pass X to OutcomeTransform (pytorch#2663)

Summary:

This enables using outcome transforms with behavior that depends on X. For example, this enables implementing a stratified standardize transform, where the the standardization is performing for distinct values of an input dimension.

Differential Revision: D67724473

botorch/acquisition/analytic.py

@@ -1116,7 +1116,7 @@ def _get_noiseless_fantasy_model(
         # Not transforming Yvar because 1e-7 is already close to 0 and it is a
         # relative, not absolute, value.
         Y_fantasized, _ = outcome_transform(
-            Y_fantasized.unsqueeze(-1), Yvar.unsqueeze(-1)
+            Y_fantasized.unsqueeze(-1), Yvar.unsqueeze(-1), X=batch_X_observed
         )
         Y_fantasized = Y_fantasized.squeeze(-1)
     input_transform = getattr(model, "input_transform", None)

botorch/models/approximate_gp.py

@@ -172,7 +172,7 @@ def posterior(
 
         posterior = GPyTorchPosterior(distribution=dist)
         if hasattr(self, "outcome_transform"):
-            posterior = self.outcome_transform.untransform_posterior(posterior)
+            posterior = self.outcome_transform.untransform_posterior(posterior, X=X)
         if posterior_transform is not None:
             posterior = posterior_transform(posterior)
         return posterior
@@ -397,7 +397,7 @@ def __init__(
                     UserInputWarning,
                     stacklevel=3,
                 )
-                train_Y, _ = outcome_transform(train_Y)
+                train_Y, _ = outcome_transform(train_Y, X=transformed_X)
             self._validate_tensor_args(X=transformed_X, Y=train_Y)
             validate_input_scaling(train_X=transformed_X, train_Y=train_Y)
             if train_Y.shape[-1] != num_outputs:

botorch/models/ensemble.py

@@ -79,7 +79,7 @@ def posterior(
         # `posterior` (as is done in GP models). This is more general since it works
         # even if the transform doesn't support `untransform_posterior`.
         if hasattr(self, "outcome_transform"):
-            values, _ = self.outcome_transform.untransform(values)
+            values, _ = self.outcome_transform.untransform(values, X=X)
         if output_indices is not None:
             values = values[..., output_indices]
         posterior = EnsemblePosterior(values=values)

botorch/models/fully_bayesian.py

@@ -373,7 +373,9 @@ def __init__(
                 X=train_X, input_transform=input_transform
             )
         if outcome_transform is not None:
-            train_Y, train_Yvar = outcome_transform(train_Y, train_Yvar)
+            train_Y, train_Yvar = outcome_transform(
+                Y=train_Y, Yvar=train_Yvar, X=transformed_X
+            )
         self._validate_tensor_args(X=transformed_X, Y=train_Y)
         validate_input_scaling(
             train_X=transformed_X, train_Y=train_Y, train_Yvar=train_Yvar

botorch/models/fully_bayesian_multitask.py

@@ -242,7 +242,9 @@ def __init__(
             )
         if outcome_transform is not None:
             outcome_transform.train()  # Ensure we learn parameters here on init
-            train_Y, train_Yvar = outcome_transform(train_Y, train_Yvar)
+            train_Y, train_Yvar = outcome_transform(
+                Y=train_Y, Yvar=train_Yvar, X=transformed_X
+            )
         if train_Yvar is not None:  # Clamp after transforming
             train_Yvar = train_Yvar.clamp(MIN_INFERRED_NOISE_LEVEL)
 

botorch/models/gp_regression.py

@@ -160,7 +160,9 @@ def __init__(
                 X=train_X, input_transform=input_transform
             )
         if outcome_transform is not None:
-            train_Y, train_Yvar = outcome_transform(train_Y, train_Yvar)
+            train_Y, train_Yvar = outcome_transform(
+                Y=train_Y, Yvar=train_Yvar, X=transformed_X
+            )
         # Validate again after applying the transforms
         self._validate_tensor_args(X=transformed_X, Y=train_Y, Yvar=train_Yvar)
         ignore_X_dims = getattr(self, "_ignore_X_dims_scaling_check", None)

