Transforms2D¶

Set of operators to perform data augmentation on 2D image tensors.

class CenterCrop(size: Union[int, Tuple[int, int]], align_corners: bool = True, resample: Union[str, int, <unknown>.Resample] = 'BILINEAR', return_transform: bool = False, p: float = 1.0, keepdim: bool = False, cropping_mode: str = 'slice')[source]¶

Crops a given image tensor at the center.

Parameters

p (float) – probability of applying the transformation for the whole batch. Default value is 1.
size (Tuple[int, int] or int) – Desired output size (out_h, out_w) of the crop. If integer, out_h = out_w = size. If Tuple[int, int], out_h = size[0], out_w = size[1].
return_transform (bool) – if True return the matrix describing the transformation applied to each. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.
cropping_mode (str) – The used algorithm to crop. slice will use advanced slicing to extract the tensor based on the sampled indices. resample will use warp_affine using the affine transformation to extract and resize at once. Use slice for efficiency, or resample for proper differentiability. Default: slice.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, out_h, out_w)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> inputs = torch.randn(1, 1, 4, 4)
>>> inputs
tensor([[[[-1.1258, -1.1524, -0.2506, -0.4339],
          [ 0.8487,  0.6920, -0.3160, -2.1152],
          [ 0.3223, -1.2633,  0.3500,  0.3081],
          [ 0.1198,  1.2377,  1.1168, -0.2473]]]])
>>> aug = CenterCrop(2, p=1.)
>>> aug(inputs)
tensor([[[[ 0.6920, -0.3160],
          [-1.2633,  0.3500]]]])

class ColorJitter(brightness: Union[torch.Tensor, float, Tuple[float, float], List[float]] = 0.0, contrast: Union[torch.Tensor, float, Tuple[float, float], List[float]] = 0.0, saturation: Union[torch.Tensor, float, Tuple[float, float], List[float]] = 0.0, hue: Union[torch.Tensor, float, Tuple[float, float], List[float]] = 0.0, return_transform: bool = False, same_on_batch: bool = False, p: float = 1.0, keepdim: bool = False)[source]¶

Applies a random transformation to the brightness, contrast, saturation and hue of a tensor image.

Parameters

p (float) – probability of applying the transformation. Default value is 1.
brightness (float or tuple) – Default value is 0.
contrast (float or tuple) – Default value is 0.
saturation (float or tuple) – Default value is 0.
hue (float or tuple) – Default value is 0.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> inputs = torch.ones(1, 3, 3, 3)
>>> aug = ColorJitter(0.1, 0.1, 0.1, 0.1, p=1.)
>>> aug(inputs)
tensor([[[[0.9993, 0.9993, 0.9993],
          [0.9993, 0.9993, 0.9993],
          [0.9993, 0.9993, 0.9993]],
<BLANKLINE>
         [[0.9993, 0.9993, 0.9993],
          [0.9993, 0.9993, 0.9993],
          [0.9993, 0.9993, 0.9993]],
<BLANKLINE>
         [[0.9993, 0.9993, 0.9993],
          [0.9993, 0.9993, 0.9993],
          [0.9993, 0.9993, 0.9993]]]])

class GaussianBlur(kernel_size: Tuple[int, int], sigma: Tuple[float, float], border_type: str = 'reflect', return_transform: bool = False, same_on_batch: bool = False, p: float = 0.5)[source]¶

Apply gaussian blur given tensor image or a batch of tensor images randomly.

Parameters

kernel_size (Tuple[int, int]) – the size of the kernel.
sigma (Tuple[float, float]) – the standard deviation of the kernel.
border_type (str) – the padding mode to be applied before convolving. The expected modes are: 'constant', 'reflect', 'replicate' or 'circular'. Default: 'reflect'.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
p (float) – probability of applying the transformation. Default value is 0.5.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> input = torch.rand(1, 1, 5, 5)
>>> blur = GaussianBlur((3, 3), (0.1, 2.0), p=1.)
>>> blur(input)
tensor([[[[0.6699, 0.4645, 0.3193, 0.1741, 0.1955],
          [0.5422, 0.6657, 0.6261, 0.6527, 0.5195],
          [0.3826, 0.2638, 0.1902, 0.1620, 0.2141],
          [0.6329, 0.6732, 0.5634, 0.4037, 0.2049],
          [0.8307, 0.6753, 0.7147, 0.5768, 0.7097]]]])

