Generalizing across Temporal Domains with Koopman Operators

[Paper]   This is the authors' official PyTorch implementation for Temporal Koopman Networks (TKNets) method in the AAAI 2024 paper Generalizing across Temporal Domains with Koopman Operators.

Prerequisites

• PyTorch >= 1.12.1 (with suitable CUDA and CuDNN version)
• torchvision >= 0.10.0
• Python3
• Numpy
• pandas

Dataset

Dataset can be downloaded using "datasets.py"

Training and test

Experiments on all datasets

python -m scripts.sweep_tdg launch\
       --data_dir=../[dataset folder]\
       --output_dir=./EXPS/AllDataset\
       --command_launcher slurm\
       --algorithms TKNets\
       --datasets TDGRPlate TDGEvolCircle TDGRotatedMNIST TDGPortrait TDGForestCover\
       --n_hparams 10\
       --n_trials 5

Run an experiment on Portrait dataset

python scripts/train_tdg.py --data_dir /datasets --algorithm TKNets --dataset TDGPortrait --test_type forward_test --seed 2 --output_dir /EXPS --save_model_every_checkpoint

Acknowledgement

This code is implemented based on the domainbed code.

Citation

If you use this code for your research, please consider citing:

@article{zeng2024generalizing,
  title={Generalizing across Temporal Domains with Koopman Operators},
  author={Zeng, Qiuhao and Wang, Wei and Zhou, Fan and Xu, Gezheng and Pu, Ruizhi and Shui, Changjian and Gagne, Christian and Yang, Shichun and Wang, Boyu and Ling, Charles X},
  journal={arXiv preprint arXiv:2402.07834},
  year={2024}
}