Use ChainRulesCore.@not_implemented and extend tests
#308
Conversation
Codecov Report
@@ Coverage Diff @@
## master #308 +/- ##
==========================================
+ Coverage 89.28% 89.31% +0.02%
==========================================
Files 12 12
Lines 2633 2640 +7
==========================================
+ Hits 2351 2358 +7
Misses 282 282
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
Yeah. |
| (hankelh2(ν - 1, x) - hankelh2(ν + 1, x)) / 2, | ||
| ), | ||
| ) | ||
| ChainRulesCore.@scalar_rule( | ||
| polygamma(m, x), | ||
| ( | ||
| ChainRulesCore.@thunk(error("not implemented")), | ||
| ChainRulesCore.DoesNotExist(), |
There was a problem hiding this comment.
Why wasn't it already the case?
There was a problem hiding this comment.
I guess I did not pay enough attention when I moved the rules from ChainRules and just replaced all NaN with @thunk(error("not implemented")) (https://github.com/JuliaDiff/ChainRules.jl/blob/bf54b2f0e77fb05d3bc42004c448b3db0432f9cd/src/rulesets/packages/SpecialFunctions.jl#L57).
This PR updates the differentials to use the new
@not_implementedmacro instead of custom thunks witherror("not implemented"). Additionally, now all differentials are tested.The update to
@not_implementedhas multiple advantages:SpecialFunctions.bessel...errors with SpecialFunctions >= 1 - support thunks? FluxML/Zygote.jl#873: now it is possible to use even incomplete rules with ZygoteI assume the rules in ChainRules should be updated in the same way @oxinabox?