class RandomAffine(degrees: Union[torch.Tensor, float, Tuple[float, float]], translate: Union[torch.Tensor, Tuple[float, float], None] = None, scale: Union[torch.Tensor, Tuple[float, float], Tuple[float, float, float, float], None] = None, shear: Union[torch.Tensor, float, Tuple[float, float], None] = None, resample: Union[str, int, <unknown>.Resample] = 'BILINEAR', return_transform: bool = False, same_on_batch: bool = False, align_corners: bool = False, padding_mode: Union[str, int, <unknown>.SamplePadding] = 'ZEROS', p: float = 0.5, keepdim: bool = False)[source]¶

Applies a random 2D affine transformation to a tensor image.

The transformation is computed so that the image center is kept invariant.

Parameters

p (float) – probability of applying the transformation. Default value is 0.5.
degrees (float or tuple) – Range of degrees to select from. If degrees is a number instead of sequence like (min, max), the range of degrees will be (-degrees, +degrees). Set to 0 to deactivate rotations.
translate (tuple, optional) – tuple of maximum absolute fraction for horizontal and vertical translations. For example translate=(a, b), then horizontal shift is randomly sampled in the range -img_width * a < dx < img_width * a and vertical shift is randomly sampled in the range -img_height * b < dy < img_height * b. Will not translate by default.
scale (tuple, optional) – scaling factor interval. If (a, b) represents isotropic scaling, the scale is randomly sampled from the range a <= scale <= b. If (a, b, c, d), the scale is randomly sampled from the range a <= scale_x <= b, c <= scale_y <= d. Will keep original scale by default.
shear (sequence or float, optional) – Range of degrees to select from. If float, a shear parallel to the x axis in the range (-shear, +shear) will be apllied. If (a, b), a shear parallel to the x axis in the range (-shear, +shear) will be apllied. If (a, b, c, d), then x-axis shear in (shear[0], shear[1]) and y-axis shear in (shear[2], shear[3]) will be applied. Will not apply shear by default.
resample (int, str or kornia.Resample) – resample mode from “nearest” (0) or “bilinear” (1). Default: Resample.BILINEAR.
padding_mode (int, str or kornia.SamplePadding) – padding mode from “zeros” (0), “border” (1) or “refection” (2). Default: SamplePadding.ZEROS.
return_transform (bool) – if True return the matrix describing the transformation applied to each. Default: False.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
align_corners (bool) – interpolation flag. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> input = torch.rand(1, 1, 3, 3)
>>> aug = RandomAffine((-15., 20.), return_transform=True, p=1.)
>>> aug(input)
(tensor([[[[0.3961, 0.7310, 0.1574],
          [0.1781, 0.3074, 0.5648],
          [0.4804, 0.8379, 0.4234]]]]), tensor([[[ 0.9923, -0.1241,  0.1319],
         [ 0.1241,  0.9923, -0.1164],
         [ 0.0000,  0.0000,  1.0000]]]))

class RandomCrop(size: Tuple[int, int], padding: Union[int, Tuple[int, int], Tuple[int, int, int, int], None] = None, pad_if_needed: Optional[bool] = False, fill: int = 0, padding_mode: str = 'constant', resample: Union[str, int, <unknown>.Resample] = 'BILINEAR', return_transform: bool = False, same_on_batch: bool = False, align_corners: bool = True, p: float = 1.0, keepdim: bool = False, cropping_mode: str = 'slice')[source]¶

Crops random patches of a tensor image on a given size.

