StratifiedStandardize OutcomeTransform (#2671)

Summary:
Pull Request resolved: #2671

see title. This allows applying stratified standardization at the model level, which will enable selecting whether to use a Single-task or multi-task model in Ax while using the appropriate transform. I.e. One could specify ModelConfigs that use 1) `SingleTaskGP` + `Standardize`, 2) `MultiTaskGP` + `StratifiedStandardize`.

Reviewed By: saitcakmak

Differential Revision: D67728920

fbshipit-source-id: ad6ee2bbed3e484288e94dcfb7b1555fbd4395e4

botorch/models/multitask.py

@@ -39,6 +39,7 @@
 from botorch.models.model import FantasizeMixin
 from botorch.models.transforms.input import InputTransform
 from botorch.models.transforms.outcome import OutcomeTransform, Standardize
+from botorch.models.utils.assorted import get_task_value_remapping
 from botorch.models.utils.gpytorch_modules import (
     get_covar_module_with_dim_scaled_prior,
     get_gaussian_likelihood_with_lognormal_prior,
@@ -82,40 +83,6 @@
 from torch import Tensor
 
 
-def get_task_value_remapping(task_values: Tensor, dtype: torch.dtype) -> Tensor | None:
-    """Construct an mapping of discrete task values to contiguous int-valued floats.
-
-    Args:
-        task_values: A sorted long-valued tensor of task values.
-        dtype: The dtype of the model inputs (e.g. `X`), which the new
-            task values should have mapped to (e.g. float, double).
-
-    Returns:
-        A tensor of shape `task_values.max() + 1` that maps task values
-        to new task values. The indexing operation `mapper[task_value]`
-        will produce a tensor of new task values, of the same shape as
-        the original. The elements of the `mapper` tensor that do not
-        appear in the original `task_values` are mapped to `nan`. The
-        return value will be `None`, when the task values are contiguous
-        integers starting from zero.
-    """
-    task_range = torch.arange(
-        len(task_values), dtype=task_values.dtype, device=task_values.device
-    )
-    mapper = None
-    if not torch.equal(task_values, task_range):
-        # Create a tensor that maps task values to new task values.
-        # The number of tasks should be small, so this should be quite efficient.
-        mapper = torch.full(
-            (int(task_values.max().item()) + 1,),
-            float("nan"),
-            dtype=dtype,
-            device=task_values.device,
-        )
-        mapper[task_values] = task_range.to(dtype=dtype)
-    return mapper
-
-
 class MultiTaskGP(ExactGP, MultiTaskGPyTorchModel, FantasizeMixin):
     r"""Multi-Task exact GP model using an ICM (intrinsic co-regionalization model)
     kernel. See [Bonilla2007MTGP]_ and [Swersky2013MTBO]_ for a reference on the

botorch/models/transforms/outcome.py

@@ -27,9 +27,11 @@
 
 import torch
 from botorch.models.transforms.utils import (
+    nanstd,
     norm_to_lognorm_mean,
     norm_to_lognorm_variance,
 )
+from botorch.models.utils.assorted import get_task_value_remapping
 from botorch.posteriors import GPyTorchPosterior, Posterior, TransformedPosterior
 from botorch.utils.transforms import normalize_indices
 from linear_operator.operators import CholLinearOperator, DiagLinearOperator
@@ -259,6 +261,46 @@ def __init__(
         self._batch_shape = batch_shape
         self._min_stdv = min_stdv
 
