Tutorial on writing custom training loops

[michaeldeistler]

Ideally, before doing this, we should swap out the flows such that the likelihood_estimator has a nicer API.

from sbi.inference import SNLE
from sbi.utils import likelihood_nn
from sbi.inference import likelihood_estimator_based_potential, MCMCPosterior


prior = BoxUniform(-torch.ones((2,)), torch.ones((2,)))
theta = prior.sample((1000,))
x = theta + torch.randn((1000, num_dim))
x_o = torch.randn((1, num_dim))

likelihood_estimator = likelihood_nn("maf")
# ... train however you want ...

potential_fn, parameter_transform = likelihood_estimator_based_potential(
    likelihood_estimator, prior, x_o
)
posterior = MCMCPosterior(
    potential_fn, proposal=prior, theta_transform=parameter_transform
)

[michaeldeistler]

Duplicate of #766