AdaNPC: Exploring Non-Parametric Classifier for Test-Time Adaptation

This codebase is the official implementation of AdaNPC: Exploring Non-Parametric Classifier for Test-Time Adaptation (ICML, 2023), and Domain-Specific Risk Minimization for Out-of-Distribution Generalization (SIGKDD, 2023)

This codebase is mainly based on DomainBed, and T3A.

Installation

Quick start

(1) Downlload the datasets

python download.py --data_dir=/my/datasets/path 

Note: set the target datasets in the download.py

(2) Train a model on source domains

python -m domainbed.scripts.train\
       --data_dir /my/datasets/path\
       --output_dir /my/pretrain/path\
       --algorithm ERM\
       --dataset PACS\
       --hparams "{\"backbone\": \"resnet50\"}" 

This scripts will produce new directory /my/pretrain/path, which include the full training log.

Note: change --dataset PACS for training on other datasets (e.g., VLCS, OfficeHome, TerraIncognita).

Note: change --hparams "{\"backbone\": \"resnet50\"}" for using other backbones (e.g., resnet18, ViT-B16, HViT).

(3) Evaluate model with test time adaptation

python -m domainbed.scripts.unsupervised_adaptation\
       --input_dir=/my/pretrain/path\
       --adapt_algorithm=T3A

This scripts will produce a new file in /my/pretrain/path, whose name is results_{adapt_algorithm}.jsonl.

Note: change --adapt_algorithm=T3A for using other test time adaptation methods (AdaNPC, AdaNPCBN, or TentClf).

(4) Evaluate model with fine-tuning classifier

python -m domainbed.scripts.supervised_adaptation\
       --input_dir=/my/pretrain/path\
       --ft_mode=clf

This scripts will produce a new file in /my/pretrain/path, whose name is results_{ft_mode}.jsonl.

Available backbones

• resnet18
• resnet50
• BiT-M-R50x3
• BiT-M-R101x3
• BiT-M-R152x2
• ViT-B16
• ViT-L16
• DeiT
• Hybrid ViT (HViT)
• MLP-Mixer (Mixer-L16)

Reproducing results for AdaNPC

Table 1 and Figure 2 (Tuned ERM and CORAL)

You can use scripts/hparam_search.sh. Specifically, for each dataset and base algorithm, you can just type a following command.

sh scripts/hparam_search.sh resnet50 PACS ERM

Note that, it automatically starts 240 jobs, and take many times to finish.

Table 2 and Figure 1 (ERM with various backbone)

You can use scripts/launch.sh. Specifically, for each backbone, you can just type following commands.

Specifically, for baselines based on ResNet-50 (PLClf, PLFull, SHOT, SHOTIM, T3A)

sh scripts/launch.sh pretrain resnet50 10 3 local ERM
sh scripts/launch.sh sup resnet50 10 3 local ERM
sh scripts/launch.sh unsup resnet50 10 3 local ERM

for baselines based on ResNet-50-BN (TentClf, TentNorm, TentFull)

sh scripts/launch.sh pretrain resnet50-BN 10 3 local ERM
sh scripts/launch.sh sup resnet50-BN 10 3 local ERM
sh scripts/launch.sh unsup resnet50-BN 10 3 local ERM

for baselines based on KNN training algorithm

sh scripts/launch.sh pretrain resnet50 10 3 local KNN
sh scripts/launch.sh sup resnet50 10 3 local AdaNPC
sh scripts/launch.sh unsup resnet50 10 3 local AdaNPC

Reproducing results for DRM

sh scripts/launch.sh pretrain resnet50 10 3 local DRM
sh scripts/launch.sh sup resnet50 10 3 local DRM
sh scripts/launch.sh unsup resnet50 10 3 local DRM

License

This source code is released under the MIT license, included here.

Citation

If you find this repo useful, please consider citing:

@inproceedings{zhang2023adanpc,
  title={Adanpc: Exploring non-parametric classifier for test-time adaptation},
  author={Zhang, Yifan and Wang, Xue and Jin, Kexin and Yuan, Kun and Zhang, Zhang and Wang, Liang and Jin, Rong and Tan, Tieniu},
  booktitle={International Conference on Machine Learning},
  pages={41647--41676},
  year={2023},
  organization={PMLR}
}


@misc{zhang2022domainspecific,
      title={Domain-Specific Risk Minimization for Out-of-Distribution Generalization}, 
      author={Yi-Fan Zhang and Jindong Wang and Jian Liang and Zhang Zhang and Baosheng Yu and Liang Wang and Dacheng Tao and Xing Xie},
      year={2023},
      archivePrefix={KDD '23: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
      primaryClass={cs.LG}
}