Closed scheidan closed 2 years ago
Thank you for your interest and comment.
Our goal here was to provide a utility function intended for training normalizing flows, which involves computing the log-likelihood of the predicted latent variables. In normalizing flows, we typically assume that latent variables are normally distributed. Using JuliaStats' Distributions.jl implementation, you can incorporate the log-likelihoods of other distributions.
By looking at the code, the docstring reads,
Log-likelihood of X for a Gaussian distribution with given mean μ and variance σ. All elements of X are assumed to be iid.
We will make this also explicit in ∇log_likelihood
and Hlog_likelihood
s docstrings.
I've just discoverer this package and it look very interesting! Thanks a lot for putting this together.
Just a small point I stumbled on: The function
log_likelihood
implicitly means the log pdf of a multivariate Gaussian distribution (up to a constant). As statistician I can imagine many non-Gaussian likelihood functions so I find this naming that rather confusing. The docstrings of∇log_likelihood
andHlog_likelihood
do not contain a reference to the Gaussian distribution at all. Usinglog_likelihood_normal
or so would clarify thinks a lot.