ignoring mcmc sampling

[Ziaeemehr]

Suppose common steps in the sbi workflow

• providing theta and data features
• train nn and make posterior
• sample from posterior Is there any way to avoid sampling going to mcmc which is usually time consuming and for example return nan as sample?

inference = SNPE(
                prior=prior, density_estimator=density_estimator, device=device)
inference = inference.append_simulations(theta, x)
estimator_ = inference.train()
posterior = inference.build_posterior(estimator_, sample_with="mcmc") # probably need to make a change here?
posterior.sample((num_samples,), x=xo)

This is needed specially when someone is putting the workflow in a loop and try on multiple dataset.

Cheers A

[michaeldeistler]

I don't understand the problem, why do you pass sample_with=mcmc if you do not want to sample with mcmc?