/EvoS

Official implementation of our NeurIPS 2023 paper (EvoS).

Primary LanguagePython

Evolving Standardization for Continual Domain Generalization over Temporal Drift [NeurIPS 2023]

openreview    Poster   

Overview

We formulate a promising but challenging problem of continual domain generalization over temporal drift (CDGTD) and propose an Evolving Standardization (EvoS) approach for CDGTD. EvoS characterizes the evolving pattern and further achieves generalization by conducting the feature standardization.

image

Prerequisites Installation

  • The code is implemented with Python 3.7.16, CUDA 12.2. To try out this project, it is recommended to set up a virtual environment first.

    # Step-by-step installation
conda create --name evos python=3.7.16
conda activate evos

# this installs the right pip and dependencies for the fresh python
conda install -y ipython pip

# install torch, torchvision and torchaudio
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html

# this installs required packages
pip install -r requirements.txt

Datasets Preparation

The data folder should be structured as follows:

```
├── datasets/
│   ├── yearbook/     
|   |   ├── yearbook.pkl
│   ├── rmnist/
|   |   ├── MNIST/
|   |   ├── rmnist.pkl
│   ├── huffpost/	
|   |   ├── huffpost.pkl
│   ├── fMoW/	
|   |   ├── fmow_v1.1/
|   |   |   |── images/
|   |   |—— fmow.pkl
│   ├── arxiv/	
|   |   ├── arxiv.pkl
```

Code Running

  • for Eval-Fix manner:

    # running for yearbook dataset:
sh script_yearbook_eval_fix.sh

# running for rmnist dataset:
sh script_rmnist_eval_fix.sh

# running for fmow dataset:
sh script_fmow_eval_fix.sh

# running for huffpost dataset:
sh script_huffpost_eval_fix.sh

# running for arxiv dataset:
sh script_arxiv_eval_fix.sh

  • for Eval-Stream manner:

    # running for yearbook dataset:
sh script_yearbook_eval_stream.sh

# running for huffpost dataset:
sh script_huffpost_eval_stream.sh

Acknowledgments

This project is mainly based on the open-source project: Wild-Time. We thank its authors for making the source code publicly available.

Citation

If you find this work helpful to your research, please consider citing the paper:

@inproceedings{xie2023evos,
  title={Evolving Standardization for Continual Domain Generalization over Temporal Drift},
  author={Mixue Xie, Shuang Li, Longhui Yuan, Chi Harold Liu, Zehui Dai},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2023}
}