HNSSpoke

Bases: tf.keras.layers.Layer

A spoke in the hub-and-spokes model with time-averaging output.

This layer is a thin wrapper around MultiInputTimeAveraging, The differences are:

1. Catering for None input by explicitly defining units.
2. For clarity, in call(), separating the original list of inputs into the clamped input: inputs and last states: last_states.

Parameters:

Name	Type	Description	Default
tau	float	Time-averaging parameter, from 0 to 1.	required
units	int	Number of units in the layer. Explicitly set the input dimension to cater for None input.	required
activation	str	Activation function to use.	'sigmoid'
use_bias	bool	Whether the layer uses a bias vector.	True
kwargs	dict	Additional keyword arguments passed to MultiInputTimeAveraging.	{}

Forward pass:

$$a_t = \tau \cdot act(x_t + \sum_i^n s_{i, t-1} + b) + (1-\tau) \cdot a_{t-1}$$

Warning

"$\cdot$" is element-wise multiplication.

• $a_t$: activation at time $t$.
• $\tau$: time constant, smaller means slower temporal dynamics.
• $act$: activation function (provided by this layer if activation is used).
• $x_{t}$: input at time $t$ coming from input $i$.
• $s_{i, t-1}$: last state from $i$ at time $t-1$.
• $n$: number of cross-tick state layers.
• $b$: bias vector, provided by this layer if use_bias is True (Default).
• $a_{t-1}$: time-averaging state at time $t-1$, stored in self.time_averaging.states.

call(inputs=None, last_states=None)

Forward pass.

Parameters:

Name	Type	Description	Default
inputs	tf.Tensor	clamped inputs to spoke.	None
last_states	List[tf.Tensor]	list of states from last ticks projection: $s_{i, t-1}$.	None

reset_states()

Reset time-averaging states.