/DomainAdaptor

Primary LanguagePython

DomainAdaptor

The implementation of ICCV 2023 paper 《 DomainAdaptor: A Novel Approach to Test-time Adaptation 》

Install packages

conda install pytorch torchvision cudatoolkit
conda install matplotlib tqdm tensorboardX

Dataset structure

PACS
├── kfold
│   ├── art_painting
│   ├── cartoon
│   ├── photo
│   └── sketch
VLCS
├── CALTECH
│   ├── crossval
│   ├── full
│   ├── test
│   └── train
├── LABELME
│   ├── crossval
│   ├── full
| ...
OfficeHome
├── Art
│   ├── Alarm_Clock
│   ├── Backpack
│   ├── Batteries
│   ├── Bed
│   ├── Bike
│   ├── Bottle
| ...

The data root can be modified in main.py or pase the args --data-root your_data_root.

Run the code

The code of DomainAdaptor is in models/DomainAdaptor.py.

The pretrained deepall models are available at Google Drive. Or you can train the deepall models by yourself with the following code:

bash script/deepall.sh

With the pretrained models, you can run the following code to evaluate with DomainAdaptor:

bash script/TTA.sh

Citation

@inproceedings{zhang2023domainadaptor,
  title={DomainAdaptor: A Novel Approach to Test-time Adaptation},
  author={Zhang, Jian and Qi, Lei and Shi, Yinghuan and Gao, Yang},
  bootitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2023}
}