This is the official repository for Class-aware Domain Knowledge Fusion and Fission for Continual Test-Time Adaptation.

conda create -n CTTA python==3.9.7
conda activate CTTA
pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

• ImageNet-C Download
• Other datasets can be downloaded automatically.

Set --data_dir to your corruption dataset path.

For the "Source Domain Statistics", besides training it by yourself, you can also use the one provided by us: ImageNet, Cifar10, Cifar100.

You can set --train_info to use the provided "Source Domain Statistics", or you can train it yourself by setting --src_data_dir.

• ImageNet-to-ImageNet-C: you can directly load it from timm.
• Cifar10-to-Cifar10-C: you can load the source model from here.
• Cifar100-to-Cifar100-C: you can load the source model from here.

For Cifar10-to-Cifar10-C/Cifar100-to-Cifar100-C, load the source model by setting --checkpoint.

bash bash/imagenet.sh
bash bash/cifar10.sh
bash bash/cifar100.sh

The results were obtained with a single NVIDIA 4090 GPU.

If you find this code useful for your research, please cite our paper.

@inproceedings{zhou2025classaware,
    title={Class-aware Domain Knowledge Fusion and Fission for Continual Test-Time Adaptation}, 
    author={Zhou, Jiahuan and Zhu, Chao and Cui, Zhenyu and Liu, Zichen and Zou, Xu and Hua, Gang}, 
    booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}, 
    year={2025} 
}

Our code is based on the PyTorch implementation of the following projects:

• CoTTA
• ViDA
• DPCore
• Robustbench