Parameters

p (float) – probability of applying the transformation for the whole batch. Default value is 1.0.
size (Tuple[int, int]) – Desired output size (out_h, out_w) of the crop. Must be Tuple[int, int], then out_h = size[0], out_w = size[1].
padding (int or sequence, optional) – Optional padding on each border of the image. Default is None, i.e no padding. If a sequence of length 4 is provided, it is used to pad left, top, right, bottom borders respectively. If a sequence of length 2 is provided, it is used to pad left/right, top/bottom borders, respectively.
pad_if_needed (boolean) – It will pad the image if smaller than the desired size to avoid raising an exception. Since cropping is done after padding, the padding seems to be done at a random offset.
fill – Pixel fill value for constant fill. Default is 0. If a tuple of length 3, it is used to fill R, G, B channels respectively. This value is only used when the padding_mode is constant
padding_mode – Type of padding. Should be: constant, edge, reflect or symmetric. Default is constant.
resample (int, str or kornia.Resample) – Default: Resample.BILINEAR
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated
same_on_batch (bool) – apply the same transformation across the batch. Default: False
align_corners (bool) – interpolation flag. Default: True.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.
cropping_mode (str) – The used algorithm to crop. slice will use advanced slicing to extract the tensor based on the sampled indices. resample will use warp_affine using the affine transformation to extract and resize at once. Use slice for efficiency, or resample for proper differentiability. Default: slice.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, out_h, out_w)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> _ = torch.manual_seed(0)
>>> inputs = torch.arange(1*1*3*3.).view(1, 1, 3, 3)
>>> aug = RandomCrop((2, 2), p=1.)
>>> aug(inputs)
tensor([[[[3., 4.],
          [6., 7.]]]])

class RandomErasing(scale: Union[torch.Tensor, Tuple[float, float]] = (0.02, 0.33), ratio: Union[torch.Tensor, Tuple[float, float]] = (0.3, 3.3), value: float = 0.0, return_transform: bool = False, same_on_batch: bool = False, p: float = 0.5, keepdim: bool = False)[source]¶

Erases a random rectangle of a tensor image according to a probability p value.

The operator removes image parts and fills them with zero values at a selected rectangle for each of the images in the batch.

The rectangle will have an area equal to the original image area multiplied by a value uniformly sampled between the range [scale[0], scale[1]) and an aspect ratio sampled between [ratio[0], ratio[1])

Parameters

p (float) – probability that the random erasing operation will be performed. Default value is 0.5.
scale (Tuple[float, float]) – range of proportion of erased area against input image.
ratio (Tuple[float, float]) – range of aspect ratio of erased area.
same_on_batch (bool) – apply the same transformation across the batch. Default: False
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> inputs = torch.ones(1, 1, 3, 3)
>>> rec_er = RandomErasing((.4, .8), (.3, 1/.3), p=0.5)
>>> rec_er(inputs)
tensor([[[[1., 0., 0.],
          [1., 0., 0.],
          [1., 0., 0.]]]])

class RandomGrayscale(return_transform: bool = False, same_on_batch: bool = False, p: float = 0.1, keepdim: bool = False)[source]¶

Applies random transformation to Grayscale according to a probability p value.

Parameters

p (float) – probability of the image to be transformed to grayscale. Default value is 0.1.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> inputs = torch.randn((1, 3, 3, 3))
>>> rec_er = RandomGrayscale(p=1.0)
>>> rec_er(inputs)
tensor([[[[-1.1344, -0.1330,  0.1517],
          [-0.0791,  0.6711, -0.1413],
          [-0.1717, -0.9023,  0.0819]],
<BLANKLINE>
         [[-1.1344, -0.1330,  0.1517],
          [-0.0791,  0.6711, -0.1413],
          [-0.1717, -0.9023,  0.0819]],
<BLANKLINE>
         [[-1.1344, -0.1330,  0.1517],
          [-0.0791,  0.6711, -0.1413],
          [-0.1717, -0.9023,  0.0819]]]])

class RandomHorizontalFlip(return_transform: bool = None, same_on_batch: bool = False, p: float = 0.5, p_batch: float = 1.0, keepdim: bool = False)[source]¶

Applies a random horizontal flip to a tensor image or a batch of tensor images with a given probability.

Input should be a tensor of shape (C, H, W) or a batch of tensors \((B, C, H, W)\). If Input is a tuple it is assumed that the first element contains the aforementioned tensors and the second, the corresponding transformation matrix that has been applied to them. In this case the module will Horizontally flip the tensors and concatenate the corresponding transformation matrix to the previous one. This is especially useful when using this functionality as part of an nn.Sequential module.

Parameters

p (float) – probability of the image being flipped. Default value is 0.5
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> input = torch.tensor([[[[0., 0., 0.],
...                         [0., 0., 0.],
...                         [0., 1., 1.]]]])
>>> seq = nn.Sequential(RandomHorizontalFlip(p=1.0, return_transform=True),
...                     RandomHorizontalFlip(p=1.0, return_transform=True))
>>> seq(input)
(tensor([[[[0., 0., 0.],
          [0., 0., 0.],
          [0., 1., 1.]]]]), tensor([[[1., 0., 0.],
         [0., 1., 0.],
         [0., 0., 1.]]]))

class RandomVerticalFlip(return_transform: bool = None, same_on_batch: bool = False, p: float = 0.5, p_batch: float = 1.0, keepdim: bool = False)[source]¶

Applies a random vertical flip to a tensor image or a batch of tensor images with a given probability.

