Domain Specific Block Selection and Paired-View Pseudo-Labeling for Online Test-Time Adaptation (DPLOT)

Official implementation of the "Domain Specific Block Selection and Paired-View Pseudo-Labeling for Online Test-Time Adaptation (CVPR 2024)".

[Paper] [Arxiv] [Poster]

Prerequisites

To use the repository, we provide a conda environment.

conda create -n tta python=3.8
conda activate tta 
pip install -r requirements.txt

Classification

Features

This repository allows to reproduce the results in the paper. A quick overview is given below:

• Datasets

  • Dataset will be automatically downloaded. If you have problem, you can manually download the corruption dataset in following links and place it in "classification/data/" folder.
  • cifar10_c CIFAR10-C
  • cifar100_c CIFAR100-C

• Models

  • For the CIFAR10-C and CIFAR100-C benchmarks, pre-trained models from RobustBench can be used.

• Settings

  • reset_each_shift Reset the model state after the adaptation to a domain.
  • continual Train the model on a sequence of domains without knowing when a domain shift occurs.
  • gradual Train the model on a sequence of gradually increasing/decreasing domain shifts without knowing when a domain shift occurs.
  • mixed_domains Train the model on one long test sequence where consecutive test samples are likely to originate from different domains.

• Methods

  • The repository currently supports the following methods: BN-0 (source), BN-1, TENT, MEMO, ETA, EATA, CoTTA, AdaContrast, LAME, SAR, RoTTA, GTTA, RMT, and ROID .

Run Experiments

To run the code for CIFAR benchmarks, run the following command first.

cd classification

Then, this code provides config files for all experiments and methods. Simply run the following Python file with the corresponding config file.

python test_time.py --cfg cfgs/[cifar10_c/cifar100_c]/[source/norm_test/tent/memo/eata/cotta/adacontrast/lame/sar/rotta/rmt/roid/dplot].yaml

Also, to reproduce the results of DPLOT. Run the following Python file with DPLOT config.

python test_time.py --cfg cfgs/[cifar10_c/cifar100_c]/dplot.yaml

Changing Configurations

Using default setting, the continual setting benchmarks would be conducted. To run DPLOT on CIFAR10-to-CIFAR10-C in the mixed_domains setting, the arguments below have to be passed.

python test_time.py --cfg cfgs/cifar10_c/dplot.yaml SETTING mixed_domains

Also, to run DPLOT on CIFAR10-to-CIFAR10-C in the gradual setting, the arguments below have to be passed.

python test_time.py --cfg cfgs/cifar10_c/dplot.yaml SETTING gradual

Changing Architectures

Using checkpoints given by RobustBench, we can simply run the adaptation method with various architectures. For example, to run DPLOT on CIFAR10-to-CIFAR10-C using the ResNet18A, the configuration have to be changed.

python test_time.py --cfg cfgs/cifar10_c/dplot.yaml SETTING gradual MODEL.ARCH Kireev2021Effectiveness_RLATAugMix

The name of the MODEL.ARCH is as follows.

For CIFAR10-C,

• WRN28-10: standard
• WRN40-2A: Hendrycks2020AugMix_WRN
• ResNet18A: Kireev2021Effectiveness_RLATAugMix

For CIFAR100-C,

• ResNext-29A: Hendrycks2020AugMix_ResNeXt
• WRN40-2A: Hendrycks2020AugMix_WRN

Acknowledgements

This work was partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2022-0-00951, Development of Uncertainty Aware Agents Learning by Asking Questions) and by LG Electronics and was collaboratively conducted with the Advanced Robotics Laboratory within the CTO Division of the company.

This code is based on the official repository for the following works:

• Robustbench official
• CoTTA official
• TENT official
• AdaContrast official
• EATA official
• LAME official
• MEMO official
• RoTTA official
• SAR official
• RMT official: Base Code for this repository.