Update use cases of deprecated fixed-noise models (#2059)

Summary: Pull Request resolved: #2059 Updates use cases of fixed noise models that were deprecated in #2052 & elsewhere. I'll update Ax usage in a separate diff. Reviewed By: esantorella Differential Revision: D50429032 fbshipit-source-id: 990536e54077dcc9d752c583a2f5456ba75999d5
pytorch · Oct 19, 2023 · 7658c2d · 7658c2d
1 parent 5302f19
commit 7658c2d
Show file tree

Hide file tree

Showing 25 changed files with 356 additions and 356 deletions.
diff --git a/botorch/acquisition/analytic.py b/botorch/acquisition/analytic.py
@@ -24,7 +24,7 @@
 from botorch.acquisition.acquisition import AcquisitionFunction
 from botorch.acquisition.objective import PosteriorTransform
 from botorch.exceptions import UnsupportedError
-from botorch.models.gp_regression import FixedNoiseGP
+from botorch.models.gp_regression import SingleTaskGP
 from botorch.models.gpytorch import GPyTorchModel
 from botorch.models.model import Model
 from botorch.utils.constants import get_constants_like
@@ -38,6 +38,7 @@
 )
 from botorch.utils.safe_math import log1mexp, logmeanexp
 from botorch.utils.transforms import convert_to_target_pre_hook, t_batch_mode_transform
+from gpytorch.likelihoods.gaussian_likelihood import FixedNoiseGaussianLikelihood
 from torch import Tensor
 from torch.nn.functional import pad
 
@@ -580,11 +581,12 @@ class LogNoisyExpectedImprovement(AnalyticAcquisitionFunction):
 
     where `X_base` are previously observed points.
 
-    Note: This acquisition function currently relies on using a FixedNoiseGP (required
-    for noiseless fantasies).
+    Note: This acquisition function currently relies on using a SingleTaskGP
+    with known observation noise. In other words, `train_Yvar` must be passed
+    to the model. (required for noiseless fantasies).
 
     Example:
-        >>> model = FixedNoiseGP(train_X, train_Y, train_Yvar=train_Yvar)
+        >>> model = SingleTaskGP(train_X, train_Y, train_Yvar=train_Yvar)
         >>> LogNEI = LogNoisyExpectedImprovement(model, train_X)
         >>> nei = LogNEI(test_X)
     """
@@ -608,10 +610,6 @@ def __init__(
                 complexity and performance).
             maximize: If True, consider the problem a maximization problem.
         """
-        if not isinstance(model, FixedNoiseGP):
-            raise UnsupportedError(
-                "Only FixedNoiseGPs are currently supported for fantasy LogNEI"
-            )
         # sample fantasies
         from botorch.sampling.normal import SobolQMCNormalSampler
 
@@ -662,11 +660,12 @@ class NoisyExpectedImprovement(ExpectedImprovement):
     `NEI(x) = E(max(y - max Y_baseline), 0)), (y, Y_baseline) ~ f((x, X_baseline))`,
     where `X_baseline` are previously observed points.
 
-    Note: This acquisition function currently relies on using a FixedNoiseGP (required
-    for noiseless fantasies).
+    Note: This acquisition function currently relies on using a SingleTaskGP
+    with known observation noise. In other words, `train_Yvar` must be passed
+    to the model. (required for noiseless fantasies).
 
     Example:
-        >>> model = FixedNoiseGP(train_X, train_Y, train_Yvar=train_Yvar)
+        >>> model = SingleTaskGP(train_X, train_Y, train_Yvar=train_Yvar)
         >>> NEI = NoisyExpectedImprovement(model, train_X)
         >>> nei = NEI(test_X)
     """
@@ -689,10 +688,6 @@ def __init__(
                 complexity and performance).
             maximize: If True, consider the problem a maximization problem.
         """
-        if not isinstance(model, FixedNoiseGP):
-            raise UnsupportedError(
-                "Only FixedNoiseGPs are currently supported for fantasy NEI"
-            )
         # sample fantasies
         from botorch.sampling.normal import SobolQMCNormalSampler
 
@@ -974,15 +969,15 @@ def logerfcx(x: Tensor) -> Tensor:
 
 
 def _get_noiseless_fantasy_model(
-    model: FixedNoiseGP, batch_X_observed: Tensor, Y_fantasized: Tensor
-) -> FixedNoiseGP:
+    model: SingleTaskGP, batch_X_observed: Tensor, Y_fantasized: Tensor
+) -> SingleTaskGP:
     r"""Construct a fantasy model from a fitted model and provided fantasies.
 
