"""
"""
import logging
from os.path import exists, join
import numpy as np
from PIL import Image
from torch.utils.data import Dataset
from torchvision.datasets.utils import download_url
log = logging.getLogger(__name__)
[docs]
class CIFAR100GAN(Dataset):
"""
Images sampled from low likelihood regions of a BigGAN trained on CIFAR 100 from the paper
*On Outlier Exposure with Generative Models*.
Can be used as auxiliary outliers, e.g. for :class:`OutlierExposure <pytorch_ood.loss.OutlierExposureLoss>` or
any of the supervised training objectives in general.
Default sample :math:`\\sigma` is 50.0. Contains 50,000 samples. Label is `-1` by default.
.. image :: https://files.kondas.de/goe-data/cifar100gan.jpg
:width: 600px
:alt: CIFAR 100 GAN Dataset
:align: center
:see Website: `GitHub <https://github.com/kkirchheim/mlsw2022-goe>`__
:see Paper: `NeurIPS MLSW <https://openreview.net/forum?id=SU7OAfhc8OM>`__
"""
filename = {2.0: "samples-2.0.npz", 50.0: "samples-50.0.npz"}
url = {
2.0: "https://files.kondas.de/goe-data/samples-2.0.npz",
50.0: "https://files.kondas.de/goe-data/samples-50.0.npz",
}
md5 = {
2.0: "f130876edbbc13ab2bdc6f7caaa1180d",
50.0: "95f1365e4c6e188595bb8476d43a82d9",
}
def __init__(self, root, transform=None, target_transform=None, download=False, sigma=50.0):
"""
:param root: where to store the dataset
:param transform: transform to apply to the data
:param target_transform: transform to apply to the target
:param download: whether to download the dataset if it is not found in root
:param sigma: sample :math:`\\sigma` used to generate dataset. Can be ``50.0`` or ``2.0``.
"""
self.datafile = join(root, self.filename[sigma])
if not exists(self.datafile):
if download:
download_url(self.url[sigma], root, md5=self.md5[sigma])
else:
raise FileNotFoundError("Missing File. Set download=True to download.")
log.debug(f"Loading data from {self.datafile}")
self.data = np.load(self.datafile)["x"]
log.debug(f"Shape of dataset: {self.data.shape}")
self.transform = transform
self.target_transform = target_transform
def __getitem__(self, index):
index = index % len(self)
img = self.data[index]
label = -1
img = np.moveaxis(img, 0, -1)
img = Image.fromarray(img)
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
label = self.target_transform(label)
return img, label
def __len__(self):
return self.data.shape[0]