Running the sampler

Defining the model

The user must define a model that inherits from nessai.model.Model that defines two parameters and two methods. This object contains the Bayesian prior and likelihood which will be used for sampling.

Parameters:

names: a list of str with names for the parameters to be sampled
bounds: a dict with a tuple for each parameter in names with defines the lower and upper bounds of the priors.

Methods:

The user MUST define these two methods, the input to both is a structured numpy array with fields defined by names.

log_prior: return the log-prior probability of a live point (and enforce the bounds)
log_likelihood: return the log-likelihood probability of a live point (must be finite)

The input to both methods is a live point x which is an instance of a structured numpy array with one field for each parameters in names and two additional fields logP and logL. Each parameter can be accessed using the name of each field like you would a dictionary.

For examples of using live points see: using live points

Example model

Here’s an example of a simple model taken from one of the examples:

class GaussianModel(Model):
    """A simple two-dimensional Gaussian likelihood."""

    def __init__(self):
        # Names of parameters to sample
        self.names = ["x", "y"]
        # Prior bounds for each parameter
        self.bounds = {"x": [-10, 10], "y": [-10, 10]}

    def log_prior(self, x):
        """
        Returns log of prior given a live point assuming uniform
        priors on each parameter.
        """
        # Check if values are in bounds, returns True/False
        # Then take the log to get 0/-inf and make sure the dtype is float
        log_p = np.log(self.in_bounds(x), dtype="float")
        # Iterate through each parameter (x and y)
        # since the live points are a structured array we can
        # get each value using just the name
        for n in self.names:
            log_p -= np.log(self.bounds[n][1] - self.bounds[n][0])
        return log_p

    def log_likelihood(self, x):
        """
        Returns log likelihood of given live point assuming a Gaussian
        likelihood.
        """
        log_l = np.zeros(x.size)
        # Use a Gaussian logpdf and iterate through the parameters
        for n in self.names:
            log_l += norm.logpdf(x[n])
        return log_l

Initialising and running the sampler

Once a modelled is defined, create an instance of nessai.flowsampler.FlowSampler. This is when the sampler and the proposal methods are configured, for example setting the number of live points (nlive) or setting the class of normalising flow to use. nessai includes a large variety of settings that control different aspects of the sampler, these can be essential to efficient sampling. See sampler configuration for an in-depth explanation of all the settings.

from nessai.flowsampler import FlowSampler

# Initialise sampler with the model
sampler = FlowSampler(GaussianModel(), output='./', nlive=1000)
# Run the sampler
sampler.run()

Sampler output

Once the sampler has converged the results and other automatically generated plots with be saved in the directory specified as output. By default this will include:

result.json: a json file which contains various fields, the most relevant of which are posterior_samples, log_evidence and information.
posterior_distribution.png: a corner-style plot of the posterior distribution.
trace.png: a trace plot which shows the nested samples for each sampled parameter for the entire sampling process.
state.png: the state plot which shows various statistics tracked by the sampler.
insertion_indices.png: the distribution of insertion indices for all of the nested samples.
logXlogL.png: the evolution of the maximum log-likelihood versus the log prior volume.
two resume files (.pkl) used for resuming the sampler.
config.json: the exact configuration used for the sampler.

For a more detail explanation of outputs and examples, see here

Complete examples

For complete examples see Simple Gaussian example and the example directory.