nessai.model

nessai.model#

Object for defining the use-defined model.

Exceptions#

`ModelError`	Exception raised when the model is not configured correctly.
`OneDimensionalModelError`	Exception raised when the model is one-dimensional.

Classes#

Model

Base class for the user-defined model being sampled.

Module Contents#

exception nessai.model.ModelError#

Bases: Exception

Exception raised when the model is not configured correctly.

Added in version 0.15.0.

exception nessai.model.OneDimensionalModelError#

Bases: ModelError

Exception raised when the model is one-dimensional.

Changed in version 0.15.0: Changed to inherit from ModelError.

class nessai.model.Model#

Bases: abc.ABC

Base class for the user-defined model being sampled.

The user must define the attributes names bounds and the methods log_likelihood and log_prior.

The user can also define the reparemeterisations here instead of in the keyword arguments passed to the sampler.

reparameterisations = None#

dict: Dictionary of reparameterisations that overrides the values specified.

likelihood_evaluations = 0#

int: Number of likelihood evaluations.

likelihood_evaluation_time#

datetime.timedelta(): Time spent evaluating the likelihood.

pool = None#

obj: Multiprocessing pool for evaluating the log-likelihood.

allow_vectorised_prior = True#

bool: Allow the model to use a vectorised prior. If True, nessai will try to check if the log-prior is vectorised and use call the method as a vectorised function. If False, nessai won’t check and, even if the log-prior is vectorised, it will only evaluate the log-prior one sample at a time.

allow_vectorised = True#

bool: Allow the model to use a vectorised likelihood. If True, nessai will try to check if the model is vectorised and use call the likelihood as a vectorised function. If False, nessai won’t check and, even if the likelihood is vectorised, it will only evaluate the likelihood one sample at a time.

likelihood_chunksize = None#

int: Chunksize to use with a vectorised likelihood. If specified the likelihood will be called with at most chunksize points at once.

allow_multi_valued_likelihood = False#

bool: Allow for a multi-valued likelihood function that will return different likelihood values for the same point in parameter space. This is only recommended when the variation is significantly smaller that the variations in the likelihood across the prior.

parallelise_prior = False#

bool: Parallelise calculating the log-prior using the multiprocessing pool.

set_rng(rng: numpy.random.Generator | None = None)#

Set the random number generator.

Parameters:

rngnp.random.Generator, optional: Random number generator. If not provided, a new generator is created.

property names#: List of the names of each parameter in the model.

property bounds#: Dictionary with the lower and upper bounds for each parameter.

property dims#: Number of dimensions in the model

property discrete_parameters#

List of discrete parameters.

None if there are no discrete parameters.

property has_discrete_parameters#: Indicates if the model contains discrete parameters.

property lower_bounds#: Lower bounds on the priors

property upper_bounds#: Upper bounds on the priors

property vectorised_likelihood#

Boolean to indicate if the likelihood is vectorised or not.

Checks that the values returned by computing the likelihood for individual samples matches those return by evaluating the likelihood in a batch. If a TypeError or ValueError are raised the likelihood is assumed to be vectorised.

This check can be prevented by setting nessai.model.Model.allowed_vectorised to False.

configure_pool(pool=None, n_pool=None)#

Configure a multiprocessing pool for the likelihood computation.

Configuration will be skipped if the pool has already been configured.

Parameters:

pool: User provided pool. Must call nessai.utils.multiprocessing.initialise_pool_variables() before creating the pool or pass said function to the initialiser with the model.
n_poolint: Number of threads to use to create an instance of multiprocessing.Pool.

close_pool(code=None)#

Close the the multiprocessing pool.

Also resets the pool configuration.

new_point(N=1)#

Create a new LivePoint, drawn from within bounds.

See new_point_log_prob if changing this method.

Parameters:

Nint, optional: Number of points to draw. By default draws one point. If N > 1 points are drawn using a faster method.

Returns:

ndarray: Numpy structured array with fields for each parameter and log-prior (logP) and log-likelihood (logL)

new_point_log_prob(x)#

Computes the proposal probability for a new point.

This does not assume the that points will be drawn according to the prior. If new_point is redefined this method must be updated to match.