+    def _get_per_input_means_stdvs(
+        self, X: Tensor, include_stdvs_sq: bool
+    ) -> tuple[Tensor, Tensor, Tensor | None]:
+        r"""Get per-input means and stdvs.
+
+        Args:
+            X: A `batch_shape x n x d`-dim tensor of input parameters.
+            include_stdvs_sq: Whether to include the stdvs squared.
+                This parameter is not used by this method
+
+        Returns:
+            A three-tuple with the  means and stdvs:
+
+            - The per-input means.
+            - The per-input stdvs.
+            - The per-input stdvs squared.
+        """
+        return self.means, self.stdvs, self._stdvs_sq
+
+    def _validate_training_inputs(self, Y: Tensor, Yvar: Tensor | None = None) -> None:
+        """Validate training inputs.
+
+        Args:
+            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.
+        """
+        if Y.shape[:-2] != self._batch_shape:
+            raise RuntimeError(
+                f"Expected Y.shape[:-2] to be {self._batch_shape}, matching "
+                f"the `batch_shape` argument to `{self.__class__.__name__}`, but got "
+                f"Y.shape[:-2]={Y.shape[:-2]}."
+            )
+        elif Y.shape[-2] < 1:
+            raise ValueError(f"Can't standardize with no observations. {Y.shape=}.")
+        elif Y.size(-1) != self._m:
+            raise RuntimeError(
+                f"Wrong output dimension. Y.size(-1) is {Y.size(-1)}; expected "
+                f"{self._m}."
+            )
+
     def forward(
         self, Y: Tensor, Yvar: Tensor | None = None, X: Tensor | None = None
     ) -> tuple[Tensor, Tensor | None]:
@@ -283,21 +325,8 @@ def forward(
             - The transformed observation noise (if applicable).
         """
         if self.training:
-            if Y.shape[:-2] != self._batch_shape:
-                raise RuntimeError(
-                    f"Expected Y.shape[:-2] to be {self._batch_shape}, matching "
-                    "the `batch_shape` argument to `Standardize`, but got "
-                    f"Y.shape[:-2]={Y.shape[:-2]}."
-                )
-            if Y.size(-1) != self._m:
-                raise RuntimeError(
-                    f"Wrong output dimension. Y.size(-1) is {Y.size(-1)}; expected "
-                    f"{self._m}."
-                )
-            if Y.shape[-2] < 1:
-                raise ValueError(f"Can't standardize with no observations. {Y.shape=}.")
-
-            elif Y.shape[-2] == 1:
+            self._validate_training_inputs(Y=Y, Yvar=Yvar)
+            if Y.shape[-2] == 1:
                 stdvs = torch.ones(
                     (*Y.shape[:-2], 1, Y.shape[-1]), dtype=Y.dtype, device=Y.device
                 )
@@ -313,9 +342,12 @@ def forward(
             self.stdvs = stdvs
             self._stdvs_sq = stdvs.pow(2)
             self._is_trained = torch.tensor(True)
-
-        Y_tf = (Y - self.means) / self.stdvs
-        Yvar_tf = Yvar / self._stdvs_sq if Yvar is not None else None
+        include_stdvs_sq = Yvar is not None
+        means, stdvs, stdvs_sq = self._get_per_input_means_stdvs(
+            X=X, include_stdvs_sq=include_stdvs_sq
+        )
+        Y_tf = (Y - means) / stdvs
+        Yvar_tf = Yvar / stdvs_sq if include_stdvs_sq else None
         return Y_tf, Yvar_tf
 
     def subset_output(self, idcs: list[int]) -> OutcomeTransform:
@@ -376,9 +408,12 @@ def untransform(
                 "(e.g. `transform(Y)`) before calling `untransform`, since "
                 "means and standard deviations need to be computed."
             )
-
-        Y_utf = self.means + self.stdvs * Y
-        Yvar_utf = self._stdvs_sq * Yvar if Yvar is not None else None
+        include_stdvs_sq = Yvar is not None
+        means, stdvs, stdvs_sq = self._get_per_input_means_stdvs(
+            X=X, include_stdvs_sq=include_stdvs_sq
+        )
+        Y_utf = means + stdvs * Y
+        Yvar_utf = stdvs_sq * Yvar if include_stdvs_sq else None
         return Y_utf, Yvar_utf
 
