Replace AD package extensions with DifferentiationInterface

[gdalle]

Fixes #49. Here are the main changes:

• Add dependency on ADTypes and DifferentiationInterface
• Implement _derivative, _gradient, _jacobian and _hessian using DI
• Remove package extensions for FiniteDiff, FiniteDifferences, ForwardDiff, ReverseDiff, Zygote
• Replace backend constructors with their ADTypes counterparts everywhere.
• Remove package extensions from docs.
• Figure out what to do about AbstractDiffBackend, since it is apparently used downstream. Replace it with Union{ADTypes.AbstractADType, ExplicitEmbeddedBackend}? Right now I removed annotations in this package, and set AbstractDiffBackend = Any for downstream uses.
• Fix broken test. It seems that we try to compute a Jacobian for a function with a scalar output?
• Bump version to v0.4.0.
• Undo changes for testing with next Manifolds version.

[mateuszbaran]

Thanks for working on this update 👍

• Figure out what to do about AbstractDiffBackend, since it is apparently used downstream. Replace it with Union{ADTypes.AbstractADType, ExplicitEmbeddedBackend}? Right now I removed annotations in this package, and set AbstractDiffBackend = Any for downstream uses.

Since this is a breaking change, I think we can just remove AbstractDiffBackend entirely and switch to ADTypes.AbstractADType in downstream packages.

• Fix broken test. It seems that we try to compute a Jacobian for a function with a scalar output?

It's technically valid but I don't think it's a feature we rely on anywhere. I think we can switch the test to use a vector-valued function.

[gdalle]

Since this is a breaking change, I think we can just remove AbstractDiffBackend entirely and switch to ADTypes.AbstractADType in downstream packages.

I agree, but this would prevent us from testing here because the test dependencies contain a version of Manifolds.jl which still requires AbstractDiffBackend.

[mateuszbaran]

Right, in this case making AbstractDiffBackend an alias for Any is useful.

[gdalle]

The fact that we use Manifolds in testing also means we cannot bump ManifoldsDiff to v0.4.0 in the current state, because it is not permitted by the Manifolds compat bound. The right way would be testing with a dev version of Manifolds where the compat v0.4.0 has been added for ManifoldsDiff. That would also make the AbstractDiffBackend = Any hotfix superfluous.

[codecov]

Codecov Report

Attention: Patch coverage is 95.00000% with 1 line in your changes missing coverage. Please review.

Project coverage is 95.69%. Comparing base (b8c3464) to head (7a50d38).
Report is 9 commits behind head on main.

Files with missing lines	Patch %	Lines
src/ManifoldDiff.jl	83.33%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #51      +/-   ##
==========================================
+ Coverage   95.46%   95.69%   +0.22%     
==========================================
  Files          21       13       -8     
  Lines         375      395      +20     
==========================================
+ Hits          358      378      +20     
  Misses         17       17              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

---

🚨 Try these New Features:

• Flaky Tests Detection - Detect and resolve failed and flaky tests

[mateuszbaran]

That's what I sometimes do locally when making breaking changes but I don't know about any way to do it on CI.

[gdalle]

The last commit gives you an example of how to do that, using my custom branch of Manifolds where the compat entry has been upgraded and the AbstractDiffBackend import has been removed.

Warning

Don't merge before undoing those changes.

[mateuszbaran]

Nice, I didn't know that trick.

[kellertuer]

This indeed looks like a very neat trick! Thanks for illustrating that here.

[gdalle]

Before we go any further here, I'm making sure that Manifolds itself can handle this ManifoldsDiff v0.4.0. I have opened a symmetric PR (JuliaManifolds/Manifolds.jl#771) which checks out my custom branck of ManifoldsDiff.

[gdalle]

All the tests are passing with my custom branch of the downstream Manifolds (except codecov). I will now remove my Pkg tricks, the tests will thus fail but it will still be good to merge.

[gdalle]

@kellertuer this is good to go

[mateuszbaran]

LGTM.

By the way, is JuliaManifolds/Manifolds.jl#772 good to go also?

[gdalle]

Yeah you can ship that one too