Online Test-time Adaptation

This is the official implementation of paper In Search of Lost Online Test-time Adaptation: A Survey. This implementation based on ViT backbones.

Note that the final version of the code is still processing. We will release it later. Thank you!

Prerequisites

To use the repository, we provide a conda environment.

conda update conda
conda env create -f environment.yml
conda activate tta 

• Datasets

  • This repository allows studying a wide range of datasets, models, settings, and methods. A quick overview is given below:
  • cifar10_c CIFAR10-C
  • cifar100_c CIFAR100-C
  • imagenet_c ImageNet-C
  • CIFAR-10-Warehouse CIFAR-10-Warehouse

• Settings

  • reset_each_shift Reset the model state after the adaptation to a domain.

• Batch Size

  • 1, 16, 32, 64, 128

• Backbone

  • ViT B-16 224

• Methods

  • Tent
  • CoTTA
  • MEMO
  • SAR
  • Conjugate PL
  • RoTTA
  • TAST
  • ROID

Changing Configurations

All the hyperparameter could be changed in the folder classification/cfgs/

Please consider cite:

@article{DBLP:journals/ijcv/WangLZCWH25,
  author       = {Zixin Wang and
                  Yadan Luo and
                  Liang Zheng and
                  Zhuoxiao Chen and
                  Sen Wang and
                  Zi Huang},
  title        = {In Search of Lost Online Test-Time Adaptation: {A} Survey},
  journal      = {Int. J. Comput. Vis.},
  volume       = {133},
  number       = {3},
  pages        = {1106--1139},
  year         = {2025},
  url          = {https://doi.org/10.1007/s11263-024-02213-5},
  doi          = {10.1007/S11263-024-02213-5},
  biburl       = {https://dblp.org/rec/journals/ijcv/WangLZCWH25.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Acknowledgements

• Test-time Adaptation official
• We would like to thank all the authors mentioned in the paper. Thank you for contributing Test-time Adaptation community.