     The fantasy model uses the hyperparameters from the original fitted model and
     assumes the fantasies are noiseless.
 
     Args:
-        model: a fitted FixedNoiseGP
+        model: A fitted SingleTaskGP with known observation noise.
         batch_X_observed: A `b x n x d` tensor of inputs where `b` is the number of
             fantasies.
         Y_fantasized: A `b x n` tensor of fantasized targets where `b` is the number of
@@ -991,11 +986,18 @@ def _get_noiseless_fantasy_model(
     Returns:
         The fantasy model.
     """
-    # initialize a copy of FixedNoiseGP on the original training inputs
-    # this makes FixedNoiseGP a non-batch GP, so that the same hyperparameters
+    if not isinstance(model, SingleTaskGP) or not isinstance(
+        model.likelihood, FixedNoiseGaussianLikelihood
+    ):
+        raise UnsupportedError(
+            "Only SingleTaskGP models with known observation noise "
+            "are currently supported for fantasy-based NEI & LogNEI."
+        )
+    # initialize a copy of SingleTaskGP on the original training inputs
+    # this makes SingleTaskGP a non-batch GP, so that the same hyperparameters
     # are used across all batches (by default, a GP with batched training data
     # uses independent hyperparameters for each batch).
-    fantasy_model = FixedNoiseGP(
+    fantasy_model = SingleTaskGP(
         train_X=model.train_inputs[0],
         train_Y=model.train_targets.unsqueeze(-1),
         train_Yvar=model.likelihood.noise_covar.noise.unsqueeze(-1),

diff --git a/botorch/models/contextual.py b/botorch/models/contextual.py
@@ -6,28 +6,29 @@
 
 from typing import Any, Dict, List, Optional
 
-from botorch.models.gp_regression import FixedNoiseGP
+from botorch.models.gp_regression import SingleTaskGP
 from botorch.models.kernels.contextual_lcea import LCEAKernel
 from botorch.models.kernels.contextual_sac import SACKernel
 from botorch.utils.datasets import SupervisedDataset
 from torch import Tensor
 
 
-class SACGP(FixedNoiseGP):
+class SACGP(SingleTaskGP):
     r"""A GP using a Structural Additive Contextual(SAC) kernel."""
 
     def __init__(
         self,
         train_X: Tensor,
         train_Y: Tensor,
-        train_Yvar: Tensor,
+        train_Yvar: Optional[Tensor],
         decomposition: Dict[str, List[int]],
     ) -> None:
         r"""
         Args:
             train_X: (n x d) X training data.
             train_Y: (n x 1) Y training data.
-            train_Yvar: (n x 1) Noise variances of each training Y.
+            train_Yvar: (n x 1) Noise variances of each training Y. If None,
+                we use an inferred noise likelihood.
             decomposition: Keys are context names. Values are the indexes of
                 parameters belong to the context. The parameter indexes are in
                 the same order across contexts.
@@ -62,7 +63,7 @@ def construct_inputs(
         }
 
 
-class LCEAGP(FixedNoiseGP):
+class LCEAGP(SingleTaskGP):
     r"""A GP using a Latent Context Embedding Additive (LCE-A) Kernel.
 
