nessai.samplers.importancesampler

nessai.samplers.importancesampler#

Importance nested sampler.

Classes#

`OrderedSamples`	Samples ordered by log-likelihood.
`ImportanceNestedSampler`

Module Contents#

class nessai.samplers.importancesampler.OrderedSamples(strict_threshold: bool = False, replace_all: bool = False, save_log_q: bool = False)#

Samples ordered by log-likelihood.

Parameters:

strict_threshold: If true, when adding new samples, only those above the current log-likelihood threshold will be added to the live points.
replace_all: If true, all samples will be remove when calling remove_samples
save_log_q: If true, log_q will be saved when the instance is pickled. This makes resuming faster but increases the disk usage. If false, the values will not be saved and must be recomputed.

property live_points: numpy.ndarray#: The current set of live points

property nested_samples: numpy.ndarray#: The current set of discarded points

update_log_likelihood_threshold(threshold: float) → None#

Update the log-likelihood threshold.

Only relevant if using strict_threshold=True.

Parameters:

thresholdfloat: New log-likelihood threshold

sort_samples(samples: numpy.ndarray, *args) → numpy.ndarray#

Correctly sort new live points.

Parameters:

x: Array to sort
args: Any extra iterables to sort in the same way as x.

add_samples(samples: numpy.ndarray, log_q: numpy.ndarray) → None#: Add samples the existing samples

add_to_nested_samples(indices: numpy.ndarray) → None#: Move a set of samples from the live points to the nested samples

remove_samples() → int#

Remove samples all samples below the current log-likelihood threshold.

Returns:

nint: The number of samples removed

update_evidence() → None#: Update the evidence estimate given the current samples

finalise() → None#: Finalise the samples by consuming all of the live points

compute_importance(importance_ratio: float = 0.5)#

Compute the importance

Parameters:

importance_factor: Relative importance of the posterior versus the evidence. 1 is only the posterior and 0 is only the evidence.

Returns:

dict: Dictionary containing the total, posterior and evidence importance as a function of iteration.

compute_evidence_ratio(threshold: float | None = None) → float#

Compute the evidence ratio given the current log-likelihood threshold.

Parameters:

threshold: Log-likelihood threshold to use instead of the current value.

class nessai.samplers.importancesampler.ImportanceNestedSampler(model: nessai.model.Model, nlive: int = 5000, n_initial: int | None = None, output: str | None = None, seed: int | None = None, rng: numpy.random.Generator | None = None, checkpointing: bool = True, checkpoint_interval: int = 600, checkpoint_on_iteration: bool = False, checkpoint_callback: Callable | None = None, save_existing_checkpoint: bool = False, save_log_q: bool = False, logging_interval: int = None, log_on_iteration: bool = True, resume_file: str | None = None, plot: bool = True, plotting_frequency: int = 5, min_iteration: int | None = None, max_iteration: int | None = None, min_samples: int = 500, min_remove: int = 1, max_samples: int | None = None, stopping_criterion: str = 'ratio', tolerance: float = 0.0, n_update: int | None = None, plot_pool: bool = False, plot_level_cdf: bool = False, plot_trace: bool = True, plot_likelihood_levels: bool = True, plot_training_data: bool = False, plot_extra_state: bool = False, trace_plot_kwargs: dict | None = None, replace_all: bool = False, threshold_method: Literal['entropy', 'quantile'] = 'entropy', threshold_kwargs: dict | None = None, n_pool: int | None = None, pool: Any | None = None, check_criteria: Literal['any', 'all'] = 'any', weighted_kl: bool = False, draw_constant: bool = True, train_final_flow: bool = False, bootstrap: bool = False, close_pool: bool = False, strict_threshold: bool = False, draw_iid_live: bool = True, **kwargs: Any)#

Bases: nessai.samplers.base.BaseNestedSampler

Parameters:

model: User-defined model.
nlive: Number of live points.
tolerance: Tolerance for determining when to stop the sampler.
stopping_criterion: Choice of stopping criterion to use.
check_criteria: If using multiple stopping criteria determines whether any or all criteria must be met.
threshold_method: Method for determining new likelihood threshold.
threshold_kwargs: Keyword arguments for function that determines the likelihood threshold.
draw_constant: If specified the sampler will always add a constant number of samples from each proposal whilst removing a variable amount. If False, the the number will depend on the level method chosen. Note that this will override the choice of live points. The number of points draw is set by the live points.
min_samples: Minimum number of samples that are used for training the next normalising flow.
min_remove: Minimum number of samples that can be removed when creating the next level. If less than one, the sampler will stop if the level method determines no samples should be removed.
max_samples: The maximum number of samples at any given iteration. If not specified, then there is no limit to the number of samples.
plot_likelihood_levels: Enable or disable plotting the likelihood levels.
trace_plot_kwargs: Keyword arguments for the trace plot.
strict_thresholdbool: If true, when drawing new samples, only those with likelihoods above the current threshold will be added to the live points. If false, all new samples are added to the live points.
save_log_qbool: Boolean that determines if the log_q array is saved when checkpointing. If False, this can help reduce the disk usage.

property posterior_effective_sample_size: float#

The effective sample size of the posterior distribution.