Parameters

p (float) – probability of the image being flipped. Default value is 0.5
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> input = torch.tensor([[[[0., 0., 0.],
...                         [0., 0., 0.],
...                         [0., 1., 1.]]]])
>>> seq = RandomVerticalFlip(p=1.0, return_transform=True)
>>> seq(input)
(tensor([[[[0., 1., 1.],
          [0., 0., 0.],
          [0., 0., 0.]]]]), tensor([[[ 1.,  0.,  0.],
         [ 0., -1.,  2.],
         [ 0.,  0.,  1.]]]))

class RandomMotionBlur(kernel_size: Union[int, Tuple[int, int]], angle: Union[torch.Tensor, float, Tuple[float, float]], direction: Union[torch.Tensor, float, Tuple[float, float]], border_type: Union[int, str, <unknown>.BorderType] = 'CONSTANT', resample: Union[str, int, <unknown>.Resample] = 'NEAREST', return_transform: bool = False, same_on_batch: bool = False, p: float = 0.5, keepdim: bool = False)[source]¶

Perform motion blur on 2D images (4D tensor).

Parameters

p (float) – probability of applying the transformation. Default value is 0.5.
kernel_size (int or Tuple[int, int]) – motion kernel size (odd and positive). If int, the kernel will have a fixed size. If Tuple[int, int], it will randomly generate the value from the range batch-wisely.
angle (float or Tuple[float, float]) – angle of the motion blur in degrees (anti-clockwise rotation). If float, it will generate the value from (-angle, angle).
direction (float or Tuple[float, float]) – forward/backward direction of the motion blur. Lower values towards -1.0 will point the motion blur towards the back (with angle provided via angle), while higher values towards 1.0 will point the motion blur forward. A value of 0.0 leads to a uniformly (but still angled) motion blur. If float, it will generate the value from (-direction, direction). If Tuple[int, int], it will randomly generate the value from the range.
border_type (int, str or kornia.BorderType) – the padding mode to be applied before convolving. CONSTANT = 0, REFLECT = 1, REPLICATE = 2, CIRCULAR = 3. Default: BorderType.CONSTANT.
resample (int, str or kornia.Resample) – Default: Resample.NEAREST.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Please set resample to 'bilinear' if more meaningful gradients wanted.

Examples

>>> rng = torch.manual_seed(0)
>>> input = torch.ones(1, 1, 5, 5)
>>> motion_blur = RandomMotionBlur(3, 35., 0.5, p=1.)
>>> motion_blur(input)
tensor([[[[0.5773, 1.0000, 1.0000, 1.0000, 0.7561],
          [0.5773, 1.0000, 1.0000, 1.0000, 0.7561],
          [0.5773, 1.0000, 1.0000, 1.0000, 0.7561],
          [0.5773, 1.0000, 1.0000, 1.0000, 0.7561],
          [0.5773, 1.0000, 1.0000, 1.0000, 0.7561]]]])

class RandomPerspective(distortion_scale: Union[torch.Tensor, float] = 0.5, resample: Union[str, int, <unknown>.Resample] = 'BILINEAR', return_transform: bool = False, same_on_batch: bool = False, align_corners: bool = False, p: float = 0.5, keepdim: bool = False)[source]¶

Applies a random perspective transformation to an image tensor with a given probability.

Parameters

p (float) – probability of the image being perspectively transformed. Default value is 0.5.
distortion_scale (float) – it controls the degree of distortion and ranges from 0 to 1. Default value is 0.5.
resample (int, str or kornia.Resample) – Default: Resample.BILINEAR.
return_transform (bool) – if True return the matrix describing the transformation applied to each. Default: False.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
align_corners (bool) – interpolation flag. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> inputs= torch.tensor([[[[1., 0., 0.],
...                         [0., 1., 0.],
...                         [0., 0., 1.]]]])
>>> aug = RandomPerspective(0.5, p=0.5)
>>> aug(inputs)
tensor([[[[0.0000, 0.2289, 0.0000],
          [0.0000, 0.4800, 0.0000],
          [0.0000, 0.0000, 0.0000]]]])

class RandomResizedCrop(size: Tuple[int, int], scale: Union[torch.Tensor, Tuple[float, float]] = (0.08, 1.0), ratio: Union[torch.Tensor, Tuple[float, float]] = (0.75, 1.3333333333333333), resample: Union[str, int, <unknown>.Resample] = 'BILINEAR', return_transform: bool = False, same_on_batch: bool = False, align_corners: bool = True, p: float = 1.0, keepdim: bool = False, cropping_mode: str = 'slice')[source]¶

Crops random patches in an image tensor and resizes to a given size.

