mask_autoencoder

Masked reconstruction with portable computation and explicit TF/Torch steps.

ReconstructionclassMasked target patches and their corresponding predicted patches.
ReconstructionLossclassReconstruction objective together with the patches used to compute it.
MaskedAutoencoderclassMasked reconstruction with Keras fit() and independently callable objectives.

Reconstruction()

Masked target patches and their corresponding predicted patches.

targets: Tensor

Input patches gathered at the selected mask indices.

predictions: Tensor

Reconstructed patches gathered at the same indices.

ReconstructionLoss()

Reconstruction objective together with the patches used to compute it.

loss: Tensor

Loss tensor computed by the configured reconstruction objective.

targets: Tensor

Masked input patches.

predictions: Tensor

Corresponding reconstructed patches.

MaskedAutoencoder(
    patch_layer: Callable[[Array], Tensor],
    patch_encoder: Callable[[Array], tuple[Tensor, Tensor, Tensor, Tensor, Tensor]],
    encoder: keras.Model,
    decoder: keras.Model,
    **kwargs: Any = {},
) -> None

Masked reconstruction with Keras fit() and independently callable objectives.

call() returns (targets, predictions). training controls layer state, not masks. See https://ambiqai.github.io/helia-edge/getting-started/backends/ for serialization, native-loop use and support limits.

patch_layer = patch_layer

patch_encoder = patch_encoder

encoder = encoder

decoder = decoder

call(inputs: Array, training: bool = False) -> Reconstruction

reconstruction_targets(x: Array, training: bool = False) -> Reconstruction

Return masked (target_patches, predicted_patches), without an objective.

calculate_loss(x: Array, test: bool = False) -> ReconstructionLoss

Return (compiled total loss, targets, predictions); preserve the old API.

compute_loss(
    x: Any = None,
    y: Any = None,
    y_pred: Any = None,
    sample_weight: Any = None,
    training: bool = True,
) -> Tensor

compute_metrics(x: Any, y: Any, y_pred: Any, sample_weight: Any = None) -> dict[str, Tensor]

train_step(data: Any) -> dict[str, Tensor]

test_step(data: Any) -> dict[str, Tensor]

get_config() -> dict[str, Any]

from_config(config: dict[str, Any], custom_objects: dict[str, Any] | None = None) -> Self

classmethod