Training interface

[michaeldeistler]

In order to give more flexibility, we could expose the entire training loop to the user:

from sbi.inference import nre_loss, ratio_estimator_based_potential, MCMCPosterior

net = classifier_nn("mdn")

data_loader = ...
for e in range(epochs):
    optim.zero_grad()
    loss = nre_loss(net, theta, x)
    loss.backward()
    optim.step()

potential = ratio_estimator_based_potential(net, prior, x_o)
posterior = MCMCPosterior(potential, proposal=prior)

This would require two things:

• a good way to deal with z-scoring
• separating the loss functions from the rest

[janfb]

related to #932