class ACDCStrongTransforms:
pretrain = SequentialWrapperTwice(
comm_transform=pil_augment.Compose([
pil_augment.RandomRotation(30),
pil_augment.RandomVerticalFlip(),
pil_augment.RandomHorizontalFlip(),
pil_augment.RandomCrop(224),
]),
img_transform=pil_augment.Compose([
transforms.ColorJitter(brightness=[0.5, 1.5],
contrast=[0.5, 1.5],
saturation=[0.5, 1.5]),
transforms.ToTensor()
]),
target_transform=pil_augment.Compose([pil_augment.ToLabel()]),
total_freedom=True)
label = SequentialWrapperTwice(
comm_transform=pil_augment.Compose([
pil_augment.RandomCrop(224),
pil_augment.RandomRotation(30),
]),
img_transform=pil_augment.Compose([transforms.ToTensor()]),
target_transform=pil_augment.Compose([pil_augment.ToLabel()]),
)
val = SequentialWrapper(comm_transform=pil_augment.CenterCrop(224))
def __init__(self, comm_transform: Callable[[Image.Image], Image.Image] = None,
img_transform: Callable[[Image.Image], Tensor] = pil_augment.ToTensor(),
target_transform: Callable[[Image.Image], Tensor] = pil_augment.ToLabel(),
total_freedom=True) -> None:
"""
:param total_freedom: if True, the two-time generated images are using different seeds for all aspect,
otherwise, the images are used different random seed only for img_seed
"""
super().__init__(comm_transform, img_transform, target_transform)
self._total_freedom = total_freedom
class ACDCTransforms:
pretrain = SequentialWrapperTwice(
comm_transform=pil_augment.Compose([
pil_augment.RandomCrop(224),
pil_augment.RandomRotation(30), # interpolation to be nearest
]),
img_transform=pil_augment.Compose([
transforms.ColorJitter(brightness=[0.8, 1.3],
contrast=[0.8, 1.3],
saturation=[0.8, 1.3]),
transforms.ToTensor()
]),
target_transform=pil_augment.Compose([pil_augment.ToLabel()]),
total_freedom=True)
label = SequentialWrapperTwice(
comm_transform=pil_augment.Compose([
pil_augment.RandomCrop(224),
pil_augment.RandomRotation(30), # interpolation to be nearest
]),
img_transform=pil_augment.Compose([transforms.ToTensor()]),
target_transform=pil_augment.Compose([pil_augment.ToLabel()]),
)
val = SequentialWrapper(comm_transform=pil_augment.CenterCrop(224))
def __init__(
self,
comm_transform: Callable[[Image.Image], Image.Image] = None,
img_transform: Callable[[Image.Image], Tensor] = pil_augment.ToTensor(),
target_transform: Callable[[Image.Image], Tensor] = pil_augment.ToLabel()
) -> None:
"""
:param comm_transform: common geo-transformation
:param img_transform: transformation only applied for images
:param target_transform: transformation only applied for targets
"""
self._comm_transform = comm_transform
self._img_transform = img_transform
self._target_transform = target_transform