Document initializing Turing #17
Conversation
Codecov Report
@@ Coverage Diff @@
## main #17 +/- ##
=======================================
Coverage 98.65% 98.65%
=======================================
Files 8 8
Lines 223 223
=======================================
Hits 220 220
Misses 3 3 Continue to review full report at Codecov.
|
docs/src/turing.md
Outdated
| @@ -0,0 +1,100 @@ | |||
| # Using Pathfinder to initializing sampling with Turing | |||
|
|
|||
| If you have a _differentiable_ Turing model, then you can use Pathfinder to initialize its parameters. | |||
There was a problem hiding this comment.
Just my two cents. Differentiable Turing model means one that do not have for loops and integers? Like a model which would run in Zygote or ReverseDiff?
Maybe explain with a footnote?
There was a problem hiding this comment.
Thanks, yeah I wrote the text hastily. Should say the technical requirements are that all parameters are continuous and the log-density can be differentiated by an AD framework.
There was a problem hiding this comment.
Yes, that what I got the intuition. But a footnote would solve this all. Thanks!
|
With #25, we can run Pathfinder on a Turing prob = optim_problem(model, MAP(); constrained=false)
pathfinder(prob, 1)It's probably worth it to add a conditionally loaded interface for a Turing model that makes it easy to draw samples into a format useful for Turing users, e.g. as an |
This PR introduces a docs page showing how to initialize Turing with Pathfinder. Note that to initialize the metric, we'd need to work with AdvancedHMC directly, but this is really clunky right now due to the complexity of getting the unconstrained log density. Hopefully there's a simpler way than how I'm currently doing it!