Parameter (`pmrf.Parameter`)

class pmrf.parameters.Parameter(value, prior=None, fixed=False, scale=1.0, name=None)[source]

Bases: Module

Overview

A container for a parameter, usually used within a Model.

This class serves as the fundamental building block for defining the tunable or fixed parameters within a paramrf Model and for fitting. It is designed to be a flexible container that behaves like a standard numerical type (e.g., a numpy.ndarray) while holding additional metadata for model fitting and analysis.

Usage

Use in mathematical operations just like a JAX/numpy array
Parameter objects are JAX PyTrees, making them compatible with JAX’s transformations (jit, grad, etc.).
Mark as fixed (which is honoured by fitting and sampling routines)
Associate priors, specified as numpyro distributions (uniform, normal etc.)

Examples

import pmrf as prf
import jax.numpy as jnp

# A simple, single-valued parameter, initialized with a float
p1 = prf.Parameter(value=1.0e-12, name='C1')

# This parameter can be used in calculations directly (scaling is done during casting)
impedance = 1 / (2j * jnp.pi * 1e9 * p1)
print(f"Impedance: {impedance}")

# A parameter that is fixed and will not be optimized during a fit
p2 = prf.Parameter(value=50.0, fixed=True, name='Z0')

# A parameter with a uniform distribution prior
# Factory functions are a convenient way to create these
p3 = prf.Uniform(min=0.9e-9, max=1.1e-9, name='L1')

# The parameter's value is initialized to the mean of the distribution
print(f"Initial value of L1: {p3.value}")

value: Array

prior: Distribution | None = None

fixed: bool = False

scale: float = 1.0

name: str | None = None

property ndim: int: The number of free dimensions for this parameter.

property min: array

The unscaled minimum value of the parameter’s distribution (MIN_PERCENTILE quantile).

Returns:: The minimum value, or -np.inf if no distribution is set.
Return type:: jnp.array

property max: array

The unscaled maximum value of the parameter’s distribution (MAX_PERCENTILE quantile).

Returns:: The maximum value, or np.inf if no distribution is set.
Return type:: jnp.array

with_value(value)[source]

Return a copy with a new unscaled value.

Parameters:: value (jnp.array) – The new unscaled value to set.
Returns:: A copy of this object with value replaced.
Return type:: Parameter

with_prior(prior)[source]

Return a copy with a new prior distribution.

Parameters:: prior (numpyro.distributions.Distribution) – The prior to associate with this parameter.
Returns:: A copy of this object with prior replaced.
Return type:: Parameter

flatten(separator='_')[source]

Flattens self, either returning a single Parameter if the internal parameter is scalar, or a list. If the internal prior cannot be separated, this will raise an Exception.

Returns:: The raveled parameters.
Return type:: ‘Parameter’ | list[‘Parameter’]

as_fixed()[source]

Returns a fixed copy of self.

Returns:: The new, fixed parameter.
Return type:: Parameter

as_free()[source]

Returns a copy of self with fixed set to False.

Returns:: The new, fixed parameter.
Return type:: Parameter

copy()[source]

Return a shallow copy.

Returns:: A copy created via dataclasses.replace.
Return type:: Parameter

to_json()[source]

Serialize the parameter to a JSON string.

Returns:: A JSON-formatted string containing value, prior, fixed, scale and name.
Return type:: str

classmethod from_json(s)[source]

Deserialize a parameter from a JSON string.

Parameters:: s (str) – The JSON string produced by to_json().
Returns:: A reconstructed Parameter instance.
Return type:: Parameter

Parameters:

value (Array)
prior (Distribution | None)
fixed (bool)
scale (float)
name (str | None)

class pmrf.parameters.ParameterGroup(param_names, prior=None)[source]

Bases: object

A metadata class that groups a set of named flat parameters and defines any relationships between them.

Construct a ParameterGroup.

Parameters:

param_names (list[str] | dict[str, Parameter]) – The names of the flattened parameters (or a mapping to parameters).
prior (numpyro.distributions.Distribution, optional) – An optional prior over the flattened parameters.

param_names: list[str]

prior: Distribution | None = None

property num_flat_params

Number of flattened parameters.

Returns:: The count of names in param_names.
Return type:: int