LearnedObjective and PairwiseGP dtype fixes and cleanup #2006
Conversation
facebook-github-bot
added
the
CLA Signed
|
@esantorella has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Codecov Report
@@ Coverage Diff @@
## main #2006 +/- ##
=========================================
Coverage 100.00% 100.00%
=========================================
Files 179 179
Lines 15771 15782 +11
=========================================
+ Hits 15771 15782 +11
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
botorch/acquisition/objective.py
Outdated
| def _get_learned_objective_pref_model_mixed_dtype_warn() -> str: | ||
| return ( | ||
| "pref_model has double-precision data, but single-precision data " | ||
| "was passed to the LearnedObjective. Upcasting to double." | ||
| ) |
There was a problem hiding this comment.
Why not make this a constant?
| datapoints: Optional[Tensor], | ||
| comparisons: Optional[Tensor], |
There was a problem hiding this comment.
Should we update the docstring with what happens when these are None?
Co-authored-by: Max Balandat <[email protected]>
|
@esantorella has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@esantorella has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@esantorella merged this pull request in 626eb41. |
Motivation
When
LearnedObjective.forwardis passed single-precision data, but it has apref_modelwith double-precision data and an input transform, the coefficients of the input transform can be downcast from double-precision to single-precision. To prevent precision loss, and to ensure that theLearnedObjectivegives the same result when called repeatedly, this PR up-casts the input data to double-precision in this situation.Changes:
PairwiseGPPairwiseGP.__init__so that they are passed explicitly rather than as part ofkwargsHave you read the Contributing Guidelines on pull requests?
Yes
Test Plan
Added units:
LearnedObjectivegives the same result when called multiple timesassertTrue(x is y)->assertIs(x, y))