     Note that the model does not support batch training. Input training
@@ -73,7 +74,7 @@ def __init__(
         self,
         train_X: Tensor,
         train_Y: Tensor,
-        train_Yvar: Tensor,
+        train_Yvar: Optional[Tensor],
         decomposition: Dict[str, List[int]],
         train_embedding: bool = True,
         cat_feature_dict: Optional[Dict] = None,
@@ -85,7 +86,8 @@ def __init__(
         Args:
             train_X: (n x d) X training data.
             train_Y: (n x 1) Y training data.
-            train_Yvar: (n x 1) Noise variance of Y.
+            train_Yvar: (n x 1) Noise variance of Y. If None,
+                we use an inferred noise likelihood.
             decomposition: Keys are context names. Values are the indexes of
                 parameters belong to the context.
             train_embedding: Whether to train the embedding layer or not. If False,

diff --git a/botorch/models/contextual_multioutput.py b/botorch/models/contextual_multioutput.py
@@ -170,6 +170,7 @@ class FixedNoiseLCEMGP(LCEMGP):
     (LCE-M) kernel, with known observation noise.
 
     DEPRECATED: Please use `LCEMGP` with `train_Yvar` instead.
+    Will be removed in a future release (~v0.11).
     """
 
     def __init__(

diff --git a/botorch/models/converter.py b/botorch/models/converter.py
@@ -15,13 +15,14 @@
 
 import torch
 from botorch.exceptions import UnsupportedError
-from botorch.models.gp_regression import FixedNoiseGP, HeteroskedasticSingleTaskGP
+from botorch.models.gp_regression import HeteroskedasticSingleTaskGP
 from botorch.models.gp_regression_fidelity import SingleTaskMultiFidelityGP
 from botorch.models.gp_regression_mixed import MixedSingleTaskGP
 from botorch.models.gpytorch import BatchedMultiOutputGPyTorchModel
 from botorch.models.model_list_gp_regression import ModelListGP
 from botorch.models.transforms.input import InputTransform
 from botorch.models.transforms.outcome import OutcomeTransform
+from gpytorch.likelihoods.gaussian_likelihood import FixedNoiseGaussianLikelihood
 from torch import Tensor
 from torch.nn import Module
 
@@ -140,7 +141,7 @@ def model_list_to_batched(model_list: ModelListGP) -> BatchedMultiOutputGPyTorch
     train_X = deepcopy(models[0].train_inputs[0])
     train_Y = torch.stack([m.train_targets.clone() for m in models], dim=-1)
     kwargs = {"train_X": train_X, "train_Y": train_Y}
-    if isinstance(models[0], FixedNoiseGP):
+    if isinstance(models[0].likelihood, FixedNoiseGaussianLikelihood):
         kwargs["train_Yvar"] = torch.stack(
             [m.likelihood.noise_covar.noise.clone() for m in models], dim=-1
         )
@@ -302,7 +303,7 @@ def batched_to_model_list(batch_model: BatchedMultiOutputGPyTorchModel) -> Model
             .clone()
             .unsqueeze(-1),
         }
-        if isinstance(batch_model, FixedNoiseGP):
+        if isinstance(batch_model.likelihood, FixedNoiseGaussianLikelihood):
             noise_covar = batch_model.likelihood.noise_covar
             kwargs["train_Yvar"] = (
                 noise_covar.noise.select(input_bdims, i).clone().unsqueeze(-1)
@@ -390,7 +391,7 @@ def batched_multi_output_to_single_output(
         "train_X": batch_mo_model.train_inputs[0].clone(),
         "train_Y": batch_mo_model.train_targets.clone().unsqueeze(-1),
     }
-    if isinstance(batch_mo_model, FixedNoiseGP):
+    if isinstance(batch_mo_model.likelihood, FixedNoiseGaussianLikelihood):
         noise_covar = batch_mo_model.likelihood.noise_covar
         kwargs["train_Yvar"] = noise_covar.noise.clone().unsqueeze(-1)
     if isinstance(batch_mo_model, SingleTaskMultiFidelityGP):

