Output gradient?

[jean997]

For parametric families, is it possible to access the gradient of the log likelihood at the maximum?

[pcarbo]

@jean997 There might be—I can check. Is there a particular parametric family you are interested in?

[jean997]

Point normal if I can only pick one!

[pcarbo]

@jean997 See the add-grad-output-eg branch for an example of how to do this:

set.seed(1)
theta <- c(rep(0, 100), rexp(100))
s <- 1
x <- theta + rnorm(200, 0, s)
fit <- ebnm(x, s, prior_family = "point_normal")
print(fit$nllik_grad)
# [1] -1.115e-07 -1.136e-08

Note that I believe this is the gradient of the negative log-likelihood (based on looking at Jason's code).

[jean997]

neat thank you!