The $lp() method extracts lp__, the total log probability
(target) accumulated in the model block of the Stan program. For
variational inference the log density of the variational approximation to
the posterior is available via the $lp_approx() method. For
Laplace approximation the unnormalized density of the approximation to
the posterior is available via the $lp_approx() method.
See the Log Probability Increment vs. Sampling Statement section of the Stan Reference Manual for details on when normalizing constants are dropped from log probability calculations.
lp()
lp_approx()
lp_approx()Value
A numeric vector with length equal to the number of (post-warmup)
draws or length equal to 1 for optimization.
Details
lp__ is the unnormalized log density on Stan's unconstrained space.
This will in general be different than the unnormalized model log density
evaluated at a posterior draw (which is on the constrained space). lp__ is
intended to diagnose sampling efficiency and evaluate approximations.
For variational inference lp_approx__ is the log density of the variational
approximation to lp__ (also on the unconstrained space). It is exposed in
the variational method for performing the checks described in Yao et al.
(2018) and implemented in the loo package.
For Laplace approximation lp_approx__ is the unnormalized density of the
Laplace approximation. It can be used to perform the same checks as in the
case of the variational method described in Yao et al. (2018).
References
Yao, Y., Vehtari, A., Simpson, D., and Gelman, A. (2018). Yes, but did it work?: Evaluating variational inference. Proceedings of the 35th International Conference on Machine Learning, PMLR 80:5581–5590.
See also
Examples
# \dontrun{
fit_mcmc <- cmdstanr_example("logistic")
head(fit_mcmc$lp())
#> [1] -64.8724 -65.7859 -65.2974 -66.5566 -66.7795 -67.0780
fit_mle <- cmdstanr_example("logistic", method = "optimize")
fit_mle$lp()
#> [1] -63.9218
fit_vb <- cmdstanr_example("logistic", method = "variational")
plot(fit_vb$lp(), fit_vb$lp_approx())
# }