     @property
@@ -433,8 +468,9 @@ def untransform_posterior(
             )
         # GPyTorchPosterior (TODO: Should we Lazy-evaluate the mean here as well?)
         mvn = posterior.distribution
-        offset = self.means
-        scale_fac = self.stdvs
+        offset, scale_fac, _ = self._get_per_input_means_stdvs(
+            X=X, include_stdvs_sq=False
+        )
         if not posterior._is_mt:
             mean_tf = offset.squeeze(-1) + scale_fac.squeeze(-1) * mvn.mean
             scale_fac = scale_fac.squeeze(-1).expand_as(mean_tf)
@@ -449,7 +485,6 @@ def untransform_posterior(
 
         if (
             not mvn.islazy
-            # TODO: Figure out attribute namming weirdness here
             or mvn._MultivariateNormal__unbroadcasted_scale_tril is not None
         ):
             # if already computed, we can save a lot of time using scale_tril
@@ -465,6 +500,197 @@ def untransform_posterior(
         return GPyTorchPosterior(mvn_tf)
 
 
+class StratifiedStandardize(Standardize):
+    r"""Standardize outcomes (zero mean, unit variance) along stratification dimension.
+
+    This module is stateful: If in train mode, calling forward updates the
+    module state (i.e. the mean/std normalizing constants). If in eval mode,
+    calling forward simply applies the standardization using the current module
+    state.
+    """
+
+    def __init__(
+        self,
+        task_values: Tensor,
+        stratification_idx: int,
+        batch_shape: torch.Size = torch.Size(),  # noqa: B008
+        min_stdv: float = 1e-8,
+        # dtype: torch.dtype = torch.double,
+    ) -> None:
+        r"""Standardize outcomes (zero mean, unit variance) along stratification dim.
+
+        Note: This currenlty only supports single output models
+        (including multi-task models that have a single output).
+
+        Args:
+            task_values: `t`-dim tensor of task values.
+            stratification_idx: The index of the stratification dimension.
+            batch_shape: The batch_shape of the training targets.
+            min_stddv: The minimum standard deviation for which to perform
+                standardization (if lower, only de-mean the data).
+        """
+        OutcomeTransform.__init__(self)
+        self._stratification_idx = stratification_idx
+        task_values = task_values.unique(sorted=True)
+        self.strata_mapping = get_task_value_remapping(task_values, dtype=torch.long)
+        if self.strata_mapping is None:
+            self.strata_mapping = task_values
+        n_strata = self.strata_mapping.shape[0]
+        self._min_stdv = min_stdv
+        self.register_buffer("means", torch.zeros(*batch_shape, n_strata, 1))
+        self.register_buffer("stdvs", torch.ones(*batch_shape, n_strata, 1))
+        self.register_buffer("_stdvs_sq", torch.ones(*batch_shape, n_strata, 1))
+        self.register_buffer("_is_trained", torch.tensor(False))
+        self._batch_shape = batch_shape
+        self._m = 1  # TODO: support multiple outputs
+        self._outputs = None
+
+    def forward(
+        self, Y: Tensor, Yvar: Tensor | None = None, X: Tensor | None = None
+    ) -> tuple[Tensor, Tensor | None]:
+        r"""Standardize outcomes.
+
+        If the module is in train mode, this updates the module state (i.e. the
+        mean/std normalizing constants). If the module is in eval mode, simply
+        applies the normalization using the module state.
+
+        Args:
+            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 input parameters.
+
+        Returns:
+            A two-tuple with the transformed outcomes:
+
+            - The transformed outcome observations.
+            - The transformed observation noise (if applicable).
+        """
+        if X is None:
+            raise ValueError("X is required for StratifiedStandardize.")
+        if self.training:
+            self._validate_training_inputs(Y=Y, Yvar=Yvar)
+            self.means = self.means.to(dtype=X.dtype, device=X.device)
+            self.stdvs = self.stdvs.to(dtype=X.dtype, device=X.device)
+            self._stdvs_sq = self._stdvs_sq.to(dtype=X.dtype, device=X.device)
+            strata = X[..., self._stratification_idx].long()
+            unique_strata = strata.unique()
+            for s in unique_strata:
+                mapped_strata = self.strata_mapping[s]
+                mask = strata != s
