Add tutorial for training interface

[michaeldeistler]

Below is the training interface for NLE:

from sbi.neural_nets.flow import build_nsf
from sbi.inference.posteriors import MCMCPosterior
from sbi.inference.potentials import likelihood_estimator_based_potential


# Build neural density estimator.
density_estimator = build_nsf(x, theta)

# Training loop.
opt = Adam(list(density_estimator.parameters()), lr=5e-4)
for _ in range(100):
    opt.zero_grad()
    losses = density_estimator.loss(x, condition=theta)
    loss = torch.mean(losses)
    loss.backward()
    opt.step()

# Build posterior and sample with MCMC.
potential, tf = likelihood_estimator_based_potential(density_estimator, prior, x_o)
posterior = MCMCPosterior(
    potential,
    proposal=prior,
    theta_transform=tf,
    num_chains=100,
    thin=1,
    method="slice_np_vectorized"
)
samples = posterior.sample((1000,), x=x_o)
_ = pairplot(samples, limits=[[-3, 3], [-3, 3]], figsize=(3, 3))

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 76.87%. Comparing base (b3254ed) to head (827df2a).
Report is 11 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #983      +/-   ##
==========================================
- Coverage   85.44%   76.87%   -8.57%     
==========================================
  Files         101      101              
  Lines        7941     7945       +4     
==========================================
- Hits         6785     6108     -677     
- Misses       1156     1837     +681     

Flag	Coverage Δ
unittests	76.87% <ø> (-8.57%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

see 25 files with indirect coverage changes

[janfb]

I pushed small fixes directly. Thanks for adding this, looks great!