streaming

Streaming Layers API

Recurrent cells for streaming models whose state is an explicit model input and output. One model call processes one step: the caller feeds each state_out_k output back as the state_in_k input of the next call, and feeds zeros at the start of an independent sequence.

Classes

Name	Description
StreamingLSTMCell	One LSTM step with explicit hidden and cell state

Functions

Name	Description
state_input	Model input for state pair k
state_output	Model output for state pair k

Machine-readable model

• StreamingLSTMCellclassOne LSTM step: [x, h, c] -> [h', c'].
• state_inputfunctionModel input stateink for state pair k.
• state_outputfunctionName x as the model output stateoutk of state pair k.

class

StreamingLSTMCell

Python

One LSTM step: [x, h, c] -> [h', c'].

helia_edge/layers/streaming.py:22

StreamingLSTMCell(units: int, use_bias: bool = True, unit_forget_bias: bool = True, **kwargs={})

One LSTM step: [x, h, c] -> [h', c'].

The weights have the layout of keras.layers.LSTMCell: kernel (features, 4 * units), recurrent_kernel (units, 4 * units) and bias (4 * units), with gates in the order input, forget, cell, output. Weights of a keras.layers.LSTM or LSTMCell load unchanged.

c’ = f * c + i * tanh(z_c) and h’ = o * tanh(c’), where z = x W + h R + b and i, f, o are the sigmoid of their slices of z.

Parameters

Parameters of StreamingLSTMCell

Name	Type	Default	Description
units	int	Required	Size of the hidden and cell state.
use_bias	bool	True	Add ``bias`` to the gates.
unit_forget_bias	bool	True	Initialize the forget gate's bias to one and the others to zero, as ``keras.layers.LSTMCell`` does.

attribute

units

Python

helia_edge/layers/streaming.py:42

units = units

attribute

use_bias

Python

helia_edge/layers/streaming.py:43

use_bias = use_bias

attribute

unit_forget_bias

Python

helia_edge/layers/streaming.py:44

unit_forget_bias = unit_forget_bias

method

build

Python

helia_edge/layers/streaming.py:46

build(input_shape)

method

call

Python

helia_edge/layers/streaming.py:66

call(inputs)

method

compute_output_shape

Python

helia_edge/layers/streaming.py:79

compute_output_shape(input_shape)

method

get_config

Python

helia_edge/layers/streaming.py:83

get_config()

function

state_input

Python

Model input stateink for state pair k.

helia_edge/layers/streaming.py:92

state_input(k: int, shape: tuple[int, ...], batch_size: int | None = None) -> keras.KerasTensor

Model input state_in_k for state pair k.

Parameters

Parameters of state_input

Name	Type	Default	Description
k	int	Required	State pair index; ``state_out_k`` is the matching output.
shape	tuple[int, ...]	Required	State shape without the batch dimension.
batch_size	int | None	None	Fixed batch size, or None.

Returns

Returns of state_input

Type	Description
keras.KerasTensor	keras.KerasTensor: The input tensor.

function

state_output

Python

Name x as the model output stateoutk of state pair k.

helia_edge/layers/streaming.py:106

state_output(k: int, x: keras.KerasTensor) -> keras.KerasTensor

Name x as the model output state_out_k of state pair k.

Parameters

Parameters of state_output

Name	Type	Default	Description
k	int	Required	State pair index; ``state_in_k`` is the matching input.
x	keras.KerasTensor	Required	The next state.

Returns

Returns of state_output

Type	Description
keras.KerasTensor	keras.KerasTensor: ``x`` through an identity layer named ``state_out_k``.