PAID: Pairwise Angular-Invariant Decomposition for Continual Test-Time Adaptation

We have released the code for our NeurIPS 2025 paper "PAID: Pairwise Angular-Invariant Decomposition for Continual Test-Time Adaptation".

Prerequisite

[CIFAR10C, CIFAR100C, ImageNetC Datasets Download link]

https://pan.baidu.com/s/19MJRlaSWDQMYpBF1ppPzhA (pwd: 44t2)

[[CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C, ImageNet-to-ImageNetC Pre-train and Warm-up Checkpoints Download link]

https://pan.baidu.com/s/1PYmEkQcbmm4Zdyi02CiOyA (pwd: bxiw)

You need to download the dataset and weight archives and extract them into the assets folder under the root directory of PAID. The correct folder structure under assets should be as follows:

├── PAID
│   ├── asserts
│   │   ├── CIFAR-10-C
│   │   ├── CIFAR-100-C
│   │   ├── ImageNet-C
│   │   ├── vit_base_patch16_224.augreg_in21k_ft_in1k
│   │   ├── vit_base_patch16_384.augreg_in21k_ft_in1k
│   │   ├── cifar10_warmup.pth
│   │   ├── cifar100_warmup.pth
│   │   ├── imagenet_warmup.pth
│   │   ├── pretrain_cifar10.pth
│   │   ├── pretrain_cifar100.pth
│   ├── output
...

Classification Experiments

CIFAR10-to-CIFAR10C-standard task

# Warmup on CIFAR10
cd PAID/cifar
pip install -r requirements.txt -i https://pypi.mirrors.ustc.edu.cn/simple/
python cifar10_pretrain.py --cfg ./cfgs/cifar10/cifar10_pretrain.yaml

# CTTA on CIFAR10C
cd PAID/cifar
pip install -r requirements.txt -i https://pypi.mirrors.ustc.edu.cn/simple/
python cifar10c_vit_ours.py --cfg ./cfgs/cifar10/dora_m_a_hra_qkvom_align.yaml --data_dir ./assets --checkpoint ./assets/cifar10_warmup.pth

CIFAR100-to-CIFAR100C-standard task

# Warmup on CIFAR100
cd PAID/cifar
pip install -r requirements.txt -i https://pypi.mirrors.ustc.edu.cn/simple/
python cifar100_pretrain.py --cfg ./cfgs/cifar100/cifar100_pretrain.yaml

# CTTA on CIFAR100C
cd PAID/cifar
pip install -r requirements.txt -i https://pypi.mirrors.ustc.edu.cn/simple/
python cifar100c_vit_ours.py --cfg ./cfgs/cifar100/dora_m_a_hra_qkvom_align.yaml --data_dir ./assets  --checkpoint ./assets/cifar100_warmup.pth

ImageNet-to-ImageNetC-standard task

# Warmup on ImageNet
cd PAID/imagenet
pip install -r requirements.txt -i https://pypi.mirrors.ustc.edu.cn/simple/
python imagenet_pretrain.py --cfg ./cfgs/vit/imagenet_pretrain.yaml

# CTTA on ImageNetC
cd PAID/imagenet
pip install -r requirements.txt -i https://pypi.mirrors.ustc.edu.cn/simple/
python imagenetc_dora_magnitude_angle_hra.py --cfg ./cfgs/vit/dora_m_a_hra_qkvom_align.yaml --data_dir ./assets --checkpoint ./assets/imagenet_warmup.pth

Important Note

During the Warmup stage and CTTA stage where source domain statistics are extracted, we use the CIFAR10, CIFAR100, and ImageNet datasets. Due to the large size of these datasets, we do not provide them here. Please download them manually to your local machine. After downloading, update the following file paths in the code according to your local dataset locations:

• cifar10_pretrain.py: lines 51 and 54
• cifar10c_vit_ours.py: line 91
• cifar100_pretrain.py: lines 51 and 54
• cifar100c_vit_ours.py: line 89
• imagenet_pretrain.py: line 76
• imagenetc_dora_magnitude_angle_hra.py: line 88