diff --git a/botorch/models/fully_bayesian_multitask.py b/botorch/models/fully_bayesian_multitask.py
@@ -182,7 +182,7 @@ class SaasFullyBayesianMultiTaskGP(MultiTaskGP):
         >>> ])
         >>> train_Y = torch.cat(f1(X1), f2(X2)).unsqueeze(-1)
         >>> train_Yvar = 0.01 * torch.ones_like(train_Y)
-        >>> mtsaas_gp = SaasFullyBayesianFixedNoiseMultiTaskGP(
+        >>> mtsaas_gp = SaasFullyBayesianMultiTaskGP(
         >>>     train_X, train_Y, train_Yvar, task_feature=-1,
         >>> )
         >>> fit_fully_bayesian_model_nuts(mtsaas_gp)

diff --git a/botorch/models/gp_regression.py b/botorch/models/gp_regression.py
@@ -202,6 +202,7 @@ class FixedNoiseGP(SingleTaskGP):
 
     DEPRECATED: `FixedNoiseGP` has been merged into `SingleTaskGP`. Please use
     `SingleTaskGP` with `train_Yvar` instead.
+    Will be removed in a future release (~v0.12).
     """
 
     def __init__(

diff --git a/botorch/models/gp_regression_fidelity.py b/botorch/models/gp_regression_fidelity.py
@@ -192,7 +192,9 @@ def __init__(
         outcome_transform: Optional[OutcomeTransform] = None,
         input_transform: Optional[InputTransform] = None,
     ) -> None:
-        r"""DEPRECATED: Use `SingleTaskMultiFidelityGP` instead."""
+        r"""DEPRECATED: Use `SingleTaskMultiFidelityGP` instead.
+        Will be removed in a future release (~v0.11).
+        """
         warnings.warn(
             "`FixedNoiseMultiFidelityGP` has been deprecated. "
             "Use `SingleTaskMultiFidelityGP` with `train_Yvar` instead.",

diff --git a/botorch/models/model.py b/botorch/models/model.py
@@ -373,9 +373,10 @@ def fantasize(
                 + self.batch_shape
                 + torch.Size([0, self.num_outputs])
             )
+            Y = torch.empty(output_shape, dtype=X.dtype, device=X.device)
             return self.condition_on_observations(
                 X=self.transform_inputs(X),
-                Y=torch.empty(output_shape, dtype=X.dtype, device=X.device),
+                Y=Y,
                 **kwargs,
             )
         propagate_grads = kwargs.pop("propagate_grads", False)

diff --git a/botorch/models/multitask.py b/botorch/models/multitask.py
@@ -327,6 +327,7 @@ class FixedNoiseMultiTaskGP(MultiTaskGP):
     r"""Multi-Task GP model using an ICM kernel, with known observation noise.
 
     DEPRECATED: Please use `MultiTaskGP` with `train_Yvar` instead.
+    Will be removed in a future release (~v0.10).
     """
 
     def __init__(

diff --git a/botorch/models/utils/parse_training_data.py b/botorch/models/utils/parse_training_data.py
@@ -14,7 +14,7 @@
 import torch
 from botorch.exceptions import UnsupportedError
 from botorch.models.model import Model
-from botorch.models.multitask import FixedNoiseMultiTaskGP, MultiTaskGP
+from botorch.models.multitask import MultiTaskGP
 from botorch.models.pairwise_gp import PairwiseGP
 from botorch.utils.datasets import RankingDataset, SupervisedDataset
 from botorch.utils.dispatcher import Dispatcher
@@ -88,7 +88,7 @@ def _parse_model_dict(
     return dispatcher(consumer, next(iter(training_data.values())))
 
 
-@dispatcher.register((MultiTaskGP, FixedNoiseMultiTaskGP), dict)
+@dispatcher.register(MultiTaskGP, dict)
 def _parse_multitask_dict(
     consumer: Model,
     training_data: Dict[Hashable, SupervisedDataset],

diff --git a/botorch/utils/datasets.py b/botorch/utils/datasets.py
@@ -154,6 +154,7 @@ class FixedNoiseDataset(SupervisedDataset):
     observations variances so that `Y[i] ~ N(f(X[i]), Yvar[i])`.
 
     NOTE: This is deprecated. Use `SupervisedDataset` instead.
+    Will be removed in a future release (~v0.11).
     """
 
     def __init__(