Closed DhairyaLGandhi closed 1 month ago
Attention: Patch coverage is 0%
with 33 lines
in your changes are missing coverage. Please review.
Project coverage is 29.16%. Comparing base (
a0fab7a
) to head (f817b52
). Report is 28 commits behind head on master.
Files | Patch % | Lines |
---|---|---|
ext/SciMLBaseZygoteExt.jl | 0.00% | 33 Missing :warning: |
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
Add your unit tests as a new downstream testset.
Note that with SciMLSensitivity.jl#dg/ss (and https://github.com/SciML/SciMLStructures.jl/pull/18) https://github.com/SciML/SciMLSensitivity.jl/blob/32f5ae7529a1957661b153f0ca9eff7e4caf0c5a/test/reversediff_output_types.jl#L14 looks like:
julia> gs = gradient(u0 -> loss(u0), u0)
([-0.7779831009550049, 0.40028226620020263],)
I've added a DAE example in the tests, but switched it off until we get SciMLSensitivity updated as well. The DC motor example fails to initialize currently. If there's a different test case, I can also hook that in.
@ChrisRackauckas SciMLSensitivity test pass with https://github.com/SciML/SciMLBase.jl/pull/689/commits/d061ce47f10d99f78850e08b820727c6228b3c8b (latest commit), but the Core (Downstream) tests get cancelled before anything runs. Is that because the Core (Python) tests fail for unrelated reasons?
So what happens here is:
1 dependency precompiled but a different version is currently loaded. Restart julia to access the new version
Both CI/ Python and CI/ Downgrade seem to be failing on master as well.
The problem I mentioned has not been fixed. It's not a problem with ADTypes per se, it's a problem with environment stacking
Is there anything left to be done in this PR?
Checklist
Additional context
Currently, ADing through observables errors, however this allows us to AD through the observable function via symbolic indexing and accumulate and return grads against
sol
This needs handling as part of when the observable symbol is in a collection (vector/ tuple/ ...), and also for various ADs like ReverseDiff and Enzyme.
Add any other context about the problem here.
Ideally, this would be handled by removing all the adjoints related to
getindex
and let AD do the heavy lifting for us. But this is faster to implement in its current form.