Parameters

size (Tuple[int, int]) – Desired output size (out_h, out_w) of each edge. Must be Tuple[int, int], then out_h = size[0], out_w = size[1].
scale – range of size of the origin size cropped.
ratio – range of aspect ratio of the origin aspect ratio cropped.
resample (int, str or kornia.Resample) – Default: Resample.BILINEAR.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
align_corners (bool) – interpolation flag. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.
cropping_mode (str) – The used algorithm to crop. slice will use advanced slicing to extract the tensor based on the sampled indices. resample will use warp_affine using the affine transformation to extract and resize at once. Use slice for efficiency, or resample for proper differentiability. Default: slice.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, out_h, out_w)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Example

>>> rng = torch.manual_seed(0)
>>> inputs = torch.tensor([[[0., 1., 2.],
...                         [3., 4., 5.],
...                         [6., 7., 8.]]])
>>> aug = RandomResizedCrop(size=(3, 3), scale=(3., 3.), ratio=(2., 2.), p=1.)
>>> aug(inputs)
tensor([[[[1.0000, 1.5000, 2.0000],
          [4.0000, 4.5000, 5.0000],
          [7.0000, 7.5000, 8.0000]]]])

class RandomRotation(degrees: Union[torch.Tensor, float, Tuple[float, float], List[float]], resample: Union[str, int, <unknown>.Resample] = 'BILINEAR', return_transform: bool = False, same_on_batch: bool = False, align_corners: bool = True, p: float = 0.5, keepdim: bool = False)[source]¶

Applies a random rotation to a tensor image or a batch of tensor images given an amount of degrees.

Parameters

p (float) – probability of applying the transformation. Default value is 0.5.
degrees (sequence or float or tensor) – range of degrees to select from. If degrees is a number the range of degrees to select from will be (-degrees, +degrees).
resample (int, str or kornia.Resample) – Default: Resample.BILINEAR.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
align_corners (bool) – interpolation flag. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> input = torch.tensor([[1., 0., 0., 2.],
...                       [0., 0., 0., 0.],
...                       [0., 1., 2., 0.],
...                       [0., 0., 1., 2.]])
>>> seq = RandomRotation(degrees=45.0, return_transform=True, p=1.)
>>> seq(input)
(tensor([[[[0.9824, 0.0088, 0.0000, 1.9649],
          [0.0000, 0.0029, 0.0000, 0.0176],
          [0.0029, 1.0000, 1.9883, 0.0000],
          [0.0000, 0.0088, 1.0117, 1.9649]]]]), tensor([[[ 1.0000, -0.0059,  0.0088],
         [ 0.0059,  1.0000, -0.0088],
         [ 0.0000,  0.0000,  1.0000]]]))

class RandomSolarize(thresholds: Union[torch.Tensor, float, Tuple[float, float], List[float]] = 0.1, additions: Union[torch.Tensor, float, Tuple[float, float], List[float]] = 0.1, same_on_batch: bool = False, return_transform: bool = False, p: float = 0.5, keepdim: bool = False)[source]¶

Solarize given tensor image or a batch of tensor images randomly.

Parameters

p (float) – probability of applying the transformation. Default value is 0.5.
thresholds (float or tuple) – Default value is 0.1. If float x, threshold will be generated from (0.5 - x, 0.5 + x). If tuple (x, y), threshold will be generated from (x, y).
additions (float or tuple) – Default value is 0.1. If float x, addition will be generated from (-x, x). If tuple (x, y), addition will be generated from (x, y).
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> input = torch.rand(1, 1, 5, 5)
>>> solarize = RandomSolarize(0.1, 0.1, p=1.)
>>> solarize(input)
tensor([[[[0.4132, 0.1412, 0.1790, 0.2226, 0.3980],
          [0.2754, 0.4194, 0.0130, 0.4538, 0.2771],
          [0.4394, 0.4923, 0.1129, 0.2594, 0.3844],
          [0.3909, 0.2118, 0.1094, 0.2516, 0.3728],
          [0.2278, 0.0000, 0.4876, 0.0353, 0.5100]]]])

class RandomPosterize(bits: Union[int, Tuple[int, int], torch.Tensor] = 3, same_on_batch: bool = False, return_transform: bool = False, p: float = 0.5, keepdim: bool = False)[source]¶

Posterize given tensor image or a batch of tensor images randomly.

