Preserving train inputs and targets through transforms #3044
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
meta-cla
bot
added
the
CLA Signed
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 14, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. TODO: - Account for non-invertible transforms Differential Revision: D84571407
hvarfner
force-pushed
the
export-D84571407
branch
from
2466e2b to
e17f003
Compare
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 15, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. TODO: - Account for non-invertible transforms Differential Revision: D84571407
hvarfner
force-pushed
the
export-D84571407
branch
from
e17f003 to
24f101c
Compare
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 15, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. Notebook explaining the effect with some plots: N8342965 Reviewed By: Balandat Differential Revision: D84571407
hvarfner
force-pushed
the
export-D84571407
branch
2 times, most recently
from
9460560 to
a4163d0
Compare
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 16, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. Notebook explaining the effect with some plots: N8342965 Reviewed By: Balandat Differential Revision: D84571407
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #3044 +/- ##
=======================================
Coverage 99.98% 99.98%
=======================================
Files 216 216
Lines 20581 20635 +54
=======================================
+ Hits 20577 20631 +54
Misses 4 4 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
hvarfner
force-pushed
the
export-D84571407
branch
from
a4163d0 to
50dc3ba
Compare
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 17, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. Notebook explaining the effect with some plots: N8342965 Reviewed By: Balandat Differential Revision: D84571407
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 17, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. Notebook explaining the effect with some plots: N8342965 Reviewed By: Balandat Differential Revision: D84571407
hvarfner
force-pushed
the
export-D84571407
branch
from
50dc3ba to
9109fe4
Compare
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 17, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. Notebook explaining the effect with some plots: N8342965 Reviewed By: Balandat Differential Revision: D84571407
hvarfner
force-pushed
the
export-D84571407
branch
from
9109fe4 to
9341715
Compare
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 17, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. Notebook explaining the effect with some plots: N8342965 Reviewed By: Balandat Differential Revision: D84571407
hvarfner
force-pushed
the
export-D84571407
branch
from
9341715 to
82223a3
Compare
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 17, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. Notebook explaining the effect with some plots: N8342965 Reviewed By: Balandat Differential Revision: D84571407
hvarfner
force-pushed
the
export-D84571407
branch
from
82223a3 to
ce7c08f
Compare
hvarfner
pushed a commit
to hvarfner/botorch
that referenced
this pull request
Oct 17, 2025
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. Notebook explaining the effect with some plots: N8342965 Reviewed By: Balandat Differential Revision: D84571407
hvarfner
force-pushed
the
export-D84571407
branch
from
ce7c08f to
2fdfe78
Compare
…3044) Summary: This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out. __Longer explanation:__ Transforms, and specifically learnable output transforms like Standardize, will currently: a. Learn the parameters at initialization of the GP b. Transform the train_Ys to the normalized space Then, when we load a state dict, we will: a. Impose new standardization parameters on already standardized data b. Potentially make the transforms re-learnable, nullifying the change made by the state dict This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, _we don't simply leave one point out, but instead we leave one out and re-standardize_. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters. Notebook explaining the effect with some plots: N8342965 Reviewed By: Balandat, saitcakmak Differential Revision: D84571407
hvarfner
force-pushed
the
export-D84571407
branch
from
2fdfe78 to
c47e45c
Compare
|
This pull request has been merged in b0d492d. |
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:
This PR preserves botorch transforms (specifically outcome_transforms, like Standardize) through state_dict loading. The fix also ensures that train_targets of a Leave-one-out model with outcome transforms will, in the default case, have the same targets as a base model, minus the point left out.
Longer explanation:
Transforms, and specifically learnable output transforms like Standardize, will currently:
a. Learn the parameters at initialization of the GP
b. Transform the train_Ys to the normalized space
Then, when we load a state dict, we will:
a. Impose new standardization parameters on already standardized data
b. Potentially make the transforms re-learnable, nullifying the change made by the state dict
This has undesired consequences for cross-validation, as all cross-validated models will effectively have different training data. In essence, we don't simply leave one point out, but instead we leave one out and re-standardize. When we have outliers in the data, this will lead to substantially different predictions when the outlier is left out, since the outlier will substantially impact the outcome transform parameters.
TODO:
Differential Revision: D84571407