GPry via Cobaya

GPry is a package for Bayesian inference of expensive likelihoods using Gaussian Process (GP) emulation. Instead of evaluating your likelihood thousands of times, GPry builds a surrogate GP model using active learning—intelligently choosing evaluation points to maximise information gain—and can reduce the number of likelihood evaluations by a factor of 100 or more. It is best suited for smooth, low-dimensional (≲20 parameters) likelihoods where each evaluation is computationally expensive.

This interface bridges Discovery models to GPry through Cobaya. It builds a Cobaya model from your Discovery likelihood and priors, then runs GPry (or another Cobaya sampler) on it.

Minimal Run

from discoverysamplers.gpry_interface import DiscoveryGPryCobayaBridge

def my_model(params):
    x, y = params['x'], params['y']
    return -0.5 * (x**2 + y**2)

priors = {'x': ('uniform', -5.0, 5.0), 'y': ('uniform', -5.0, 5.0)}

# Create the bridge (accepts callable or object with .logL attribute)
bridge = DiscoveryGPryCobayaBridge(
    discovery_model=my_model,
    priors=priors,
    latex_labels={'x': r'$x$', 'y': r'$y$'},
    like_name='my_like',
)

info, sampler = bridge.run_sampler(max_samples=5000)  # defaults to GPry
products = sampler.products()
samples = products["sample"]
print(samples.mean('x'), samples.std('x'))

Using with Discovery Likelihoods

import discovery as ds

# Create Discovery likelihood
psr = ds.Pulsar.read_feather('path/to/pulsar.feather')
likelihood = ds.PulsarLikelihood([...])

# Pass the likelihood object directly
bridge = DiscoveryGPryCobayaBridge(
    discovery_model=likelihood,  # or likelihood.logL - both work
    priors=priors,
    like_name='pulsar_likelihood'
)

Priors

Uniform: ("uniform", min, max) → Cobaya prior: {min, max}
Log-uniform: ("loguniform", a, b) → Cobaya prior: {dist: loguniform, a, b}
Normal: ("normal", mean, sigma[, min, max]) → Cobaya prior: {dist: norm, loc, scale}
Fixed: ("fixed", value) → fixed parameter
Callable priors are not supported here.

Key Options

sampler: choose Cobaya sampler ('gpry' default, or 'mcmc', 'polychord' if installed).
max_samples: total samples for GPry.
alternative_paramnames: remap model names to Cobaya names.
latex_labels: used for GetDist outputs.

Manual Access

info = bridge.info              # Cobaya info dict
model = bridge.model            # Cobaya model
loglike = model.loglike({'x': 0.0, 'y': 0.0})

Tips

Keep priors bounded; GPry and Cobaya expect finite ranges for uniform/loguniform.
Use GetDist (already in Cobaya) for plots and summaries: products = sampler.products(); samples = products["sample"].
When using alternative names, ensure you still provide priors keyed by the original Discovery names.