This is the official implementation of CVPR 2022 paper Contrastive Test-Time Adaptation, authored by Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Please also see the project page.

The code is tested with Pytorch 1.7.1, CUDA 10.1. Please also make sure to install the following packages:

pip install hydra-core numpy omegaconf sklearn tqdm wandb

Please download the VisDA-C dataset, and put it under ${DATA_ROOT}. For your convenience we also compiled .txt files based on the the image labels, provided under ./datasets/VISDA-C/. The prepared directory would look like:

${DATA_ROOT}
├── VISDA-C
│   ├── train
│   ├── validation
│   ├── train_list.txt
│   ├── validation_list.txt

${DATA_ROOT} is set to ./datasets/ by default, which can be modified in configs/data/basic.yaml or via hydra command line interface data.data_root=${DATA_ROOT}.

We use hydra as the configuration system. By default, the working directory is ./output, which can be changed directly from configs/root.yaml or via hydra command line interface workdir=${WORK_DIR}.

VISDA-C experiments are done for train to validation adaptation. Before the test-time adaptation, we should have the source model. You may train the source model with script train_VISDA-C_source.sh as shown below. We also provide the pre-trained source models from 3 seeds (2020, 2021, 2022) which can be downloaded from here.

After obtaining the source models, put them under ${SRC_MODEL_DIR} and run train_VISDA-C_target.sh to execute test-time adaptation.

export CUDA_VISIBLE_DEVICES=0,1,2,3

# train source model
bash train_VISDA-C_source.sh

# train TTA
bash train_VISDA-C_target.sh <SRC_MODEL_DIR>

This will reproduce Table. 1 from the main paper:

Please download the DomainNet dataset (cleaned version), and put it under ${DATA_ROOT}. Notice that we follow MME to use a subset that contains 126 classes from 4 domains, so we also compiled .txt files for your convenience based on the the image labels, provided under ./datasets/domainnet-126/. The prepared directory would look like:

${DATA_ROOT}
├── domainnet-126
│   ├── real
│   ├── sketch
│   ├── clipart
│   ├── painting
│   ├── real_list.txt
│   ├── sketch_list.txt
│   ├── clipart_list.txt
│   ├── painting_list.txt

${DATA_ROOT} is set to ./datasets/ by default, which can be modified in configs/data/basic.yaml or via hydra command line interface data.data_root=${DATA_ROOT}.

We use hydra as the configuration system. By default, the working directory is ./output, which can be changed directly from configs/root.yaml or via hydra command line interface workdir=${WORK_DIR}.

DomainNet-126 experiments are done for 7 domain shifts constructed from combinations of Real, Sketch, Clipart, and Painting. Before the test-time adaptation, we should have the source model. You may train the source model with script train_domainnet-126_source.sh as shown below. We also provide the pre-trained source models from 3 seeds (2020, 2021, 2022) which can be downloaded from here.

After obtaining the source models, put them under ${SRC_MODEL_DIR} and run train_domainnet-126_target.sh to execute test-time adaptation.

export CUDA_VISIBLE_DEVICES=0,1,2,3

# train source model
# example: bash train_VISDA-C_source.sh real
bash train_domainnet-126_source.sh <SOURCE_DOMAIN>

# train TTA
# example: bash train_VISDA-C_target.sh real sketch <SRC_MODEL_DIR>
bash train_domainnet-126_target.sh <SOURCE_DOMAIN> <TARGET_DOMAIN> <SRC_MODEL_DIR>

This will reproduce Table. 2 from the main paper:

If you find this work helpful to your own work, please consider citing us:

@inproceedings{chen2022contrastive,
  title={Contrastive Test-time Adaptation},
  author={Chen, Dian and Wang, Dequan and Darrell, Trevor and Ebrahimi, Sayna},
  booktitle={CVPR},
  year={2022}
}