICLR 2025.
Create environment and install dependencies
conda create -n cirt python==3.9
conda activate cirt
pip install -r requirements.txt
The traning data is downloaded from WeatherBench2
python data_download/training_data/step1_pressure_level_download.py
python data_download/training_data/step2_single_level_download.py
python data_download/training_data/step3_compute_climatology.py --dataset_name pressure_level_1.5
python data_download/training_data/step3_compute_climatology.py --dataset_name single_level_1.5
Update `CIRT/configs/CirT.yaml` field: data_dir: <YOUR_DATA_DIR>
python train.py
If you find any of the code useful, feel free to cite these works.
@inproceedings{
liu2025cirt,
title={CirT: Global Subseasonal-to-Seasonal Forecasting with Geometry-inspired Transformer},
author={Yang Liu and Zinan Zheng and Jiashun Cheng and Fugee Tsung and Deli Zhao and Yu Rong and Jia Li},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025}
}
We use the code from the repository ChaosBench