DeSSL.transforms package

Submodules

autoaugment module

_images/original.png

The example of original data.
_images/autoaugment.png

The example of AutoAugment Augmentation.
class CIFAR10Policy(fillcolor=(128, 128, 128))

Bases: object

Randomly choose one of the best 25 Sub-policies on CIFAR10.

Example

>>> policy = CIFAR10Policy()
>>> transformed = policy(image)

Example as a PyTorch Transform:

>>> transform=transforms.Compose([
>>>     transforms.Resize(256),
>>>     CIFAR10Policy(),
>>>     transforms.ToTensor()])

cutmix module

cutmix_for_integer(input: torch.Tensor, target: torch.Tensor, gamma: float, indices: Optional[torch.Tensor] = None) → Tuple[torch.Tensor, float, torch.Tensor, torch.Tensor]

CutMix Augmentation for Tensor.

Parameters

  • input – input tensor of data.

  • target – integer label of data.

  • gamma – The expected interpolation coefficient.

  • indices – indices[i] denotes the cutmix target of data i. If indices=None, then cutmix generates indices by torch.perm.

Example

>>> from DeSSL.transforms import cutmix_for_integer
>>> input = torch.randn([256, 3, 64, 64])
>>> target = torch.rand([256, 10]).argmax(dim=-1)
>>> gamma = 0.995
>>> mix_input, real_gamma, target, perm_target = cutmix_for_integer(input, target, gamma)
Returns

Mixed tensor, real gamma, targets and perm_targets.

cutmix_for_one_hot(input: torch.Tensor, target: torch.Tensor, gamma: float, indices: Optional[torch.Tensor] = None) → Tuple[torch.Tensor, torch.Tensor]

CutMix Augmentation for Tensor.

Parameters

  • input – input tensor of data.

  • target – one-hot label of data.

  • gamma – The expected interpolation coefficient.

  • indices – indices[i] denotes the cutmix target of data i. If indices=None, then cutmix generates indices by torch.perm.

Example

>>> from DeSSL.transforms import cutmix_for_one_hot
>>> input = torch.randn([256, 3, 64, 64])
>>> target = torch.rand([256, 10])
>>> gamma = 0.995
>>> mix_input, mix_target = cutmix_for_one_hot(input, target, gamma)
Returns

Mixed tensor and mixed targets.

cutout module

_images/cutout.png

The example of CutOut Augmentation.

class ImageCutout(length: int, color: Tuple[int, int, int] = (127, 127, 127))

Bases: object

Cutout Augmentation for Image.

Parameters

  • length – side length of cutout part.

  • color – color of cutout part.

Example

>>> from torchvision import transforms as tf
>>> transforms = tf.Compose([ImageCutout(10), tf.ToTensor()])
Returns

An augmented image.

class TensorCutout(length: int)

Bases: object

Cutout Augmentation for Tensor.

Parameters

length – side length of cutout part.

Example

>>> from torchvision import transforms as tf
>>> transforms = tf.Compose([tf.ToTensor(), TensorCutout(10)])
Returns

An augmented tensor.

mixup module

class IntegerMixLoss

Bases: torch.nn.modules.module.Module

Creates a criterion that measures the cross entropy loss between each element in the input and integer target.

forward(mix_input: torch.Tensor, gamma: float, target: torch.LongTensor, perm_target: torch.LongTensor)
Parameters

  • mix_input – The input tensor.

  • gamma – The interpolation coefficient.

  • target – The target tensor.

  • perm_target – The perm_target tensor.

Returns

The cross entropy loss.

training: bool
class OneHotMixLoss

Bases: torch.nn.modules.module.Module

Creates a criterion that measures the cross entropy loss between each element in the input and one-hot target.

forward(mix_input: torch.Tensor, mix_target: torch.Tensor) → torch.Tensor
Parameters

  • mix_input – The input tensor.

  • mix_target – The target tensor.

Returns

The cross entropy loss.

training: bool
mixup_for_integer(input: torch.Tensor, target: torch.Tensor, gamma: float, indices: Optional[torch.Tensor] = None) → Tuple[torch.Tensor, torch.Tensor, torch.Tensor]

Mixup Augmentation for Tensor.

Parameters

  • input – input tensor of data.

  • target – integer label of data.

  • gamma – The interpolation coefficient.

  • indices – indices[i] denotes the mixup target of data i. If indices=None, then mixup generates indices by torch.perm.

Example

>>> from allinone.transforms import mixup_for_integer
>>> input = torch.randn([256, 3, 64, 64])
>>> target = torch.rand([256, 10]).argmax(dim=-1)
>>> gamma = 0.995
>>> mix_input, target, perm_target = mixup_for_integer(input, target, gamma)
Returns

Mixed tensor, targets and perm_targets.

mixup_for_one_hot(input: torch.Tensor, target: torch.Tensor, gamma: Union[float, torch.Tensor], indices: Optional[torch.Tensor] = None) → Tuple[torch.Tensor, torch.Tensor]

Mixup Augmentation for Tensor.

Parameters

  • input – input tensor of data.

  • target – one-hot label of data.

  • gamma – The interpolation coefficient.

  • indices – indices[i] denotes the mixup target of data i. If indices=None, then mixup generates indices by torch.perm.

Example

>>> from allinone.transforms import mixup_for_one_hot
>>> input = torch.randn([256, 3, 64, 64])
>>> target = torch.rand([256, 10])
>>> gamma = 0.995
>>> mix_input, mix_target = mixup_for_one_hot(input, target, gamma)
Returns

Mixed tensor and mixed targets.

one_hot_mix_loss(mix_input, mix_target, is_logit=True)

randaugment module

class RandAugment(n: int, m: int)

Bases: object

The RandAugment reproduced by pytorch. This code is completely duplicated from Github.

Parameters

  • n – Number of augmentation transformations to apply sequentially.

  • m – Magnitude for all the transformaations.

DeSSL.transforms.transform_many_times module

class ManyTimes(transform: Callable, n: int)

Bases: object

This class transfers an image to a series of augmented images.

Parameters

  • transform – The transform for augmentation and normalization of images.

  • n – The times that the transform performs.

Returns

The tuple of augmented images.

Twice(transform: Callable) → DeSSL.transforms.transform_many_times.ManyTimes

The easy call method of ManyTimes(transform, 2).

Parameters

transform – The transform for augmentation and normalization of images.

Returns

The class of ManyTimes(transform, 2).

Module contents