Parameters

p (float) – probability of applying the transformation. Default value is 0.5.
bits (int or tuple) – Integer that ranged from (0, 8], in which 0 gives black image and 8 gives the original. If int x, bits will be generated from (x, 8). If tuple (x, y), bits will be generated from (x, y). Default value is 3.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> input = torch.rand(1, 1, 5, 5)
>>> posterize = RandomPosterize(3, p=1.)
>>> posterize(input)
tensor([[[[0.4706, 0.7529, 0.0627, 0.1255, 0.2824],
          [0.6275, 0.4706, 0.8784, 0.4392, 0.6275],
          [0.3451, 0.3765, 0.0000, 0.1569, 0.2824],
          [0.5020, 0.6902, 0.7843, 0.1569, 0.2510],
          [0.6588, 0.9098, 0.3765, 0.8471, 0.4078]]]])

class RandomSharpness(sharpness: Union[torch.Tensor, float, Tuple[float, float]] = 0.5, same_on_batch: bool = False, return_transform: bool = False, p: float = 0.5, keepdim: bool = False)[source]¶

Sharpen given tensor image or a batch of tensor images randomly.

Parameters

p (float) – probability of applying the transformation. Default value is 0.5.
sharpness (float or tuple) – factor of sharpness strength. Must be above 0. Default value is 0.5.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> input = torch.rand(1, 1, 5, 5)
>>> sharpness = RandomSharpness(1., p=1.)
>>> sharpness(input)
tensor([[[[0.4963, 0.7682, 0.0885, 0.1320, 0.3074],
          [0.6341, 0.4810, 0.7367, 0.4177, 0.6323],
          [0.3489, 0.4428, 0.1562, 0.2443, 0.2939],
          [0.5185, 0.6462, 0.7050, 0.2288, 0.2823],
          [0.6816, 0.9152, 0.3971, 0.8742, 0.4194]]]])

class RandomEqualize(same_on_batch: bool = False, return_transform: bool = False, p: float = 0.5, keepdim: bool = False)[source]¶

Equalize given tensor image or a batch of tensor images randomly.

Parameters

p (float) – Probability to equalize an image. Default value is 0.5.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False.

Shape:

Input: \((C, H, W)\) or \((B, C, H, W)\), Optional: \((B, 3, 3)\)
Output: \((B, C, H, W)\)

Note

Input tensor must be float and normalized into [0, 1] for the best differentiability support. Additionally, this function accepts another transformation tensor (\((B, 3, 3)\)), then the applied transformation will be merged int to the input transformation tensor and returned.

Examples

>>> rng = torch.manual_seed(0)
>>> input = torch.rand(1, 1, 5, 5)
>>> equalize = RandomEqualize(p=1.)
>>> equalize(input)
tensor([[[[0.4963, 0.7682, 0.0885, 0.1320, 0.3074],
          [0.6341, 0.4901, 0.8964, 0.4556, 0.6323],
          [0.3489, 0.4017, 0.0223, 0.1689, 0.2939],
          [0.5185, 0.6977, 0.8000, 0.1610, 0.2823],
          [0.6816, 0.9152, 0.3971, 0.8742, 0.4194]]]])

class RandomMixUp(lambda_val: Union[torch.Tensor, Tuple[float, float], None] = None, same_on_batch: bool = False, p: float = 1.0, keepdim: bool = False)[source]¶

Apply MixUp augmentation to a batch of tensor images.

Implemention for mixup: BEYOND EMPIRICAL RISK MINIMIZATION [ZnYNDLP18].

The function returns (inputs, labels), in which the inputs is the tensor that contains the mixup images while the labels is a \((B, 3)\) tensor that contains (label_batch, label_permuted_batch, lambda) for each image.

