base_augmentation

Unified public-Keras hooks for deterministic preprocessing and augmentation.

BaseAugmentationParamsclassShared construction-time configuration; never validates live tensors.
BaseAugmentationclassBase for rank-specific transforms, with one sampling/application contract.
BaseAugmentation1DclassOne-dimensional signals with optional batch axis.
BaseAugmentation2DclassTwo-dimensional images with optional batch axis.

BaseAugmentationParams()

Shared construction-time configuration; never validates live tensors.

model_config = ConfigDict(extra='forbid', frozen=True)

seed: int | None = None

auto_vectorize: bool = False

data_format: Literal['channels_first', 'channels_last'] = 'channels_last'

device: str = 'cpu'

aligned_targets: tuple[str, ...] | None = None

aligned_masks: tuple[str, ...] | None = None

BaseAugmentation(
    seed=None,
    auto_vectorize=False,
    data_format=None,
    device='cpu',
    aligned_targets=None,
    aligned_masks=None,
    **kwargs={},
)

Base for rank-specific transforms, with one sampling/application contract.

Override augment_samples or augment_sample and optionally get_random_transformations. Sample parameters once per signal batch; joint transforms reuse them across aligned signals, targets and masks. Per-example parameter leaves must have a leading batch axis. Deterministic subclasses set training_only=False. Bases are not serialized transforms.

SAMPLES = 'data'

LABELS = 'labels'

TARGETS = 'targets'

TRANSFORMS = 'transforms'

NDIMS = 4

training_only = True

joint = False

seed = params.seed

auto_vectorize = params.auto_vectorize

data_format = params.data_format

device = params.device

aligned_targets = params.aligned_targets

aligned_masks = params.aligned_masks

generator = keras.random.SeedGenerator(self.seed)

random_generator

ch_axis

data_axis

height_axis

width_axis

get_random_transformations(input_shape)

Return batched parameter tensors, or None for deterministic transforms.

augment_sample(inputs)

augment_samples(inputs)

augment_targets(inputs)

Apply geometry to selected targets using the signal parameters.

augment_masks(inputs)

Apply geometry to selected masks using the signal parameters.

batch_augment(inputs, transformations=None)

Apply to a tensor or schema, with optional pre-sampled parameters.

call(inputs, training=None, transformations=None)

get_config()

BaseAugmentation1D(
    seed=None,
    auto_vectorize=False,
    data_format=None,
    device='cpu',
    aligned_targets=None,
    aligned_masks=None,
    **kwargs={},
)

One-dimensional signals with optional batch axis.

Base class: BaseAugmentation.

Inherited from BaseAugmentation: augment_masks(), augment_sample(), augment_samples(), augment_targets(), batch_augment(), call(), get_config(), get_random_transformations().

NDIMS = 3

BaseAugmentation2D(
    seed=None,
    auto_vectorize=False,
    data_format=None,
    device='cpu',
    aligned_targets=None,
    aligned_masks=None,
    **kwargs={},
)

Two-dimensional images with optional batch axis.

Base class: BaseAugmentation.

Inherited from BaseAugmentation: augment_masks(), augment_sample(), augment_samples(), augment_targets(), batch_augment(), call(), get_config(), get_random_transformations().

NDIMS = 4