Returns the value for the posterior samples from the resampling step if they are available, otherwise falls back to the samples from the initial sampling.

property samples_entropy: float#

Differential entropy of all of the samples (nested + live).

Notes

Compute the Monte Carlo approximation of

\[-\int W(x) \log W(x) dx\]

where \(W(x) = \pi(x)/Q(x)\).

property current_proposal_entropy: float#: Differential entropy of the current proposal

property live_points_unit: numpy.ndarray#: The current set of live points

property nested_samples_unit: numpy.ndarray#: The current set of discarded points

property reached_tolerance: bool#

Indicates if tolerance has been reached.

Checks if any or all of the criteria have been met, this depends on the value of check_criteria.

property stopping_criteria: List[str]#: The stopping criteria used by the sampler

static add_fields()#: Add extra fields logW, logQ, logU

configure_stopping_criterion(stopping_criterion: str | List[str], tolerance: float | List[float], check_criteria: Literal['any', 'all']) → None#: Configure the stopping criterion

get_proposal(subdir: str = 'levels', **kwargs)#: Configure the proposal.

update_output(output: str) → None#

Update the output directory and resume file.

Parameters:

output: str: Path to the output directory

configure_iterations(min_iteration: int | None = None, max_iteration: int | None = None) → None#

Configure the minimum and maximum iterations.

Note: will override any existing values when called.

check_configuration() → bool#

Check sampler configuration is valid.

Returns true if all checks pass.

populate_live_points() → None#

Draw the initial live points from the prior.

The live points are automatically sorted and assigned the iteration number -1.

initialise() → None#

Initialise the nested sampler.

Draws live points, initialises the proposal.

initialise_history() → None#: Initialise the dictionary to store history

update_history() → None#: Update the history dictionary

determine_threshold_quantile(samples: numpy.ndarray, q: float = 0.8, include_likelihood: bool = False) → int#

Determine where the next likelihood threshold should be located.

Computes the q’th quantile based on log-likelihood and log-weights.

Parameters:

samplesnp.ndarray: An array of samples
qfloat: Quantile to use. Defaults to 0.8
include_likelihoodbool: If True, the likelihood is included in the weights.

Returns:

int: The number of live points to discard.

determine_threshold_entropy(samples: numpy.ndarray, q: float = 0.5, include_likelihood: bool = False, use_log_weights: bool = True) → int#

Determine where the next likelihood threshold should be located using the entropy method.

Parameters:

q: Fraction by which to shrink the current level.
include_likelihood: Boolean to indicate whether the likelihood is included in the weights for each samples.
use_log_weights: Boolean to determine if the CDF is computed using the weights or log-weights.

plot_level_cdf(log_likelihood_values: numpy.ndarray, cdf: numpy.ndarray, threshold: float, q: float, filename: str | None = None) → matplotlib.figure.Figure | None#

Plot the CDF of the log-likelihood

Parameters:

log_likelihood_valuesnp.ndarray: The log-likelihood values for the CDF
cdfnp.ndarray: The CDF to plot
filenameOptional[str]: Filename for saving the figure. If not specified the figure will be returned instead.

Returns:

matplotlib.figure.Figure: Level CDF figure. Only returned when the filename is not specified.

determine_log_likelihood_threshold(samples: numpy.ndarray, method: Literal['entropy', 'quantile'] = 'entropy', **kwargs) → int#

Determine the next likelihood threshold

Parameters:

samplesnumpy.ndarray: An array of samples
methodLiteral[“entropy”, “quantile”]: The method to use
kwargs: Keyword arguments passed to the function for the chosen method.

Returns:

float: The log-likelihood threshold

add_new_proposal()#: Add a new proposal to the meta proposal

draw_n_samples(n: int, **kwargs)#

Draw n samples from the current proposal

Includes computing the log-likelihood of the samples

Parameters:

nint: Number of samples to draw
kwargs: Keyword arguments passed to the draw method of the proposal class.

compute_leakage(samples: numpy.ndarray, weights: bool = True) → float#

Compute the leakage for a number of samples.

Parameters:

samplesnumpy.ndarray: Array of samples.
weightsbool: If True, the weight of each sample is accounted for in the calculation.

Returns:

float: The leakage as a fraction of the total number of samples (or effective sample size if weights is True).

add_and_update_points(n: int)#

Add new points to the current set of live points.

Parameters:

nint: The number of points to add.

remove_samples() → int#: Remove samples from the current set of live points.

adjust_final_samples(n_batches=5)#: Adjust the final samples

finalise() → None#: Finalise the sampling process.

add_level_post_sampling(samples: numpy.ndarray, n: int) → None#: Add a level to the nested sampler after initial sampling has completed.

compute_stopping_criterion() → List[float]#

Compute the stopping criterion.

The method used will depend on how the sampler was configured.

checkpoint(periodic: bool = False, force: bool = False)#: Checkpoint the sampler.

log_state()#: Log the state of the sampler

update_proposal_weights()#

Update the proposal weights based on the current sample counts.

Uses the values from sample_counts.