Yige Yuan, Bingbing Xu, Liang Hou, Fei Sun, Huawei Shen, Xueqi Cheng

The IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR), 2024

This is an official PyTorch implementation of paper TEA: Test-time Energy Adaptation.

CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/cifar10/energy.yaml

The default model using trained WRN-28-10 from RobustBench.

core/config.py defines all default settings, you can specify particular settings in cfgs/xx.yaml

Our code supports running other baselines with a one-line script, the supported baselines include:

• Source: model without any adaptation
• PL: Pseudo-Label-The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks (ICMLW 2013)
• SHOT: Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation (ICML 2020)
• BN: Improving robustness against common corruptions by covariate shift adaptation (NeurIPS 2020)
• TENT: Tent: Fully Test-Time Adaptation by Entropy Minimization (ICLR 2021)
• ETA: Efficient Test-Time Model Adaptation without Forgetting (ICML 2022)
• EATA: Efficient Test-Time Model Adaptation without Forgetting (ICML 2022)
• SAR: Towards Stable Test-time Adaptation in Dynamic Wild World (ICLR 2023)

# Baselines
CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/cifar10/source.yaml
CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/cifar10/norm.yaml
CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/cifar10/tent.yaml
CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/cifar10/eta.yaml
CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/cifar10/eata.yaml
CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/cifar10/sar.yaml
CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/cifar10/pl.yaml
CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/cifar10/shot.yaml

If you find our work useful, please consider citing our paper:

@article{yuan2023tea,
  title={TEA: Test-time Energy Adaptation},
  author={Yuan, Yige and Xu, Bingbing and Hou, Liang and Sun, Fei and Shen, Huawei and Cheng, Xueqi},
  journal={arXiv preprint arXiv:2311.14402},
  year={2023}
}