Practical recipes

Checking the prior when evaluating the likelihood

By default, Aspire samplers always evaluate the log-prior before the log-likelihood. This allows users to check the prior support and skip likelihood evaluations for samples that lie outside the prior bounds.

import aspire
import numpy as np


def log_likelihood(samples: aspire.Samples) -> np.ndarray:
    if samples.log_prior is None:
        raise RuntimeError("log-prior has not been evaluated!")
    # Return -inf for samples with invalid prior
    logl = np.full(samples.n_samples, -np.inf, dtype=float)
    # Only evaluate the likelihood where the prior is finite
    mask = np.isfinite(samples.log_prior, dtype=bool)
    # Valid samples
    x = samples.x[mask, :]
    logl[mask] = -np.sum(x**2, axis=1)  # Example likelihood
    return logl

Checking the flow distribution

It can be useful to inspect the flow-based proposal distribution before sampling from the posterior. You can do this by drawing samples from the flow after fitting and comparing them to the initial samples:

from aspire import Aspire, Samples
from aspire.plot import plot_comparison

# Define the initial samples
initial_samples = Samples(...)

# Define the Aspire instance
aspire = Aspire(
    log_likelihood=log_likelihood,
    log_prior=log_prior,
    ...
)

# Fit the flow to the initial samples
fit_history = aspire.fit(initial_samples)

# Draw samples from the flow
flow_samples = aspire.sample_flow(10_000)

# Plot a comparison between initial samples and flow samples
fig = plot_comparison(
    initial_samples,
    flow_samples,
    per_samples_kwargs=[
        dict(include_weights=False, color="C0"),
        dict(include_weights=False, color="C1"),
    ],
    labels=["Initial samples", "Flow samples"],
)
# Save or show the figure
fig.savefig("flow_comparison.png")