Surrogate Posteriors #309
Merged
Surrogate Posteriors #309
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
Scienfitz
reviewed
Jul 9, 2024
23f121c to
e1279bf
Compare
AVHopp
reviewed
Jul 10, 2024
AVHopp
reviewed
Jul 10, 2024
e1279bf to
9d53245
Compare
9d53245 to
13e3353
Compare
13e3353 to
32e2ef5
Compare
AVHopp
approved these changes
Jul 12, 2024
Scienfitz
approved these changes
Jul 12, 2024
AVHopp
reviewed
Jul 12, 2024
Merged
AdrianSosic
added a commit
that referenced
this pull request
Aug 29, 2024
Completes the surrogate factoring, which extended over #278, #309, #315, #325, #337. ### Most important changes * The transition point from experimental to computational representation has been moved from the recommender to the surrogate. From an architecture/responsibility perspective, this is reasonable since the recommend should not have to bother about algorithmic/computational details. * The desired consequence is that public `Surrogate` methods like `posterior` and `fit` can now operate on dataframes in experimental representation, meaning they can also be exposed directly to the user. * The new posterior methods now all return a general `Posterior` object instead of implicitly assuming Gaussian distributions. This paves the way for arbitrary surrogate extensions, such as Bernoulli/Categorical surrogates, etc. At the moment, this introduces an explicit coupling to botorch, which is fine because botorch remains a core dependency and the only backend used for complex surrogate modeling. In the future, this can be further abstracted by introducing our own `Posterior` class. * The `Surrogate` layout has been refined such that the extracted `SurrogateProtocol`, which now defines the formal interface for all surrogates, imposes minimal requirements to the user. * Scaling has been completely redesigned, offering the possibility to configure input/output scaling down to the level of individual parameters and targets. The configuration is currently class-specific, but can be extended to allow surrogate instance specific rules in the future.
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 is a next step toward a lean surrogate class layout:
Surrogate.(_)posteriornow returns aPosteriorobjectSurrogate._fitno longer expects theSearchSpaceas an argument, which brings us closer to the state that.fitand.posterioroperate on the user/dataframe/context-level while_posteriorand_fitoperate on the purely mathematical level. This means that a user who writes their own surrogate class effectively only needs to implement the corresponding mathematical model in the latter two methods. Optional context information that may be required for this implementation (like the dimension index of theTaskParameterin the passedTensorobject) is now encapsulated into a surrogate-specificcontextobject, that can be arbitrarily populated by the surrogate class, but whose logic is now cleanly separated from the actual fitting logic.GaussianSurrogatebase class for (most our other) models that come with the implicit Gaussian noise assumption and effectively only implement mean and (co-)variance estimation.Removes current scaling logic --> will be redesigned in upcoming PRcatch_constant_target, re-enabling slots forSurrogatesRemovesregister_custom_architecture--> will replaced withSurrogateProtocolin upcoming PR