Parameter

class nyx.core.parameter.Parameter

Bases: Module

A trainable physical parameter with an explicit characteristic scale.

Variables:

• factor (jax.Array) – The only pytree leaf; O(1) for a well-chosen scale.

• scale (float) – Characteristic scale in the unconstrained space, static for JIT. Change it via autoscale().

• per_obs (bool) – Whether the parameter carries a leading (nobs, ...) axis after tile_per_obs().

• frozen (bool) – Whether the parameter is excluded from optimization.

• transform (str or None) – None (unconstrained), 'log' or 'softplus' for a positive quantity, or 'tanh' for one confined to (-1, 1).

Attributes Summary

dvalue_dfactor	d value / d factor, elementwise; scale when untransformed.
factor
frozen
per_obs
scale
transform
value	Physical value transform(factor * scale).

Methods Summary

freeze()	Return a copy with frozen=True.
from_value(value[, scale, per_obs, frozen, ...])	Construct from a physical value.
unfreeze()	Return a copy with frozen=False.

Attributes Documentation

dvalue_dfactor

d value / d factor, elementwise; scale when untransformed.

factor = <dataclasses._MISSING_TYPE object>

frozen = False

per_obs = False

scale = 1.0

transform = None

value

Physical value transform(factor * scale).

Methods Documentation

freeze()

Return a copy with frozen=True.

classmethod from_value(value, scale=None, per_obs=False, frozen=False, transform=None)

Construct from a physical value.

Parameters:

• value (array-like) – Physical value; must lie in the domain of transform.

• scale (float or None) – Characteristic scale. When None, 10**floor(log10(max|value|)) of the unconstrained value; 1.0 if that is zero or transform is set.

• per_obs (bool)

• frozen (bool)

• transform (str or None)

Return type:

Parameter

Raises:

ValueError – If transform is unknown, or value lies outside its domain.

unfreeze()

Return a copy with frozen=False.

__init__(factor, scale=1.0, per_obs=False, frozen=False, transform=None)

static __new__(cls, *args, **kwargs)