botorch/models/gpytorch.py

@@ -198,7 +198,7 @@ def posterior(
                     mvn = self.likelihood(mvn, X)
         posterior = GPyTorchPosterior(distribution=mvn)
         if hasattr(self, "outcome_transform"):
-            posterior = self.outcome_transform.untransform_posterior(posterior)
+            posterior = self.outcome_transform.untransform_posterior(posterior, X=X)
         if posterior_transform is not None:
             return posterior_transform(posterior)
         return posterior
@@ -244,7 +244,7 @@ def condition_on_observations(
             # (unless we've already trasnformed if BatchedMultiOutputGPyTorchModel)
             if not isinstance(self, BatchedMultiOutputGPyTorchModel):
                 # `noise` is assumed to already be outcome-transformed.
-                Y, _ = self.outcome_transform(Y=Y, Yvar=Yvar)
+                Y, _ = self.outcome_transform(Y=Y, Yvar=Yvar, X=X)
         # Validate using strict=False, since we cannot tell if Y has an explicit
         # output dimension. Do not check shapes when fantasizing as they are
         # not expected to match.
@@ -467,7 +467,7 @@ def posterior(
 
         posterior = GPyTorchPosterior(distribution=mvn)
         if hasattr(self, "outcome_transform"):
-            posterior = self.outcome_transform.untransform_posterior(posterior)
+            posterior = self.outcome_transform.untransform_posterior(posterior, X=X)
         if posterior_transform is not None:
             return posterior_transform(posterior)
         return posterior
@@ -511,7 +511,7 @@ def condition_on_observations(
         if hasattr(self, "outcome_transform"):
             # We need to apply transforms before shifting batch indices around.
             # `noise` is assumed to already be outcome-transformed.
-            Y, _ = self.outcome_transform(Y)
+            Y, _ = self.outcome_transform(Y, X=X)
         # Do not check shapes when fantasizing as they are not expected to match.
         if fantasize_flag.off():
             self._validate_tensor_args(X=X, Y=Y, Yvar=noise, strict=False)
@@ -924,7 +924,7 @@ def posterior(
             )
             posterior = GPyTorchPosterior(distribution=mtmvn)
         if hasattr(self, "outcome_transform"):
-            posterior = self.outcome_transform.untransform_posterior(posterior)
+            posterior = self.outcome_transform.untransform_posterior(posterior, X=X)
         if posterior_transform is not None:
             return posterior_transform(posterior)
         return posterior

botorch/models/higher_order_gp.py

@@ -91,7 +91,7 @@ def _return_to_output_shape(self, tsr: Tensor) -> Tensor:
         return out
 
     def forward(
-        self, Y: Tensor, Yvar: Tensor | None = None
+        self, Y: Tensor, Yvar: Tensor | None = None, X: Tensor | None = None
     ) -> tuple[Tensor, Tensor | None]:
         Y = self._squeeze_to_single_output(Y)
         if Yvar is not None:
@@ -107,7 +107,7 @@ def forward(
         return Y_out, Yvar_out
 
     def untransform(
-        self, Y: Tensor, Yvar: Tensor | None = None
+        self, Y: Tensor, Yvar: Tensor | None = None, X: Tensor | None = None
     ) -> tuple[Tensor, Tensor | None]:
         Y = self._squeeze_to_single_output(Y)
         if Yvar is not None:
@@ -121,7 +121,7 @@ def untransform(
         return Y, Yvar
 
     def untransform_posterior(
-        self, posterior: HigherOrderGPPosterior
+        self, posterior: HigherOrderGPPosterior, X: Tensor | None = None
     ) -> TransformedPosterior:
         # TODO: return a HigherOrderGPPosterior once rescaling constant
         # muls * LinearOperators won't force a dense decomposition rather than a
@@ -227,7 +227,7 @@ def __init__(
                     output_shape=train_Y.shape[-num_output_dims:],
                     batch_shape=batch_shape,
                 )
-            train_Y, _ = outcome_transform(train_Y)
+            train_Y, _ = outcome_transform(train_Y, X=train_X)
 
