🍦EATTA🍦

This is the official project repository for Effortless Active Labeling for Long-Term Test-Time Adaptation by Guowei Wang and Changxing Ding* (CVPR 2025).

🍦EATTA🍦 relieves the increasing annotation burden in ATTA as online data scale grows by requiring only one labeled sample per batch, or even per multi-batches.

• 🍦EATTA🍦 identifies the most valuable sample in each batch for labeling from the perspective of single-step optimization.
• 🍦EATTA🍦 introduces a gradient norm-based debiasing method to balance the training objectives of the labeled and unlabeled data.

Environments

Driver Version: 550.67 | CUDA Version: 12.4 | Python Version：3.10.12

To use the repository, we provide a conda environment.

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

Datasets

• ImageNet-C
• ImageNet-R
• ImageNet-K
• ImageNet-A

After downloading these datasets, you may need to change the path to the root directory in the file conf.py. For example,

# change path
_C.DATA_DIR = "/guowei/data" --> _C.DATA_DIR = "/usr/data"

Baselines

• Baseline. We create an ATTA baseline based on TENT (ICLR 2021). It andomly selects a specified number of samples from each online batch for manual/large model annotation, and perform adaptation using Eq.2.
• SimATTA (ICLR 2024).
• CEMA (ICLR 2024).
• HILTTA (TMLR)

Example

cd classification
CUDA_VISIBLE_DEVICES=0 python test_time.py --cfg cfgs/imagenet_c/eatta.yaml

Considering modifying the following parameters in './cfgs/imagenet_c/eatta.yaml' to test ATTA methods with different parameters. For example,

MODEL:
  ADAPTATION: eatta
  EDGE_CLOUD: True # No need to change
  CLOUD_ARCH: 'vit_l_16' # Architecture of the large model (annotator). Only works when 'LARGE_MODELS' is called.
  EDGE_ARCH: 'resnet50' # Architecture of the adapted model.
  EDGE_ARCH_WEIGHTS: 'IMAGENET1K_V1' # No need to change
  HUMAN_OR_LARGE_MODEL: 'HUMAN' # Switch the annotator: 'HUMAN'-->'LARGE_MODELS'.
  ORACLE_NUM: 1  # Change the number of samples for annotation in each batch: 1-->3.
  BUFFER: False # Not use or use the buffer: False-->True.

---

Correspondence

Please contact Guowei Wang by eegw.wang [at] mail.scut.edu.cn

Citation

If you find 🍦EATTA🍦 is helpful in your research, please considering citing our paper:

@inproceedings{wang2025effortless,
  title={Effortless active labeling for long-term test-time adaptation},
  author={Wang, Guowei and Ding, Changxing},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={25633--25642},
  year={2025}
}

Acknowledgements

We appreciate the contribution of a brilliant codebase https://github.com/mariodoebler/test-time-adaptation, which is developed by Robert A. Marsden and Mario Döbler.