A Code Release for VIPAug

Official PyTorch implementation of "Domain Generalization with Vital Phase Augmentation"

arXiv:2312.16451, Accepted by AAAI-24

Setup

#Clone the repo.
git clone https://github.com/excitedkid/vipaug.git

#Prepare fractal iamges
Make a new directory "fractals"  
Unzip the zip file of fractal images in ./fractals  

#Build environment by docker 
docker pull excitedkid/vipaug:0

#Structure of dataset directory
CIFAR-10
{dataset path}/cifar10/cifar-10-batches-py ...
{dataset path}/CIFAR-10-C ...

CIFAR-100
{dataset path}/cifar100/cifar-100-python ...
{dataset path}/CIFAR-100-C ...

ImageNet
{dataset path}/train ...

You can download fractal images here.

Fractal images are from DeviantArt.

Running

CIFAR

Train

python3 main.py --dataset cifar10 --aug vipaug --vital 0.001 --nonvital 0.014 --data {dataset path} 

Evaluation

python3 main.py --dataset cifar10 --aug vipaug --vital 0.001 --nonvital 0.014 --data {dataset path} --data-c {corrupted dataset path} --eval eval

ImageNet

Train

python3 imagenet.py --gpu 0 --data {dataset path}

Pretrained Models

You can download pretrained models.

CIFAR-10

CIFAR-100

ImageNet

*VIPAug code is based on APR GitHub.

Citation

@misc{vipaug2024,
      title={Domain Generalization with Vital Phase Augmentation}, 
      author={Ingyun Lee and Wooju Lee and Hyun Myung},
      journal={AAAI},
      year={2024}
}