mml.core.data_loading.augmentations.albumentations

class AlbumentationsAugmentationModule

Bases: AugmentationModule

Albumentations augmentation module.

Basic supported dict entries are [‘image’, ‘mask’, ‘bboxes’, ‘keypoints’].

An AutoAlbument generated pipeline is available via the Identifier “ImageNetAA” (no parameters). “RandAugment” with parameters is also provided on top. See get_rand_augment() for details.

Note that once bboxes and keypoints will be supported than composition will include the respective parameters https://albumentations.ai/docs/api_reference/core/composition/. Futhermore “additional_targets” might need to be defined. To check whether a certain augmentation supports a specific target type see https://albumentations.ai/docs/getting_started/transforms_and_targets/.

__init__(device: str, cfg: ListConfig, is_first: bool, is_last: bool, means: RGBInfo | None, stds: RGBInfo | None, floatify: bool = False, tensorize: bool = True)

static compose_albumentations_transforms(pp_trans: List[BasicTransform], aug_trans: List[BasicTransform], mean: RGBInfo | None = None, std: RGBInfo | None = None) → Compose

Composes transforms from preprocessing, data augmentation, normalization and tensorization.

Parameters:

• pp_trans – list of preprocessing transforms, can be empty

• aug_trans – list of data augmentation transforms, can be empty

• mean – mean channel values to normalize input data

• std – standard deviation values per channel to normalize input data

Returns:

a composed albumentation pipeline

static from_cfg(aug_config: ListConfig) → List[Callable[[Dict[str, Any]], Dict[str, Any]]]

Takes a config and returns a list of corresponding transforms.

Parameters:

aug_config (DictConfig) – see configs/augmentations/default.yaml for an example, the “cpu” attribute is what has to be passed to this function.

Returns:

a list of albumentation transforms

static get_rand_augment(number: int, magnitude: int, p: float = 1.0, mode: str = 'all', cut_out: bool = False) → BaseCompose

Gets RandAugment transform. For details see https://arxiv.org/abs/1909.13719.

Parameters:

• number – number of transforms to be applied (excluding cut_out if active)

• magnitude – int between 0 and 9 determining strength of transformation

• p – probability to apply RandAugment

• mode – either ‘geo’ for geometrical transforms, ‘color’ for color transforms or ‘all’ for both of them

• cut_out – indicating if cutout should be applied

Returns:

an albumentations transform