fit_sample

pmrf.fitting.fit_sample(model: ModelT, data: Array | Network | NetworkCollection, frequency: Frequency | None = None, solver: AbstractJointSampler | AbstractSplitSampler | AbstractHypercubeSampler | None = None, *, features: str | list[str] | Callable = 's', likelihood: Callable[[Array], AbstractDistribution] | list[Callable[[Array], AbstractDistribution]] = None, noise: Param | Callable[[Array], Array] = None, loss: Callable[[Array, Array], Array] = None, discrepancy: Callable[[Array, Array], AbstractDistribution] | None = None, temperature: float = None, **kwargs) → FitResult[ModelT]

Conditions an RF model on measured data using Bayesian sampling.

This high-level function handles data format formatting (e.g., extracting arrays from scikit-rf Networks) and forwards to pmrf.infer.sample().

Parameters:

  • model (Model) – The RF model to fit.

  • data (jnp.ndarray | skrf.Network | NetworkCollection) – The data to condition on. Can either be a JAX array, a skrf.Network, or a pmrf.NetworkCollection.

  • frequency (Frequency | None, default=None) – The frequency sweep. Required if data is a raw array; otherwise automatically extracted from the Network object.

  • solver (AbstractSampler, optional) – The sampler to use. See pmrf.infer for available solvers.

  • features (str | list[str] | Callable[[Model, Frequency], jnp.ndarray], default='s') – The RF features to condition on. Can either be function, a callable PyTree with optional parameters, or a string, in which case a ‘feature’ evaluator is created (see pmrf.evaluators.Feature). Defaults to all S-parameters.

  • likelihood (Callable[[jnp.ndarray], AbstractDistribution], optional) – The likelihood model, which accepts a model prediction (in event space) and returns a distribution representing the probability of observing the data. Can be a function or a callable PyTree with optional parameters. See pmrf.likelihoods for common likelihoods. Mutually exclusive with loss.

  • noise (prf.Param | Callable[[jnp.ndarray], jnp.ndarray], optional) – Likelihood noise (variance), either a fixed parameter, or a callable that accepts a model prediction (in event space) and returns noise parameters for a Gaussian likelihood. Mutually exclusive with likelihood and loss. For the function case, can be a callable PyTree with optional parameters. See pmrf.noise_models for built-in noise models. Defaults to None, in which case uniform variance from 0.0 to 0.1 is constructed internally.

  • loss (Callable[[jnp.ndarray, jnp.ndarray], jnp.ndarray], optional) – A loss function between the model prediction and the data to construct a Gibbs measure. Can be a function or a callable PyTree with optional parameters. Mutually exclusive with likelihood and noise. If neither loss nor likelihood is passed, a pmrf.likelihoods.GaussianLikelihood is constructed.

  • discrepancy (Callable[[jnp.ndarray, jnp.ndarray], jnp.ndarray | AbstractDistribution], optional) – A discrepancy model, which caters for the discrepancy between the model and measured data. Can either be a function, or a callable PyTree with optional parameters. To use a Gaussian process as a discrepancy model, see pmrf.discrepancys.GaussianProcess.

  • temperature (float, optional) – The temperature value for generalized Bayesian optimization. Only used when loss is not None. Defaults to 1.0 internally.

  • **kwargs (dict) – Additional keyword arguments passed to the underlying solver.

Returns:

The result containing the model maximum likelikhood estimate model with an empirical posterior.

Return type:

FitResult