         self._aug_batch_shape = batch_shape
         self._num_dimensions = num_output_dims + 1
@@ -416,7 +416,7 @@ def condition_on_observations(
         """
         if hasattr(self, "outcome_transform"):
             # we need to apply transforms before shifting batch indices around
-            Y, noise = self.outcome_transform(Y=Y, Yvar=noise)
+            Y, noise = self.outcome_transform(Y=Y, Yvar=noise, X=X)
         # Do not check shapes when fantasizing as they are not expected to match.
         if fantasize_flag.off():
             self._validate_tensor_args(X=X, Y=Y, Yvar=noise, strict=False)
@@ -540,7 +540,7 @@ def posterior(
                 num_outputs=self._num_outputs,
             )
             if hasattr(self, "outcome_transform"):
-                posterior = self.outcome_transform.untransform_posterior(posterior)
+                posterior = self.outcome_transform.untransform_posterior(posterior, X=X)
             return posterior
 
     def make_posterior_variances(

botorch/models/latent_kronecker_gp.py

@@ -81,7 +81,7 @@ def __init__(
         self._use_min = use_min
 
     def forward(
-        self, Y: Tensor, Yvar: Tensor | None = None
+        self, Y: Tensor, Yvar: Tensor | None = None, X: Tensor | None = None
     ) -> tuple[Tensor, Tensor | None]:
         r"""Standardize outcomes.
 
@@ -93,6 +93,7 @@ def forward(
             Y: A `batch_shape x n x m`-dim tensor of training targets.
             Yvar: A `batch_shape x n x m`-dim tensor of observation noises
                 associated with the training targets (if applicable).
+            X: A `batch_shape x n x d`-dim tensor of training inputs (if applicable).
 
         Returns:
             A two-tuple with the transformed outcomes:
@@ -240,7 +241,9 @@ def __init__(
             outcome_transform = MinMaxStandardize(batch_shape=batch_shape)
         if outcome_transform is not None:
             # transform outputs once and keep the results
-            train_Y = outcome_transform(train_Y.unsqueeze(-1))[0].squeeze(-1)
+            train_Y = outcome_transform(train_Y.unsqueeze(-1), X=transformed_X)[
+                0
+            ].squeeze(-1)
 
         ExactGP.__init__(
             self,
@@ -506,7 +509,7 @@ def _rsample_from_base_samples(
         )
         # samples.shape = (*sample_shape, *broadcast_shape, n_test_x, n_t)
         if hasattr(self, "outcome_transform") and self.outcome_transform is not None:
-            samples, _ = self.outcome_transform.untransform(samples)
+            samples, _ = self.outcome_transform.untransform(samples, X=X)
         return samples
 
     def condition_on_observations(

botorch/models/model_list_gp_regression.py

@@ -117,7 +117,7 @@ def condition_on_observations(
             else:
                 noise_i = torch.cat([noise[..., k] for k in range(i, j)], dim=-1)
             if hasattr(model, "outcome_transform"):
-                y_i, noise_i = model.outcome_transform(y_i, noise_i)
+                y_i, noise_i = model.outcome_transform(y_i, noise_i, X=X_i)
                 if noise_i is not None:
                     noise_i = noise_i.squeeze(0)
             targets.append(y_i)

botorch/models/multitask.py

@@ -219,7 +219,9 @@ def __init__(
         if outcome_transform == DEFAULT:
             outcome_transform = Standardize(m=1, batch_shape=train_X.shape[:-2])
         if outcome_transform is not None:
-            train_Y, train_Yvar = outcome_transform(Y=train_Y, Yvar=train_Yvar)
+            train_Y, train_Yvar = outcome_transform(
+                Y=train_Y, Yvar=train_Yvar, X=transformed_X
+            )
 
         # squeeze output dim
         train_Y = train_Y.squeeze(-1)
@@ -464,7 +466,7 @@ def __init__(
                 X=train_X, input_transform=input_transform
             )
         if outcome_transform is not None:
-            train_Y, _ = outcome_transform(train_Y)
+            train_Y, _ = outcome_transform(train_Y, X=transformed_X)
 
         self._validate_tensor_args(X=transformed_X, Y=train_Y)
         self._num_outputs = train_Y.shape[-1]
@@ -772,7 +774,7 @@ def posterior(
         )
 
         if hasattr(self, "outcome_transform"):
-            posterior = self.outcome_transform.untransform_posterior(posterior)
+            posterior = self.outcome_transform.untransform_posterior(posterior, X=X)
         return posterior
 
     def train(self, val=True, *args, **kwargs):