Introduce an EmbeddedObjective #286
Conversation
Codecov Report
@@ Coverage Diff @@
## master #286 +/- ##
========================================
Coverage 99.73% 99.74%
========================================
Files 76 77 +1
Lines 6892 7017 +125
========================================
+ Hits 6874 6999 +125
Misses 18 18
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
There was a problem hiding this comment.
I think the general idea looks good, I think just a few minor comments.
This PR resolves #217 by proposing an EmbeddedObjective– maybe one could also better all it EmbeddedManifoldObjective?
I think EmbeddedManifoldObjective would be more consistent with the other names?
which functions can we extend this to? Hessian for sure, maybe also proximal maps.
Is there actually a generic way to convert a Euclidean proximal map to a Riemannian one?
Co-authored-by: Mateusz Baran <mateuszbaran89@gmail.com>
That's what I thought as well when already writing this PR ;)
You might be right, that that is not the case – but the constraints and their gradients, the differentials, they should be possible as well. |
…lds/Manopt.jl into kellertuer/EmbeddedObjective
|
I think I now also sketched the hessian, before doing that for the constraints – since that is quite some work (though more in writing cases than in much thinking) I would like to discuss/check whether
edit: also this one for now has to wait for JuliaManifolds/ManifoldDiff.jl#30 |
|
In order to have a nice tutorial, this probably also should wait for JuliaManifolds/Manifolds.jl#644 |
kellertuer
added
enhancement
WIP
# Conflicts: # Project.toml # docs/make.jl # src/plans/plan.jl
|
This is just missing a bit of testing, but the tutorial is at least basically finished – it needs JuliaManifolds/Manifolds.jl#644, and I have to check while for now the Hessian conversion on Grassmann is not (yet) correct. |
kellertuer
added
Ready-for-Review
|
Sorry for the tag a bit too early, I missed the “automatic conversion to embedded objectives” for sub solvers (and was wondering why embedded worked so poorly with ARC). |
|
One thing I am a bit confused about it, that on https://manoptjl.org/previews/PR286/tutorials/EmbeddingObjectives/ the second run (with the euclidean ones converted) sometimes does work (yesterday evening) sometimes not (today) – on my machine that consistently needs 7 iterations and not 40 (which is the ARC default max iter). Not sure what the CI is doing there. |
|
Ah, the main point is that I am stupid. Quarto renders on the last registered version. So sure that is the old one and hence fails. I will trick that a bit for this tutorial now. |
This PR resolves #217 by proposing an
EmbeddedManifoldObjective.Idea
embed(M,p)get_gradient– but then also callsriemannian_gradienton the gradient the original objective returnsInterface
since we have the
decorate_objective!chain already available, just introducing theobjective_type=:Euclideankeyword there makes it possible for any objective to be wrapped. Currently also:Embeddedis equivalent, since it actually works more general then just Euclidean – but:Euclideanis easier to understand for most users.Internals
The
EmbeddedManifoldObjectivehas two (optionally used) fieldspandXfor a point and tangent vector in the embedding. That way one can useembed!orget_gradient!in the embedding. This should reduce allocations tremendously.ToDo
get_costget_gradient/get_gradient!get_hessian/get_hessian!get_grad_equality_constraint/get_grad_equality_constraint!(the the plural)get_grad_inequality_constraint/get_grad_inequality_constraint!(and the plural)