Datasets Preparation

We conduct experiments on ImageNet, CIFAR100, CelebA, and SUN397. Please follow the instructions to set up the dataset folders. By default, our ${DATASET}$ is set to be datasets.

${DATASET}/
|--- imagenet/
    |--- ILSCRC2012_devkit_t12
    |--- train/
    |--- val/
|--- cifar-100-python3/
    |--- meta/
    |--- train/
    |--- test/
|--- sun397/
    |--- Partitions/
    |--- SUN397/
        |--- a/
        |--- b/
        |--- ...
|--- celeba/
    |--- zips/
        |--- CelebA-20231025T061928Z-001.zip
        |--- CelebA-20231025T061928Z-002.zip
        |--- ...
    |--- img_align_celeba/
    |--- list_attr_celeba.txt

ImageNet

Download the dataset from ILSVRC2012. Please use the scripts/tools/prepare_imagenet_train.sh and scripts/tools/prepare_imagenet_val.sh modified from Preparation of ImageNet (ILSVRC2012) to prepare the imagenet folder.

bash scripts/tools/prepare_imagenet_train.sh ${DATASET} ILSVRC2012_img_train.tar
bash scripts/tools/prepare_imagenet_val.sh ${DATASET} ILSVRC2012_img_val.tar

We provide the superclass setting we used in data/imagenet/superclass-10.json. The split can be generated by running the following code.

python3 -m data.imagenet.generate_superclass \
    --image-dir ${DATASET}/imagenet \
    --num-superclasses 10

CIFAR100

python3 tools/download_datasets cifar100 ${DATASET}

We provide the superclass setting we used in data/cifar100/superclass-20.json. The split is acquired from the official CIFAR100 page.

SUN397

python3 tools/download_datasets sun397 ${DATASET}

We provide the superclass setting we used in data/sun397/superclass-15.json. The split can be generated by running the following code.

python3 -m data.sun397.generate_superclass

CelebA

Please download the dataset from Google Drive or other means. Run the following command to unsip the zip files to the destination.

7z x ${DATASET}/celeba/zips/CelebA-20231025T061928Z-001.zip -o${DATASET}