Discrete Prior Selection

[Saheli2001]

Can we use discrete priors as proposal density? I am having an error when i use bernoulli distribution as proposal priors. Any idea regarding this would be helpful.

P.S. I am new to the theory simulation based inference and this sbi package.

[michaeldeistler]

Hi there,

generally, sbi does not yet support discrete density estimators (and thereby also no discrete priors).

As a workaround, you can simply use a continuous density estimator and get the best-possible continuous approximation to the discrete posterior. When doing this, do not pass the prior at initialization (see below). Also, I would recommend to use an NSF as density estimator for this:

inference = SNPE(density_estimator="nsf") # Do not pass prior
_ = inference.append_simulations(theta, x).train()
posterior = inference.build_posterior()