Expand model space for numeric ordered ChoiceParameters (#5210)#5210
Open
saitcakmak wants to merge 1 commit into
Open
Expand model space for numeric ordered ChoiceParameters (#5210)#5210saitcakmak wants to merge 1 commit into
saitcakmak wants to merge 1 commit into
Conversation
meta-cla
Bot
added
the
CLA Signed
|
@saitcakmak has exported this pull request. If you are a Meta employee, you can view the originating Diff in D81179672. |
Summary: Pull Request resolved: facebook#5210 Closes facebook#5209. Extends `Adapter._set_model_space` to expand `ChoiceParameter.values` for numeric ordered choice parameters whose training data falls outside the declared `values` list, mirroring the existing behavior for `RangeParameter` bounds. Previously, only `RangeParameter` was expanded and out-of-design choice values were filtered, which was surprising for users who expect symmetric treatment. Expansion is gated to numeric ordered choice parameters. Unordered (one-hot-encoded) and string choice parameters are left alone because the downstream `OneHot` and `OrderedChoiceToIntegerRange` transforms would otherwise produce a transformed model space that is inconsistent with the transformed search space used at generation time, corrupting `extract_search_space_digest` and `Adapter.parameters`. With D87903513 enforcing `sort_values=True` for numeric ordered parameters, the gated case is safe to expand and always sortable. Implementation notes: - New observed values are deduped and merged with existing `p.values`, then sorted before being passed to `set_values`. `set_values` (unlike `ChoiceParameter.__init__`) does not dedupe its input, and duplicates would corrupt the integer encodings produced downstream by `OrderedChoiceToIntegerRange`. - If a numeric ordered `ChoiceParameter` is declared as `INT` but an observed value is non-integer (e.g. `2.5`), the model-space copy of the parameter is relaxed to `FLOAT` so the observation is preserved exactly rather than truncated by `_cast_values`. This is safe because `_model_space` is adapter-local and never persisted; the user-facing `experiment.search_space` is untouched. Downstream consumers of `ChoiceParameter` (`OrderedChoiceToIntegerRange`, `ChoiceToNumericChoice`, etc.) do not branch on INT vs FLOAT. Differential Revision: D81179672
saitcakmak
force-pushed
the
export-D81179672
branch
from
5caebcd to
c4bf40e
Compare
meta-codesync
Bot
changed the title
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #5210 +/- ##
=======================================
Coverage 96.50% 96.51%
=======================================
Files 617 617
Lines 69626 69694 +68
=======================================
+ Hits 67194 67262 +68
Misses 2432 2432 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Summary:
Closes #5209.
Extends
Adapter._set_model_spaceto expandChoiceParameter.valuesfor numeric ordered choice parameters whose training data falls outside the declaredvalueslist, mirroring the existing behavior forRangeParameterbounds. Previously, onlyRangeParameterwas expanded and out-of-design choice values were filtered, which was surprising for users who expect symmetric treatment.Expansion is gated to numeric ordered choice parameters. Unordered (one-hot-encoded) and string choice parameters are left alone because the downstream
OneHotandOrderedChoiceToIntegerRangetransforms would otherwise produce a transformed model space that is inconsistent with the transformed search space used at generation time, corruptingextract_search_space_digestandAdapter.parameters. With D87903513 enforcingsort_values=Truefor numeric ordered parameters, the gated case is safe to expand and always sortable.Implementation notes:
p.values, then sorted before being passed toset_values.set_values(unlikeChoiceParameter.__init__) does not dedupe its input, and duplicates would corrupt the integer encodings produced downstream byOrderedChoiceToIntegerRange.ChoiceParameteris declared asINTbut an observed value is non-integer (e.g.2.5), the model-space copy of the parameter is relaxed toFLOATso the observation is preserved exactly rather than truncated by_cast_values. This is safe because_model_spaceis adapter-local and never persisted; the user-facingexperiment.search_spaceis untouched. Downstream consumers ofChoiceParameter(OrderedChoiceToIntegerRange,ChoiceToNumericChoice, etc.) do not branch on INT vs FLOAT.Differential Revision: D81179672