-
-
Notifications
You must be signed in to change notification settings - Fork 2.1k
Closed
Description
While reading through the stan code to double check the new implementation of multinomial sampling, I noticed that the termination criterion in stan is different to the one in pymc3. There have been recent changes in stan (stan-dev/stan#1854, stan-dev/stan#2180), but even in the older code ("classic nuts") it is different. pymc3 fails to take into account the metric and always assumes a unit mass matrix. This is probably because the original paper on nuts used the standard euclidean space.
Activity
twiecki commentedon Feb 14, 2017
Yes, we haven't tracked their changes consistently. Should definitely bring our NUTS as much in sync with theirs as possible.
aseyboldt commentedon Mar 12, 2017
This was fixed in #1769