(botorch) fix noise term in PairwiseGP and add ScaleKernel #571
Conversation
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
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D23854124 |
Balandat
pushed a commit
to Balandat/botorch
that referenced
this pull request
Sep 28, 2020
Summary: Pull Request resolved: meta-pytorch#571 In a probit model, the noise is simply a scaling parameter on the function. In the current implementation, the scale is implemented via the noise, which has two disadvantages: 1. A user coming from anywhere else in gpytorch/botorch wouldn't necessarily know not to use ScaleKernel here. 2. When running prediction / computing posteriors, by default we don't include the noise (consistent with behavior elsewhere), so our prediction of f(x) is off by an arbitrary constant factor. This changes the default so that we use ScaleKernel and fix the noise. Differential Revision: D23854124 fbshipit-source-id: d8ab08264cfc7c09ab027f308ea39921b8147daa
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D23854124 |
Balandat
force-pushed
the
export-D23854124
branch
from
41a5119 to
613af36
Compare
Codecov Report
@@ Coverage Diff @@
## master #571 +/- ##
=========================================
Coverage 100.00% 100.00%
=========================================
Files 84 84
Lines 5503 5485 -18
=========================================
- Hits 5503 5485 -18
Continue to review full report at Codecov.
|
…pytorch#571) Summary: Pull Request resolved: meta-pytorch#571 In a probit model, the noise is not identifiable separately from a scale on latent function. In the current implementation, this scale is implemented via a noise term, which has three disadvantages: 1. A user coming from anywhere else in gpytorch/botorch wouldn't necessarily know not to use ScaleKernel here. 2. When running prediction / computing posteriors, by default we don't include the noise (consistent with behavior elsewhere), so our prediction of f(x) is off by an arbitrary constant factor. 3. When computing posteriors with `noise=True` the noise is added to the covariance, but this doesn't scale the function value either. This means if I use a batched single-outcome acquisition function to do pairwise acquisition (which is reasonable), I'm not doing acquisition on the same model I'm using for interpolation (since now I have noise on individual items rather than on the comparison). Explicitly pairwise MC acquisition functions that take draws of pairs and do something with them should still be correct here, I think. This PR changes the default so that we use ScaleKernel and remove the noise term. I've added docs to this effect in a few places, but this is a breaking change API-wise. Reviewed By: Balandat Differential Revision: D23854124 fbshipit-source-id: da802d9df05ac1cb873b4f7e4fe6ff9bfa67da8a
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D23854124 |
Balandat
force-pushed
the
export-D23854124
branch
from
613af36 to
4c56e7c
Compare
Contributor
|
This pull request has been merged in dd17a4c. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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:
In a probit model, the noise is simply a scaling parameter on the function. In the current implementation, the scale is implemented via the noise, which has two disadvantages:
This changes the default so that we use ScaleKernel and fix the noise.
Differential Revision: D23854124