Replace Differential implementation with simplified Dual (2/N) #163
Conversation
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #163 +/- ##
==========================================
- Coverage 87.75% 87.68% -0.07%
==========================================
Files 100 99 -1
Lines 15982 15811 -171
Branches 856 852 -4
==========================================
- Hits 14025 13864 -161
+ Misses 1101 1095 -6
+ Partials 856 852 -4 ☔ View full report in Codecov by Sentry. |
sritchie
changed the title
| @@ -530,8 +531,11 @@ | |||
| (letfn [(process-term [term] | |||
| (g/simplify | |||
| (s/mapr (fn rec [x] | |||
| (if (d/differential? x) | |||
| (d/map-coefficients rec x) | |||
| (if (d/dual? x) | |||
There was a problem hiding this comment.
does this come up often? it looks like a kind of fmap over Duals
There was a problem hiding this comment.
This has always seemed like a hack to me... but maybe not. yes, it's like an fmap for sure, and it's what we do for simplify. We'll have a similar line to add for tapes. What do you think? Should we add something more general?
| *active-tags*) | ||
| (apply max tags))) | ||
|
|
||
| (defn tag+perturbation |
There was a problem hiding this comment.
could you just return the dual number? it has the tag inside it already?
There was a problem hiding this comment.
YES that is true... I added this on my scratch branch when I still had Differential and it was a touch more expensive to get the tag out, so I wanted to save the effort.
Can we look at this together when I massage this to be compatible with tapes too?
| (defmethod g/one-like [::dual] [_] 1) | ||
| (defmethod g/identity-like [::dual] [_] 1) | ||
| (defmethod g/freeze [::dual] [d] | ||
| `[~'Dual |
There was a problem hiding this comment.
I swore a saw a dict-like implementation of this somewhere
There was a problem hiding this comment.
I had a pprint implementation for tapes that did that:
https://github.com/mentat-collective/emmy/blob/main/src/emmy/tape.cljc#L285-L296
but I didn't use it for freeze. Honestly I don't think these should ever make it out into a situation where we'd inspect the results of a freeze...
This PR
Differentialgeneralized dual and its term list algebra with a simplifiedDualnumber typeemmy.util.vector-setand tests, as these are no longer usednilimplementation forextract-tangent, meaning thatnil-valued functions now work withDThis new approach works because the
emmy.differential/*active-tags*stack allows us to make sure that lifted binary operations always wrap their output in a newDualwith the tag assigned by the inner-most derivative call.