Adding new notebook for conjugate sampling #4199
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
|
I can't (yet) pretend to be able to comment on whether it's a good idea to "combine NUTS with Gibbs sampling from a closed form posterior for a hierarchical Dirichlet multinomial model" (maybe attending PyMCon will help 😄 ), but I'll just leave a quick comment regarding the CI failure (in case it's not clear): you can fix it with |
Ah darn it - I didn't use the pre-commit process right. Thanks for the command. |
Codecov Report
@@ Coverage Diff @@
## master #4199 +/- ##
==========================================
+ Coverage 88.93% 88.95% +0.01%
==========================================
Files 92 92
Lines 14806 14806
==========================================
+ Hits 13168 13170 +2
+ Misses 1638 1636 -2
|
|
Might be interesting to mention if anything changes with the introduction of the Dirichlet-Multinomial distribution (PR: #3639). Does having an explicit compound distribution change performance to a similar extent as this strategy? |
These two topics are closely related but this notebook shows how to sample from the Dirichlet random vector used to parameterize the multinomial distribution while the compound distribution marginalizes over that random vector instead. It is probably worth updating this notebook to mention that using the DCM is likely much faster if you are not interested in the low-level Dirichlet draws. |
There was a problem hiding this comment.
This looks really good @ckrapu! Thanks for the initiative, and sorry for the long time to review, the last few weeks have been hectic...
I like the way you introduce, contextualize and explain the problem 👌 It's almost ready to go; I just left some comments and suggestions below -- feel free to ask if anything's unclear 😉
Note you'll have to merge master into this branch to fix the conflict.
Regarding the Dirichlet-Multinomial PR, I agree with you:
It is probably worth updating this notebook to mention that using the DCM is likely much faster if you are not interested in the low-level Dirichlet draws.
BTW this PR has unfortunately stalled, so if you or @ricardoV94 feel like pushing it over the finish line that'd really be a great addition 🤩
| @@ -0,0 +1,634 @@ | |||
| { | |||
There was a problem hiding this comment.
There was a problem hiding this comment.
You forgot this typo ;)
There was a problem hiding this comment.
Thanks @ckrapu, that looks really good now! Just one last typo you forgot and one last question, and it'll be good to go 😉
Thanks for the review! I really appreciate the attention to detail. |
There was a problem hiding this comment.
All good, thanks for this nice notebook @ckrapu ! I'll merge once tests pass 🎉
PS: I relaunched some of the jobs, which weirdly failed with a very cryptic message 🤷♂️
There was an error in the .travis.yml file from which we could not recover.
Unfortunately, we do not know much about this error.
Please review https://docs.travis-ci.com, or contact us at support@travis-ci.com with the error ID: 38be43ecc95a469f8a3b50222008d246
|
Should add this to release notes too, it's a good one. |
|
Ah right! Will merge once you've added that then @ckrapu 😉 |
|
Thanks @ckrapu, this is a great contribution. |
This is a short piece of documentation showing how to combine NUTS with Gibbs sampling from a closed form posterior for a hierarchical Dirichlet multinomial model.
I acknowledge this may be a somewhat controversial item to include in the documentation because of the frequency with which users inquire about replacing NUTS with something that draws samples more rapidly despite being less efficient in ESS / second. In the particular case that is discussed in this notebook, it happens to be more efficient to do alternating NUTS / conjugate sampling updates in a compound Gibbs sampling scheme.
Any thoughts?