The implementation is on top of the following repository: https://github.com/hongyi-zhang/mixup/blob/master/cifar/utils.py.

The loss and accuracy are computed as:

def loss_mixup(y, logits):
    criterion = F.cross_entropy
    loss_a = criterion(logits, y[:, 0].long(), reduction='none')
    loss_b = criterion(logits, y[:, 1].long(), reduction='none')
    return ((1 - y[:, 2]) * loss_a + y[:, 2] * loss_b).mean()

def acc_mixup(y, logits):
    pred = torch.argmax(logits, dim=1).to(y.device)
    return (1 - y[:, 2]) * pred.eq(y[:, 0]).float() + y[:, 2] * pred.eq(y[:, 1]).float()

Parameters

p (float) – probability for applying an augmentation to a batch. This param controls the augmentation probabilities batch-wisely.
lambda_val (float or torch.Tensor, optional) – min-max value of mixup strength. Default is 0-1.
same_on_batch (bool) – apply the same transformation across the batch. This flag will not maintain permutation order. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False

Inputs:

Input image tensors, shape of \((B, C, H, W)\).
Label: raw labels, shape of \((B)\).

Returns

Adjusted image, shape of \((B, C, H, W)\).
Raw labels, permuted labels and lambdas for each mix, shape of \((B, 3)\).

Return type

Tuple[torch.Tensor, torch.Tensor]

Note

This implementation would randomly mixup images in a batch. Ideally, the larger batch size would be preferred.

Examples

>>> rng = torch.manual_seed(1)
>>> input = torch.rand(2, 1, 3, 3)
>>> label = torch.tensor([0, 1])
>>> mixup = RandomMixUp()
>>> mixup(input, label)
(tensor([[[[0.7576, 0.2793, 0.4031],
          [0.7347, 0.0293, 0.7999],
          [0.3971, 0.7544, 0.5695]]],
<BLANKLINE>
<BLANKLINE>
        [[[0.4388, 0.6387, 0.5247],
          [0.6826, 0.3051, 0.4635],
          [0.4550, 0.5725, 0.4980]]]]), tensor([[0.0000, 0.0000, 0.1980],
        [1.0000, 1.0000, 0.4162]]))

class RandomCutMix(height: int, width: int, num_mix: int = 1, cut_size: Union[torch.Tensor, Tuple[float, float], None] = None, beta: Union[float, torch.Tensor, None] = None, same_on_batch: bool = False, p: float = 1.0, keepdim: bool = False)[source]¶

Apply CutMix augmentation to a batch of tensor images.

Implemention for CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features [YHO+19].

The function returns (inputs, labels), in which the inputs is the tensor that contains the mixup images while the labels is a \((\text{num_mixes}, B, 3)\) tensor that contains (label_permuted_batch, lambda) for each cutmix.

The implementation referred to the following repository: https://github.com/clovaai/CutMix-PyTorch.

The onehot label may be computed as:

def onehot(size, target):
    vec = torch.zeros(size, dtype=torch.float32)
    vec[target] = 1.
    return vec

def cutmix_label(labels, out_labels, size):
    lb_onehot = onehot(size, labels)
    for out_label in out_labels:
        label_permuted_batch, lam = out_label[:, 0], out_label[:, 1]
        label_permuted_onehot = onehot(size, label_permuted_batch)
        lb_onehot = lb_onehot * lam + label_permuted_onehot * (1. - lam)
    return lb_onehot

Parameters

height (int) – the width of the input image.
width (int) – the width of the input image.
p (float) – probability for applying an augmentation to a batch. This param controls the augmentation probabilities batch-wisely.
num_mix (int) – cut mix times. Default is 1.
beta (float or torch.Tensor, optional) – hyperparameter for generating cut size from beta distribution. Beta cannot be set to 0 after torch 1.8.0. If None, it will be set to 1.
cut_size ((float, float) or torch.Tensor, optional) – controlling the minimum and maximum cut ratio from [0, 1]. If None, it will be set to [0, 1], which means no restriction.
same_on_batch (bool) – apply the same transformation across the batch. This flag will not maintain permutation order. Default: False.
keepdim (bool) – whether to keep the output shape the same as input (True) or broadcast it to the batch form (False). Default: False

Inputs:

Input image tensors, shape of \((B, C, H, W)\).
Raw labels, shape of \((B)\).

