AFHQ - xyfJASON/image-datasets GitHub Wiki

Links

Official website | Papers with Code | Dropbox

Brief introduction

Copied from paperswithcode.

Animal FacesHQ (AFHQ) is a dataset of animal faces consisting of 15,000 high-quality images at 512 × 512 resolution. The dataset includes three domains of cat, dog, and wildlife, each providing 5000 images. By having multiple (three) domains and diverse images of various breeds (≥ eight) per each domain, AFHQ sets a more challenging image-to-image translation problem. All images are vertically and horizontally aligned to have the eyes at the center. The low-quality images were discarded by human effort.

Statistics

Numbers: 15,803

Splits (train / test):

• Total: 14,336 / 1,467
• Cat: 5,065 / 493
• Dog: 4,678 / 491
• Wild: 4,593 / 483

Resolution: 512×512

:warning: The stats above are for AFHQv2, an upgrade from AFHQv1 that used a better resampling method (nearest neighbor -> lanczos), removed about 2% of the images, and saved them in PNG format.

Usage

File structure

Please organize the downloaded dataset in the following file structure:

root
├── afhq_v2.zip    (7.0 GB)
├── train          (extracted from afhq_v2.zip)
│   ├── cat        (contains 5065 images)
│   ├── dog        (contains 4678 images)
│   └── wild       (contains 4593 images)
└── test           (extracted from afhq_v2.zip)
    ├── cat        (contains 493 images)
    ├── dog        (contains 491 images)
    └── wild       (contains 483 images)

API Reference

AFHQ(root: str, split: str = 'train', transforms: Optional[Callable] = None)

• root: Root directory of dataset.
• split: One of {'train', 'test'}.
• transforms: A function/transform that takes in an PIL image and returns a transformed version.

Example

from image_datasets import AFHQ

root = '~/data/AFHQ/'  # path to downloaded dataset
train_set = AFHQ(root=root, split='train')
test_set = AFHQ(root=root, split='test')
print(len(train_set))  # 14336
print(len(test_set))   # 1467
print(train_set[0])    # (<PIL.PngImagePlugin.PngImageFile image mode=RGB size=512x512 at 0x7F5FE8955D10>, 0)