Parameters:

xndarray: Points in a structured array

Returns:

ndarray: Log proposal probability for each point

in_bounds(x)#

Check if a set of live point are within the prior bounds.

Parameters:

xnumpy.ndarray: Structured array of live points. Must contain all of the parameters in the model.

Returns:

Array of bool: Array with the same length as x where True indicates the point is within the prior bounds.

in_unit_hypercube(x: numpy.ndarray) → numpy.ndarray#

Check if samples are within the unit hypercube

Parameters:

xnumpy.ndarray: Structured array of live points. Must contain all of the parameters in the model.

Returns:

Array of bool: Array with the same length as x where True indicates the point is within the prior bounds.

abstract sample_parameter(name, n=1)#

Draw samples for a specific parameter from the prior.

Should be implemented by the user and return a numpy array of length n. The array should NOT be a structured array. This method is not required for standard sampling with nessai. It is intended for use with nessai.conditional.ConditionalFlowProposal.

Parameters:

namestr: Name for the parameter to sample
nint, optional: Number of samples to draw.

parameter_in_bounds(x, name)#

Check if an array of values for specific parameter are in the prior bounds.

Parameters:

xnumpy:ndarray: Array of values. Not a structured array.

Returns:

Array of bool: Array with the same length as x where True indicates the value is within the prior bounds.

sample_unit_hypercube(n: int = 1) → numpy.ndarray#

“Sample from the unit hypercube.

Parameters:

nint: Number of samples

Returns:

numpy.ndarray: Structured array of samples

log_prior_unit_hypercube(x) → numpy.ndarray#

Compute the log-prior in the unit hypercube.

By default this returns log(1) and checks if the values are in the unit hypercube.

abstract from_unit_hypercube(x)#

Map from the unit hypercube to the priors.

Not implemented by default.

abstract to_unit_hypercube(x)#

Map from the prior space to the unit hypercube.

Not implemented by default.

abstract log_prior(x)#: Returns log-prior, must be defined by the user.

abstract log_likelihood(x)#: Returns the log-likelihood, must be defined by the user.

evaluate_log_likelihood(x)#

Evaluate the log-likelihood and track the number of calls.

Returns:

float: Log-likelihood value

batch_evaluate_log_likelihood(x: numpy.ndarray, unit_hypercube: bool = False) → numpy.ndarray#

Evaluate the likelihood for a batch of samples.

Uses the pool if available.

Parameters:

xnumpy.ndarray: Array of samples
unit_hypercubebool: Indicates if input samples are from the unit hypercube or not.

Returns:

numpy.ndarray: Array of log-likelihood values

batch_evaluate_log_prior(x: numpy.ndarray, unit_hypercube: bool = False) → numpy.ndarray#

Evaluate the log-prior for a batch of samples.

Uses the pool if available.

Parameters:

xnumpy.ndarray: Array of samples
unit_hypercubebool: Indicates if input samples are from the unit hypercube or not.

Returns:

numpy.ndarray: Array of log-prior values

batch_evaluate_log_prior_unit_hypercube(x: numpy.ndarray) → numpy.ndarray#

Evaluate the log-prior in the unit hypercube for a batch of samples.

Uses the pool if available.

Parameters:

xnumpy.ndarray: Array of samples

Returns:

numpy.ndarray: Array of log-prior values

unstructured_view(x)#

An unstructured view of point(s) x that only contains the parameters in the model.

This is quicker than converting to a unstructured array and does not create a copy of the array.

Calls nessai.livepoint.unstructured_view() with a pre-computed dtype.

Warning

Will only work if all of the model parameters use the same dtype.

Parameters:

xstructured_array: Structured array of points

Returns:

numpy.ndarray: View of x as an unstructured array that contains only the parameters in the model. Shape is (x.size, self.dims).

classmethod check_new_point_methods()#

Check if the new_point methods have been redefined.

Added in version 0.15.0.

Raises:

ModelError: If the new_point methods have been redefined but not both.

verify_model()#

Verify that the model is correctly setup. This includes checking the names, bounds and log-likelihood.

Returns:

bool: True if the model was verified as valid.