Robust Test-Time Adaptation for Zero-Shot Prompt Tuning (ADAPROMPT)

This repository provides the official PyTorch implementation of our AAAI 2024 paper:

Robust Test-Time Adaptation for Zero-Shot Prompt Tuning Authors: Ding-Chu Zhang*, Zhi Zhou*, Yu-Feng Li

For more details, please check out our paper.

Environment

Install pip environment

pip install -r requirements.txt

Install conda environment

conda install --yes --file requirements.txt

Datasets

Download the datasets CIFAR10-C, CIFAR100-C,ImageNet-R and TinyImageNet-C.

Download cross-validation datasets by TPT.

Run

1. Place the dataset in the ./MyDATA folder

2. Run Adaprompt

   python ./test.py ./MyDATA --test_sets CIFAR10_C -a ViT-B/16 -b 64 --gpu 0 --tpt --ctx_init a_photo_of_a --result-dir ./results/ours --method-config ./configs/methods/ours.yaml

3. Run TPT

   python ./test.py ./MyDATA --test_sets CIFAR10_C -a ViT-B/16 -b 64 --gpu 0 --tpt --ctx_init a_photo_of_a --result-dir ./results/tpt --method-config ./configs/methods/tpt.yaml

4. Run Source

   python ./test.py ./MyDATA --test_sets CIFAR10_C -a ViT-B/16 -b 64 --gpu 0 --tpt --ctx_init a_photo_of_a --result-dir ./results/source --method-config ./configs/methods/source.yaml

5. The results will be printed and stored in ./results/.

Acknowledgment

We thank the authors for the following repositories for code reference: TPT.