Skip to content

Parameter initialization

Specialized parameter initializers, extending jax.nn.initializers.

For some initialization schemes, the bias depends on the number of input features, which cannot be determined from the shape of the bias array (see, e.g, Hoedt normal initializer). To handle such cases klax provides a custom klax.Initializer protocol that generalizes upon the JAX API, while ensuring compatibility with all jax.nn.initializers.

klax.Initializer ¤

Protocol for initializers, generalizing jax.nn.initializers.Initializer.

Some advanced initialization schemes initialize the bias depending on the number of input features (fan_in). However, from the bias shape alone fan_in cannot be computed. This protocol specifies an initializer that is supplied with fan_in explicitly, enabling advanced bias initialization.

__call__(key, shape, fan_in, dtype=float64) ¤


klax.hoedt_normal(in_axis=-2, dtype=float64) ¤

Build a Hoedt normal initializer (for positivity constrained weights).

A Hoedt normal initializer is designed for weights with positivity constraint, such as in klax.nn.FICNN. It samples weights according to a log-normal distribution derived by studying signal propagation through layers with non-negative weights.

Tip

This initiailzation should be paired with the klax.hoedt_bias initializer for biases of constrained layers.

PARAMETER DESCRIPTION
in_axis

Axis of the input dimension in the weights array.

TYPE: int DEFAULT: -2

dtype

The dtype of the weights.

TYPE: Any DEFAULT: float64

RETURNS DESCRIPTION
Initializer

A jax.nn.initializers.Initializer.


klax.hoedt_bias() ¤

Build a Hoedt bias initializer (for layers with positivity constrained weights).

A Hoedt bias initializer is designed for the unconstrained biases in linear layers where the weights are constrained to be positive, such as in klax.nn.FICNN. It intializes biases to vectors of a constant value, computed from the number of input features (fan_in).

Tip

This initialization should be paired with the klax.hoedt_normal initializer for the positivity constraint weights.

RETURNS DESCRIPTION
Initializer

An Initializer.