def __init__(self,
root,
N,
noise_level=1e-2,
train=True,
low_light=False,
transform=None,
target_transform=None,
download=False):
# transform = BlockGaussian(N)
transform = th_transforms.Compose([
torchvision.transforms.Resize(size=32),
th_transforms.ToTensor(),
ScaleZeroMean(),
AddGaussianNoiseSetN2N(N, (0, 50.))
])
th_trans = th_transforms.Compose(
[torchvision.transforms.Resize(size=32),
th_transforms.ToTensor()])
self.__class__.__name__ = "mnist"
super(DisentMNISTv1, self).__init__(root,
train=train,
transform=transform,
target_transform=target_transform,
download=download)
self.th_trans = th_trans
def _get_msg_noise(self, params, N):
"""
Noise Type: Multi-scale Gaussian (MSG)
- Each N images has it's own noise level
"""
resize = torchvision.transforms.Resize(size=32)
to_tensor = th_transforms.ToTensor()
szm = ScaleZeroMean()
gaussian_n2n = AddGaussianNoiseSetN2N(N, (0, 50.))
comp = [resize, to_tensor, szm, gaussian_n2n]
t = th_transforms.Compose(comp)
return t
def __init__(self,
root: str,
N: int,
noise_level: float,
random_crop: bool,
split: str = 'train',
transform=None,
target_transform=None):
root = Path(root) / Path("imagenet/")
if random_crop:
transform = th_transforms.Compose([
torchvision.transforms.Resize(size=256),
torchvision.transforms.RandomCrop(size=256),
th_transforms.ToTensor(),
ScaleZeroMean(),
AddGaussianNoiseSetN2N(N, (0, 50.))
])
else:
transform = th_transforms.Compose([
torchvision.transforms.Resize(size=256),
th_transforms.ToTensor(),
ScaleZeroMean(),
AddGaussianNoiseSetN2N(N, (0, 50.))
])
th_trans = th_transforms.Compose([
torchvision.transforms.Resize(size=256),
th_transforms.ToTensor(),
ScaleZeroMean(),
])
super(DisentImageNetv1,
self).__init__(root,
split,
None,
transform=transform,
target_transform=target_transform)
self.th_trans = th_trans