Initial integration with ExponentialFamilyProjection #408
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
albertpod
approved these changes
Jul 19, 2024
Nimrais
reviewed
Jul 19, 2024
| end | ||
|
|
||
| # # gradient function | ||
| ## I think this is wrong. This is not a gradient on the manifolds. It is just Euclidean gradient. |
There was a problem hiding this comment.
We did projection on the last line of this function, X after is actually laying in the Tangent space of the point p.
| ef = convert(ExponentialFamilyDistribution, M, p) | ||
| ifisher = cholinv(Hermitian(fisherinformation(ef))) | ||
| X = ExponentialFamilyProjection.ExponentialFamilyManifolds.partition_point(M, ifisher * ForwardDiff.gradient((p) -> targetfn(M, p, data), p)) | ||
| X = ExponentialFamilyProjection.ExponentialFamilyManifolds.ManifoldsBase.project(M, p, X) |
There was a problem hiding this comment.
Maybe ManifoldsBase.project!(M, X, p, X)?
| ## Option 3 | ||
| # T = ExponentialFamily.exponential_family_typetag(q_out) | ||
| # s = sampling_optimized(q_out) | ||
| # d = fit_mle(typeof(s), q_out_samples) |
| @@ -279,14 +279,15 @@ end | |||
| ## We create a lambda-like callable structure to improve type inference and make it more stable | |||
| ## However it is not fully inferrable due to dynamic tags and variable constraints, but still better than just a raw lambda callback | |||
|
|
|||
| struct MessageMapping{F, T, C, N, M, A, X, R} | |||
| struct MessageMapping{F, T, C, N, M, A, X, R, K} | |||
There was a problem hiding this comment.
I think it could be valuable to group fields here into smaller structures. Like there is a certainly cluster of names.
But this is minor comment.
| (FactorizedJoint((Exponential(0.5),)), Exponential(3), Rayleigh(1 / (sqrt(2 * 3))), x -> sqrt(x)) | ||
| # (FactorizedJoint((Exponential(0.5),)), Exponential(3), Geometric(1 - exp(-0.5)), x->ceil(x)) | ||
| # (FactorizedJoint((Exponential(0.5),)), Exponential(3), Pareto(3, 0.5), x->3*exp(x) ) ##exponential family projection errors | ||
| # (FactorizedJoint((Exponential(0.5),)), Exponential(30), Beta(0.5, 1), x->exp(-x) ) ##exponential family projection errors |
There was a problem hiding this comment.
This test does work # (FactorizedJoint((Exponential(0.5),)), Exponential(30), Beta(0.5, 1), x->exp(-x) ) ##exponential family projection errors
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR adds initial support for
ExponentialFamilyProjectioninReactiveMP.jl. It also introduces a new feature for the inference backend that uses fallback rules when explicit message-passing update rules are not available. This functionality acts globally for all nodes, similar tometa. Currently, only a simple fallback is implemented, and it is applicable only to stochastic nodes.Remaining items:
CVIProjectionThis PR is accompanied by a corresponding PR to
RxInfer.jl. Inference tests are available inRxInfer, and this PR includes tests for individual rules only.As a note, @ismailsenoz has an improved version of the rules, which require additional time for testing and setup due to dependencies on the
AdvancedHMClibrary. The plan is to integrateAdvancedHMC-based rules following this PR.