deepdraw.data.utils#

Common utilities.

Functions

`invert_mode1_image`(img)	Inverts a binary PIL image (mode == `"1"`)
`overlayed_image`(img, label[, mask, ...])	Creates an image showing existing labels and mask.
`subtract_mode1_images`(img1, img2)	Returns a new image that represents `img1 - img2`

Classes

`SSLDataset`(labelled, unlabelled)	PyTorch dataset wrapper around labelled and unlabelled sample lists.
`SampleListDataset`(samples[, transforms])	PyTorch dataset wrapper around Sample lists.

deepdraw.data.utils.invert_mode1_image(img)[source]#: Inverts a binary PIL image (mode == "1")

deepdraw.data.utils.subtract_mode1_images(img1, img2)[source]#: Returns a new image that represents img1 - img2

deepdraw.data.utils.overlayed_image(img, label, mask=None, label_color=(0, 255, 0), mask_color=(0, 0, 255), alpha=0.4)[source]#

Creates an image showing existing labels and mask.

This function creates a new representation of the input image img overlaying a green mask for labelled objects, and a red mask for parts of the image that should be ignored (negative mask). By looking at this representation, it shall be possible to verify if the dataset/loader is yielding images correctly.

Parameters:

img (PIL.Image.Image) – An RGB PIL image that represents the original image for analysis
label (PIL.Image.Image) – A PIL image in any mode that represents the labelled elements in the image. In case of images in mode “L” or “1”, white pixels represent the labelled object. Black-er pixels represent background.
mask (py:class:PIL.Image.Image, Optional) – A PIL image in mode “1” that represents the mask for the image. White pixels indicate where content should be used, black pixels, content to be ignored.
label_color (py:class:tuple, Optional) – A tuple with three integer entries indicating the RGB color to be used for labels. Only used if label.mode is “1” or “L”.
mask_color (py:class:tuple, Optional) – A tuple with three integer entries indicating the RGB color to be used for the mask-negative (black parts in the original mask).
alpha (py:class:float, Optional) – A float that indicates how much of blending should be performed between the label, mask and the original image.

Returns:

image (PIL.Image.Image) – A new image overlaying the original image, object labels (in green) and what is to be considered parts to be masked-out (i.e. a representation of a negative of the mask).

class deepdraw.data.utils.SampleListDataset(samples, transforms=[])[source]#

Bases: Dataset

PyTorch dataset wrapper around Sample lists.

A transform object can be passed that will be applied to the image, ground truth and mask (if present).

It supports indexing such that dataset[i] can be used to get the i-th sample.

Parameters:

samples (list) – A list of deepdraw.data.sample.Sample objects
transforms (list, Optional) – a list of transformations to be applied to both image and ground-truth data. Notice a last transform (deepdraw.data.transforms.ToTensor) is always applied - you do not need to add that.

property transforms#

copy(transforms=None)[source]#

Returns a deep copy of itself, optionally resetting transforms.

Parameters:: transforms (list, Optional) – An optional list of transforms to set in the copy. If not specified, use self.transforms.

keys()[source]#: Generator producing all keys for all samples.

all_keys_match(other)[source]#: Compares all keys to other, return True if all match.

class deepdraw.data.utils.SSLDataset(labelled, unlabelled)[source]#

Bases: Dataset

PyTorch dataset wrapper around labelled and unlabelled sample lists.

Yields elements of the form:

[key, image, ground-truth, [mask,] unlabelled-key, unlabelled-image]

The size of the dataset is the same as the labelled dataset.

Indexing works by selecting the right element on the labelled dataset, and randomly picking another one from the unlabelled dataset

Parameters:

labelled (torch.utils.data.Dataset) – Labelled dataset (must have “mask” and “label” entries for every sample)
unlabelled (torch.utils.data.Dataset) – Unlabelled dataset (may have “mask” and “label” entries for every sample, but are ignored)

keys()[source]#: Generator producing all keys for all samples.

all_keys_match(other)[source]#: Compares all keys to other, return True if all match.