Hi everyone, I have a strangely specific question. Is there a way to evaluate pointwise the joint probability density function (lets assume all my latents are continuous) and particularly its gradient on the constrained (i.e., not reparametrized) space? I know most of the Pyro internals for MCMC and SVI reparametrize the underlying space first so everything happens in R^d, but for a current research project I would like to use Pyro for, I need to work directly on the constrained space. Any tips would be appreciated!
@Robsal In Pyro, the joint probability density is in constrained space. You can get it with
pyro.poutine.trace(model).get_trace(data).log_prob_sum()
In mcmc and autoguides, we generate samples in unconstrained space, so we will adjust some log_abs_det_jacobian term to correct it. You don’t have to worry about it here I guess.