Returns

Adjusted image, shape of \((B, C, H, W)\).
Raw labels, permuted labels and lambdas for each mix, shape of \((B, num_mix, 3)\).

Return type

Tuple[torch.Tensor, torch.Tensor]

Note

This implementation would randomly cutmix images in a batch. Ideally, the larger batch size would be preferred.

Examples

>>> rng = torch.manual_seed(3)
>>> input = torch.rand(2, 1, 3, 3)
>>> input[0] = torch.ones((1, 3, 3))
>>> label = torch.tensor([0, 1])
>>> cutmix = RandomCutMix(3, 3)
>>> cutmix(input, label)
(tensor([[[[0.8879, 0.4510, 1.0000],
          [0.1498, 0.4015, 1.0000],
          [1.0000, 1.0000, 1.0000]]],
<BLANKLINE>
<BLANKLINE>
        [[[1.0000, 1.0000, 0.7995],
          [1.0000, 1.0000, 0.0542],
          [0.4594, 0.1756, 0.9492]]]]), tensor([[[0.0000, 1.0000, 0.4444],
         [1.0000, 0.0000, 0.4444]]]))

class RandomInvert(max_val: torch.Tensor = tensor(1.), return_transform: bool = False, same_on_batch: bool = False, p: float = 0.5)[source]¶

Invert the tensor images values randomly.

Parameters

max_val (torch.Tensor) – The expected maximum value in the input tensor. The shape has to according to the input tensor shape, or at least has to work with broadcasting. Default: 1.0.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
p (float) – probability of applying the transformation. Default value is 0.5.

Examples

>>> rng = torch.manual_seed(0)
>>> img = torch.rand(1, 1, 5, 5)
>>> inv = RandomInvert()
>>> inv(img)
tensor([[[[0.4963, 0.7682, 0.0885, 0.1320, 0.3074],
          [0.6341, 0.4901, 0.8964, 0.4556, 0.6323],
          [0.3489, 0.4017, 0.0223, 0.1689, 0.2939],
          [0.5185, 0.6977, 0.8000, 0.1610, 0.2823],
          [0.6816, 0.9152, 0.3971, 0.8742, 0.4194]]]])

class RandomChannelShuffle(return_transform: bool = False, same_on_batch: bool = False, p: float = 0.5)[source]¶

Shuffles the channels of a batch of multi-dimensional images.

Parameters

return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
p (float) – probability of applying the transformation. Default value is 0.5.

Examples

>>> rng = torch.manual_seed(0)
>>> img = torch.arange(1*2*2*2.).view(1,2,2,2)
>>> RandomChannelShuffle()(img)
tensor([[[[4., 5.],
          [6., 7.]],
<BLANKLINE>
         [[0., 1.],
          [2., 3.]]]])

class RandomGaussianNoise(mean: float = 0.0, std: float = 1.0, return_transform: bool = False, same_on_batch: bool = False, p: float = 0.5)[source]¶

Add gaussian noise to a batch of multi-dimensional images.

Parameters

mean (float) – The mean of the gaussian distribution. Default: 0.
std (float) – The standard deviation of the gaussian distribution. Default: 1.
return_transform (bool) – if True return the matrix describing the transformation applied to each input tensor. If False and the input is a tuple the applied transformation wont be concatenated.
same_on_batch (bool) – apply the same transformation across the batch. Default: False.
p (float) – probability of applying the transformation. Default value is 0.5.

Examples

>>> rng = torch.manual_seed(0)
>>> img = torch.ones(1, 1, 2, 2)
>>> RandomGaussianNoise(mean=0., std=1., p=1.)(img)
tensor([[[[ 2.5410,  0.7066],
          [-1.1788,  1.5684]]]])

apply_adjust_brightness(input: torch.Tensor, params: Dict[str, torch.Tensor]) → torch.Tensor[source]¶

Apply brightness adjustment.

Wrapper for adjust_brightness for Torchvision-like param settings.

Parameters

input (torch.Tensor) – Tensor to be transformed with shape \((*, C, H, W)\).
params (Dict[str, torch.Tensor]) –
- params[‘brightness_factor’]: Brightness adjust factor per element in the batch. 0 gives a black image, 1 does not modify the input image and 2 gives a white image, while any other number modify the brightness.

Returns

Adjusted image with shape \((B, C, H, W)\).

Return type

Transforms2D¶

Transforms3D¶

Normalizations¶