deepdraw.data.transforms#

Image transformations for our pipelines.

Differences between methods here and those from torchvision.transforms is that these support multiple simultaneous image inputs, which are required to feed segmentation networks (e.g. image and labels or masks). We also take care of data augmentations, in which random flipping and rotation needs to be applied across all input images, but color jittering, for example, only on the input image.

Classes

`AutoLevel16to8`()	Converts multiple 16-bit images to 8-bit representations using "auto- level".
`CenterCrop`(size)
`ColorJitter`([p])	Randomly applies a color jitter transformation on the first image.
`Compose`(transforms)
`Crop`(i, j, h, w)	Crops multiple images at the given coordinates.
`GaussianBlur`([p])	Randomly applies a gaussian blur transformation on the first image.
`GroundTruthCrop`([reference, extra_area])	Crop image in a square keeping only the area with the ground truth.
`Pad`(padding[, fill, padding_mode])
`RandomHorizontalFlip`([p])	Randomly flips all input images horizontally.
`RandomRotation`([p])	Randomly rotates all input images by the same amount.
`RandomVerticalFlip`([p])	Randomly flips all input images vertically.
`Resize`(size[, interpolation, max_size, ...])
`ShrinkIntoSquare`([reference, threshold])	Crops black borders and then resize to a square with minimal padding.
`SingleAutoLevel16to8`()	Converts a 16-bit image to 8-bit representation using "auto-level".
`SingleCrop`(i, j, h, w)	Crops one image at the given coordinates.
`SingleToRGB`()	Converts from any input format to RGB, using an ADAPTIVE conversion.
`ToRGB`()	Converts from any input format to RGB, using an ADAPTIVE conversion.
`ToTensor`()
`TupleMixin`()	Adds support to work with tuples of objects to torchvision transforms.

class deepdraw.data.transforms.TupleMixin[source]#

Bases: object

Adds support to work with tuples of objects to torchvision transforms.

class deepdraw.data.transforms.CenterCrop(size)[source]#: Bases: TupleMixin, CenterCrop

class deepdraw.data.transforms.Pad(padding, fill=0, padding_mode='constant')[source]#: Bases: TupleMixin, Pad

class deepdraw.data.transforms.Resize(size, interpolation=InterpolationMode.BILINEAR, max_size=None, antialias='warn')[source]#: Bases: TupleMixin, Resize

class deepdraw.data.transforms.ToTensor[source]#: Bases: TupleMixin, ToTensor

class deepdraw.data.transforms.Compose(transforms)[source]#: Bases: Compose

class deepdraw.data.transforms.SingleCrop(i, j, h, w)[source]#

Bases: object

Crops one image at the given coordinates.

i#

upper pixel coordinate.

Type:: int

j#

left pixel coordinate.

Type:: int

h#

height of the cropped image.

Type:: int

w#

width of the cropped image.

Type:: int

class deepdraw.data.transforms.Crop(i, j, h, w)[source]#

Bases: TupleMixin, SingleCrop

Crops multiple images at the given coordinates.

i#

upper pixel coordinate.

Type:: int

j#

left pixel coordinate.

Type:: int

h#

height of the cropped image.

Type:: int

w#

width of the cropped image.

Type:: int

class deepdraw.data.transforms.SingleAutoLevel16to8[source]#

Bases: object

Converts a 16-bit image to 8-bit representation using “auto-level”.

This transform assumes that the input image is gray-scaled.

To auto-level, we calculate the maximum and the minimum of the image, and consider such a range should be mapped to the [0,255] range of the destination image.

class deepdraw.data.transforms.AutoLevel16to8[source]#

Bases: TupleMixin, SingleAutoLevel16to8

Converts multiple 16-bit images to 8-bit representations using “auto- level”.

This transform assumes that the input images are gray-scaled.

To auto-level, we calculate the maximum and the minimum of the image, and consider such a range should be mapped to the [0,255] range of the destination image.

class deepdraw.data.transforms.SingleToRGB[source]#

Bases: object

Converts from any input format to RGB, using an ADAPTIVE conversion.

This transform takes the input image and converts it to RGB using py:method:PIL.Image.Image.convert, with mode=’RGB’ and using all other defaults. This may be aggressive if applied to 16-bit images without further considerations.

class deepdraw.data.transforms.ToRGB[source]#

Bases: TupleMixin, SingleToRGB

Converts from any input format to RGB, using an ADAPTIVE conversion.

class deepdraw.data.transforms.RandomHorizontalFlip(p=0.5)[source]#

Bases: RandomHorizontalFlip

Randomly flips all input images horizontally.

class deepdraw.data.transforms.RandomVerticalFlip(p=0.5)[source]#

Bases: RandomVerticalFlip

Randomly flips all input images vertically.

class deepdraw.data.transforms.RandomRotation(p=0.5, **kwargs)[source]#

Bases: RandomRotation

Randomly rotates all input images by the same amount.

Unlike the current torchvision implementation, we also accept a probability for applying the rotation.

Parameters:

p (float, Optional) – probability at which the operation is applied
**kwargs (dict) –
passed to parent. Notice that, if not set, we use the following defaults here for the underlying transform from torchvision:
- degrees: 15
- interpolation: torchvision.transforms.functional.InterpolationMode.BILINEAR

class deepdraw.data.transforms.ColorJitter(p=0.5, **kwargs)[source]#

Bases: ColorJitter

Randomly applies a color jitter transformation on the first image.

Notice this transform extension, unlike others in this module, only affects the first image passed as input argument. Unlike the current torchvision implementation, we also accept a probability for applying the jitter.

Parameters:

p (float, Optional) – probability at which the operation is applied
**kwargs (dict) –
passed to parent. Notice that, if not set, we use the following defaults here for the underlying transform from torchvision:
- brightness: 0.3
- contrast: 0.3
- saturation: 0.02
- hue: 0.02

class deepdraw.data.transforms.ShrinkIntoSquare(reference=0, threshold=0)[source]#

Bases: object

Crops black borders and then resize to a square with minimal padding.

This transform can crop all the images by removing the black pixels in the width and height until it finds a non-black pixel. Then, expands the image back until it makes a square with minimal size.

Parameters:

reference (int, Optional) – Which reference part of the sample to use for cropping black borders. If not set, use the first object on the sample (typically, the image).
threshold (int, Optional) – Threshold to use for when considering what is a “black” border

class deepdraw.data.transforms.GaussianBlur(p=0.5, **kwargs)[source]#

Bases: GaussianBlur

Randomly applies a gaussian blur transformation on the first image.

Parameters:

p (float, Optional) – probability at which the operation is applied
**kwargs (dict) –
passed to parent. Notice that, if not set, we use the following defaults here for the underlying transform from torchvision:
- kernel_size: (5, 5)
- sigma: (0.1, 5)

class deepdraw.data.transforms.GroundTruthCrop(reference=1, extra_area=0.0)[source]#

Bases: object

Crop image in a square keeping only the area with the ground truth.

This transform can crop all images given a ground-truth mask as reference. Notice that the crop will result in a square image at the end, which means that it will keep the bigger dimension and adjust the smaller one to fit into a square. There’s an option to add extra area around the gt bounding box. If resulting dimensions are larger than the boundaries of the image, minimal padding will be done to keep the image in a square shape.

Parameters:

reference (int, Optional) – Which reference part of the sample to use for getting coordinates. If not set, use the second object on the sample (typically, the mask).
extra_area (float, Optional) – Multiplier that will add the extra area around the ground-truth bounding box. Example: 0.1 will result in a crop with dimensions of the largest side increased by 10%. If not set, the default will be 0 (only the ground-truth box).