normalization

Normalization Layers API

This module provides classes to build normalization layers.

Classes

Name	Description
`LayerNormalization`	Keras layer normalization that also runs on the Torch backend for any axes

Functions

Name	Description
`layer_normalization`	Layer normalization
`batch_normalization`	Batch normalization
`normalization`	Normalization builder layer

Please check Keras Normalization Layers for additional layers.

Machine-readable model

LayerNormalizationclassKeras LayerNormalization that also normalizes non-trailing axes on the Torch backend.
layer_normalizationfunctionLayer normalization
batch_normalizationfunctionBatch normalization
normalizationfunctionCreates normalization layer based on type

LayerNormalization()

Keras LayerNormalization that also normalizes non-trailing axes on the Torch backend.

Keras 3’s Torch backend normalizes trailing axes only, so axes such as (1, 2) of a (batch, height, width, channels) input fail there. On Torch, this layer moves the normalized axes last, normalizes and moves them back. On other backends it is the Keras layer, with the same weights, configuration and graph.

Saved models record this class as helia_edge>LayerNormalization: load them with helia_edge.models.load_model, or call helia_edge.register_keras_serializables() before keras.saving.load_model.

call(inputs)

layer_normalization(name: str | None = None, axis: int | tuple[int] | None = None, scale: bool = True) -> keras.Layer

Layer normalization

If axis is None, this layer will infer based on the input tensor shape:

If rank is 4 (B, H, W, C), normalize over H, W
If rank is 3 (B, T, C), normalize over T
If rank is 2 (B, C), normalize over C

Parameters

Parameters of layer_normalization
Name	Type	Default	Description
`name`	`str \| None`	None	Layer name. Defaults to None.
`axis`	`int \| tuple[int] \| None`	None	Axis. Defaults to None.
`scale`	`bool`	True	Scale. Defaults to True.

Returns

Returns of layer_normalization
Type	Description
`keras.Layer`	keras.Layer: Layer

batch_normalization(name: str | None = None, momentum=0.9, epsilon=0.001, axis: int | None = -1) -> keras.Layer

Batch normalization

Parameters

Parameters of batch_normalization
Name	Type	Default	Description
`name`	`str \| None`	None	Layer name. Defaults to None.
`momentum`	`float`	0.9	Momentum. Defaults to 0.9.
`epsilon`	`float`	0.001	Epsilon. Defaults to 1e-3.
`axis`	`int \| None`	-1	Axis. Defaults to None.

Returns

Returns of batch_normalization
Type	Description
`keras.Layer`	keras.Layer: Layer

normalization(norm: str, name: str, **kwargs={}) -> keras.Layer

Creates normalization layer based on type

Parameters

Parameters of normalization
Name	Type	Default	Description
`norm`	`str`	Required	Normalization type
`name`	`str`	Required	Name

Returns

Returns of normalization
Value	Type	Description
`KerasLayer`	`keras.Layer`	Layer