Get log probabilities of samples using AutoBNAFNormal

Is it possible to get log_prob estimates of samples from a trained autoregressive flow, e.g. AutoBNAFNormal? I understand that in the literature the inverse Jacobian of the BNAF was not available. From 1904.0467:

…note that, while B-NAFs and NAFs are invertible in principle, their inverses are not available in closed form.

Further, when attempting to do this in numpyro, I get the error: NotImplementedError: Block neural autoregressive transform does not have an analytic inverse implemented. One of the benefits of (general) NFs are tractable densities – I am wondering since it seems this is not true for autoregressive flows, if it is otherwise possible to get exact densities of samples from the posterior (without the inverse Jacobian), or if people have strong preferences towards other methods (KDE, MCMC, etc…). Thanks!

it is pretty easy to compute the log density for generated samples but harder to compute the log density for arbitrary points since the latter requires an (approximate) inverse

What do you mean by ‘generated samples’? I’m referencing points from the flow retrieved via sample_posterior. Are there any code examples that show this?

i believe if you use the _get_posterior method you can directly sample and compute log_prob