I was wondering why NumPyro distributions only provided a
log_prob method, which implies computing the normalized distribution, and not also an unnormalized log-density method which doesn’t have the overhead of computing the normalization constant?
Using the unnormalized density could speed up computations by avoiding computing any normalization constants.
Is there any reason why using unnormalized densities would not work in NumPyro? Or are they already in use and have I missed something?