API of implemented methods¶

This notebook spells out the API for all algorithms implemented in the sbi toolbox:

Posterior estimation (SNPE)
Likelihood estimation (SNLE)
Likelihood-ratio estimation (SNRE)
Utilities

Posterior estimation (SNPE)¶

Fast ε-free Inference of Simulation Models with Bayesian Conditional Density Estimation
by Papamakarios & Murray (NeurIPS 2016)
[PDF] [BibTeX]

from sbi.inference import SNPE_A

inference = SNPE_A(prior)
proposal = prior
for _ in range(rounds):
    theta = proposal.sample((num_sims,))
    x = simulator(theta)
    _ = inference.append_simulations(theta, x, proposal=proposal).train()
    posterior = inference.build_posterior().set_default_x(x_o)
    proposal = posterior

Automatic posterior transformation for likelihood-free inference
by Greenberg, Nonnenmacher & Macke (ICML 2019)
[PDF]

from sbi.inference import SNPE

inference = SNPE(prior)
proposal = prior
for _ in range(rounds):
    theta = proposal.sample((num_sims,))
    x = simulator(theta)
    _ = inference.append_simulations(theta, x, proposal=proposal).train()
    posterior = inference.build_posterior().set_default_x(x_o)
    proposal = posterior

Truncated proposals for scalable and hassle-free simulation-based inference
by Deistler, Goncalves & Macke (NeurIPS 2022)
[Paper]

from sbi.inference import SNPE
from sbi.utils import get_density_thresholder, RestrictedPrior

inference = SNPE(prior)
proposal = prior
for _ in range(rounds):
    theta = proposal.sample((num_sims,))
    x = simulator(theta)
    _ = inference.append_simulations(theta, x).train(force_first_round_loss=True)
    posterior = inference.build_posterior().set_default_x(x_o)

    accept_reject_fn = get_density_thresholder(posterior, quantile=1e-4)
    proposal = RestrictedPrior(prior, accept_reject_fn, sample_with="rejection")

Likelihood estimation (SNLE)¶

Sequential neural likelihood: Fast likelihood-free inference with autoregressive flows
by Papamakarios, Sterratt & Murray (AISTATS 2019)
[PDF] [BibTeX]

from sbi.inference import SNLE

inference = SNLE(prior)
proposal = prior
for _ in range(rounds):
    theta = proposal.sample((num_sims,))
    x = simulator(theta)
    _ = inference.append_simulations(theta, x).train()
    posterior = inference.build_posterior().set_default_x(x_o)
    proposal = posterior

Variational methods for simulation-based inference
by Glöckler, Deistler, Macke (ICLR 2022)
[Paper]

from sbi.inference import SNLE

inference = SNLE(prior)
proposal = prior
for _ in range(rounds):
    theta = proposal.sample((num_sims,))
    x = simulator(theta)
    _ = inference.append_simulations(theta, x).train()
    posterior = inference.build_posterior(sample_with="vi", vi_method="fKL").set_default_x(x_o)
    proposal = posterior

Flexible and efficient simulation-based inference for models of decision-making
by Boelts, Lueckmann, Gao, Macke (Elife 2022)
[Paper]

from sbi.inference import MNLE

inference = MNLE(prior)
theta = prior.sample((num_sims,))
x = simulator(theta)
_ = inference.append_simulations(theta, x).train()
posterior = inference.build_posterior().set_default_x(x_o)

Likelihood-ratio estimation (SNRE)¶

Likelihood-free MCMC with Amortized Approximate Likelihood Ratios
by Hermans, Begy & Louppe (ICML 2020)
[PDF]

from sbi.inference import SNRE_A

inference = SNRE_A(prior)
theta = prior.sample((num_sims,))
x = simulator(theta)
_ = inference.append_simulations(theta, x).train()
posterior = inference.build_posterior().set_default_x(x_o)

On Contrastive Learning for Likelihood-free Inference
Durkan, Murray & Papamakarios (ICML 2020)
[PDF].

from sbi.inference import SNRE

inference = SNRE(prior)
proposal = prior
for _ in range(rounds):
    theta = proposal.sample((num_sims,))
    x = simulator(theta)
    _ = inference.append_simulations(theta, x).train()
    posterior = inference.build_posterior().set_default_x(x_o)
    proposal = posterior

Towards Reliable Simulation-Based Inference with Balanced Neural Ratio Estimation
by Delaunoy, Hermans, Rozet, Wehenkel & Louppe (NeurIPS 2022)
[PDF]

from sbi.inference import BNRE

inference = BNRE(prior)
theta = prior.sample((num_sims,))
x = simulator(theta)
_ = inference.append_simulations(theta, x).train(regularization_strength=100.)
posterior = inference.build_posterior().set_default_x(x_o)

Contrastive Neural Ratio Estimation
Benjamin Kurt Miller, Christoph Weniger, Patrick Forré (NeurIPS 2022)
[PDF]

# The main feature of NRE-C is producing an exact ratio of densities at optimum, even when using multiple contrastive pairs (classes).

from sbi.inference import SNRE_C

# Amortized inference
inference = SNRE_C(prior)
proposal = prior
theta = proposal.sample((num_sims,))
x = simulator(theta)
_ = inference.append_simulations(theta, x).train(
    num_classes=5,  # SNRE_C sees `2 * num_classes - 1` marginally drawn contrastive pairs.
    gamma=1.0,  # SNRE_C can control the weight between terms in its loss function.
)
posterior = inference.build_posterior().set_default_x(x_o)

Utilities¶

Simulation-based calibration
by Talts, Betancourt, Simpson, Vehtari, Gelman (arxiv 2018)
[Paper])

from sbi.analysis import run_sbc, sbc_rank_plot

thetas = prior.sample((1_000,))
xs = simulator(thetas)

ranks, dap_samples = run_sbc(
    thetas, xs, posterior, num_posterior_samples=1_000
)

_ = sbc_rank_plot(
    ranks=ranks,
    num_posterior_samples=num_posterior_samples,
    plot_type="hist",
    num_bins=None,
)

Restriction estimator
by Deistler, Macke & Goncalves (PNAS 2022)
[Paper]

from sbi.inference import SNPE
from sbi.utils import RestrictionEstimator

restriction_estimator = RestrictionEstimator(prior=prior)
proposal = prior

for _ in range(num_rounds):
    theta = proposal.sample((num_sims,))
    x = simulator(theta)
    restriction_estimator.append_simulations(theta, x)
    classifier = restriction_estimator.train()
    proposal = restriction_estimator.restrict_prior()

all_theta, all_x, _ = restriction_estimator.get_simulations()

inference = SNPE(prior)
density_estimator = inference.append_simulations(all_theta, all_x).train()
posterior = inference.build_posterior()

Expected coverage (sample-based)
as computed in Deistler, Goncalves, Macke (Neurips 2022) [Paper] and in Rozet, Louppe (2021) [Paper]

from sbi.analysis import run_sbc, sbc_rank_plot

thetas = prior.sample((1_000,))
xs = simulator(thetas)

ranks, dap_samples = run_sbc(
    thetas, xs, posterior, num_posterior_samples=1_000, reduce_fns=posterior.log_prob
)

_ = sbc_rank_plot(
    ranks=ranks,
    num_posterior_samples=num_posterior_samples,
    plot_type="hist",
    num_bins=None,
)