+                Y_strata = Y.clone()
+                Y_strata[..., mask, :] = float("nan")
+                stdvs = (
+                    torch.ones_like(Y_strata)
+                    if Y.shape[-2] == 1
+                    else nanstd(X=Y_strata, dim=-2)
+                )
+                stdvs = stdvs.where(
+                    stdvs >= self._min_stdv, torch.full_like(stdvs, 1.0)
+                )
+                means = Y_strata.nanmean(dim=-2)
+                self.means[..., mapped_strata, :] = means
+                self.stdvs[..., mapped_strata, :] = stdvs
+                self._stdvs_sq[..., mapped_strata, :] = stdvs.pow(2)
+            self._is_trained = torch.tensor(True)
+        training = self.training
+        self.training = False
+        tf_Y, tf_Yvar = super().forward(Y=Y, Yvar=Yvar, X=X)
+        self.training = training
+        return tf_Y, tf_Yvar
+
+    def _get_per_input_means_stdvs(
+        self, X: Tensor, include_stdvs_sq: bool
+    ) -> tuple[Tensor, Tensor, Tensor | None]:
+        r"""Get per-input means and stdvs.
+
+        Args:
+            X: A `batch_shape x n x d`-dim tensor of input parameters.
+            include_stdvs_sq: Whether to include the stdvs squared.
+
+        Returns:
+            A three-tuple with the per-input means and stdvs:
+
+            - The per-input means.
+            - The per-input stdvs.
+            - The per-input stdvs squared.
+        """
+        strata = X[..., self._stratification_idx].long()
+        mapped_strata = self.strata_mapping[strata].unsqueeze(-1)
+        # get means and stdvs for each strata
+        n_extra_batch_dims = mapped_strata.ndim - 2 - len(self._batch_shape)
+        expand_shape = mapped_strata.shape[:n_extra_batch_dims] + self.means.shape
+        means = torch.gather(
+            input=self.means.expand(expand_shape),
+            dim=-2,
+            index=mapped_strata,
+        )
+        stdvs = torch.gather(
+            input=self.stdvs.expand(expand_shape),
+            dim=-2,
+            index=mapped_strata,
+        )
+        if include_stdvs_sq:
+            stdvs_sq = torch.gather(
+                input=self._stdvs_sq.expand(expand_shape),
+                dim=-2,
+                index=mapped_strata,
+            )
+        else:
+            stdvs_sq = None
+        return means, stdvs, stdvs_sq
+
+    def subset_output(self, idcs: list[int]) -> OutcomeTransform:
+        r"""Subset the transform along the output dimension.
+
+        Args:
+            idcs: The output indices to subset the transform to.
+
+        Returns:
+            The current outcome transform, subset to the specified output indices.
+        """
+        raise NotImplementedError
+
+    def untransform(
+        self, Y: Tensor, Yvar: Tensor | None = None, X: Tensor | None = None
+    ) -> tuple[Tensor, Tensor | None]:
+        r"""Un-standardize outcomes.
+
+        Args:
+            Y: A `batch_shape x n x m`-dim tensor of standardized targets.
+            Yvar: A `batch_shape x n x m`-dim tensor of standardized observation
+                noises associated with the targets (if applicable).
+            X: A `batch_shape x n x d`-dim tensor of input parameters.
+
+        Returns:
+            A two-tuple with the un-standardized outcomes:
+
+            - The un-standardized outcome observations.
+            - The un-standardized observation noise (if applicable).
+        """
+        if X is None:
+            raise ValueError("X is required for StratifiedStandardize.")
+        return super().untransform(Y=Y, Yvar=Yvar, X=X)
+
+    def untransform_posterior(
+        self, posterior: Posterior, X: Tensor | None = None
+    ) -> GPyTorchPosterior | TransformedPosterior:
+        r"""Un-standardize the posterior.
+
+        Args:
+            posterior: A posterior in the standardized space.
+            X: A `batch_shape x n x d`-dim tensor of training inputs (if applicable).
+
+        Returns:
+            The un-standardized posterior. If the input posterior is a
+            `GPyTorchPosterior`, return a `GPyTorchPosterior`. Otherwise, return a
+            `TransformedPosterior`.
+        """
+        if X is None:
+            raise ValueError("X is required for StratifiedStandardize.")
+        return super().untransform_posterior(posterior=posterior, X=X)
+
+
 class Log(OutcomeTransform):
     r"""Log-transform outcomes.