Update the default SingleTaskGP prior (pytorch#2449)

Summary:
X-link: facebook/Ax#2610

Pull Request resolved: pytorch#2449

Update of the default hyperparameter priors for the SingleTaskGP.

Switch from the conventional Scale-Matern kernel with Gamma(3, 6) lengthscale prior is substituted for an RBF Kernel (without a ScaleKernel), and a change from the high-noise Gamma(1.1, 0.05) noise prior of the GaussianLikelihood to a LogNormal prior that prefers lower values. The change is made in accordance with the findings of [1].

 The change is made to improve the out-of-the-box performance of the BoTorch models on high-dimensional problems.

[1] Carl Hvarfner, Erik Orm Hellsten, Luigi Nardi. _Vanilla Bayesian Optimization Performs Great in High Dimensions_. ICML, 2024.

Reviewed By: saitcakmak

Differential Revision: D60080819

botorch/models/gp_regression.py

@@ -40,8 +40,8 @@
 from botorch.models.transforms.outcome import Log, OutcomeTransform
 from botorch.models.utils import validate_input_scaling
 from botorch.models.utils.gpytorch_modules import (
-    get_gaussian_likelihood_with_gamma_prior,
-    get_matern_kernel_with_gamma_prior,
+    get_covar_module_with_dim_scaled_prior,
+    get_gaussian_likelihood_with_lognormal_prior,
     MIN_INFERRED_NOISE_LEVEL,
 )
 from botorch.utils.containers import BotorchContainer
@@ -67,9 +67,13 @@
 class SingleTaskGP(BatchedMultiOutputGPyTorchModel, ExactGP, FantasizeMixin):
     r"""A single-task exact GP model, supporting both known and inferred noise levels.
 
-    A single-task exact GP using relatively strong priors on the Kernel
-    hyperparameters, which work best when covariates are normalized to the unit
-    cube and outcomes are standardized (zero mean, unit variance).
+    A single-task exact GP which, by default, utilizes hyperparameter priors
+    from [Hvarfner2024vanilla]_. These priors are fairly weak and designed to be
+    agnostic to the dimensionality of the problem. Moreover, they suggest a
+    moderately low level of noise. Importantly, The model works best when covariates
+    are normalized to the unit cube and outcomes are standardized (zero mean, unit
+    variance). For a detailed discussion on the hyperparameter priors, see
+    https://github.com/pytorch/botorch/discussions/2451.
 
     This model works in batch mode (each batch having its own hyperparameters).
     When the training observations include multiple outputs, this model will use
@@ -174,7 +178,7 @@ def __init__(
         )
         if likelihood is None:
             if train_Yvar is None:
-                likelihood = get_gaussian_likelihood_with_gamma_prior(
+                likelihood = get_gaussian_likelihood_with_lognormal_prior(
                     batch_shape=self._aug_batch_shape
                 )
             else:
@@ -190,14 +194,13 @@ def __init__(
             mean_module = ConstantMean(batch_shape=self._aug_batch_shape)
         self.mean_module = mean_module
         if covar_module is None:
-            covar_module = get_matern_kernel_with_gamma_prior(
+            covar_module = get_covar_module_with_dim_scaled_prior(
                 ard_num_dims=transformed_X.shape[-1],
                 batch_shape=self._aug_batch_shape,
             )
             self._subset_batch_dict = {
                 "mean_module.raw_constant": -1,
-                "covar_module.raw_outputscale": -1,
-                "covar_module.base_kernel.raw_lengthscale": -3,
+                "covar_module.raw_lengthscale": -3,
             }
             if train_Yvar is None:
                 self._subset_batch_dict["likelihood.noise_covar.raw_noise"] = -2

botorch/utils/gp_sampling.py

@@ -143,10 +143,9 @@ def __init__(
         """
         if not isinstance(kernel, ScaleKernel):
             base_kernel = kernel
-            outputscale = torch.tensor(
-                1.0,
-                dtype=base_kernel.lengthscale.dtype,
-                device=base_kernel.lengthscale.device,
+            outputscale = torch.ones(kernel.batch_shape).to(
+                dtype=kernel.lengthscale.dtype,
+                device=kernel.lengthscale.device,
             )
         else:
             base_kernel = kernel.base_kernel

botorch_community/models/gp_regression_multisource.py

@@ -20,6 +20,7 @@
 
 from __future__ import annotations
 
+from copy import deepcopy
 from typing import Optional
 
 import torch
@@ -231,9 +232,9 @@ def _init_fit_gp(
             train_X,
             train_Y,
             train_Yvar,
-            likelihood=likelihood,
-            covar_module=covar_module,
-            mean_module=mean_module,
+            likelihood=deepcopy(likelihood),
+            covar_module=deepcopy(covar_module),
+            mean_module=deepcopy(mean_module),
             outcome_transform=outcome_transform,
             input_transform=input_transform,
         )

test/models/test_converter.py

@@ -21,9 +21,10 @@
 )
 from botorch.models.transforms.input import AppendFeatures, Normalize
 from botorch.models.transforms.outcome import Standardize
+from botorch.models.utils.gpytorch_modules import get_matern_kernel_with_gamma_prior
 from botorch.utils.test_helpers import SimpleGPyTorchModel
 from botorch.utils.testing import BotorchTestCase
-from gpytorch.kernels import RBFKernel
+from gpytorch.kernels import MaternKernel, RBFKernel
 from gpytorch.likelihoods import GaussianLikelihood
 from gpytorch.likelihoods.gaussian_likelihood import FixedNoiseGaussianLikelihood
 from gpytorch.priors import LogNormalPrior
@@ -133,14 +134,19 @@ def test_model_list_to_batched(self):
             with self.assertRaises(UnsupportedError):
                 model_list_to_batched(ModelListGP(gp1, gp2))
             # check scalar agreement
-            gp2 = SingleTaskGP(train_X, train_Y2)
-            gp2.likelihood.noise_covar.noise_prior.rate.fill_(1.0)
+            # modified to check the scalar agreement in a parameter that is accessible
+            # since the error is going to slip through for the non-parametrizable
+            # priors regardless (like the LogNormal)
+            with self.assertRaises(UnsupportedError):
+                model_list_to_batched(ModelListGP(gp1, gp2))
+
+            gp2.likelihood.noise_covar.raw_noise_constraint.lower_bound.fill_(1e-3)
             with self.assertRaises(UnsupportedError):
                 model_list_to_batched(ModelListGP(gp1, gp2))
             # check tensor shape agreement
             gp2 = SingleTaskGP(train_X, train_Y2)
-            gp2.covar_module.raw_outputscale = torch.nn.Parameter(
-                torch.tensor([0.0], device=self.device, dtype=dtype)
+            gp2.likelihood.noise_covar.raw_noise = torch.nn.Parameter(
+                torch.tensor([[0.42]], device=self.device, dtype=dtype)
             )
             with self.assertRaises(UnsupportedError):
                 model_list_to_batched(ModelListGP(gp1, gp2))
@@ -155,14 +161,15 @@ def test_model_list_to_batched(self):
             with self.assertRaises(NotImplementedError):
                 model_list_to_batched(ModelListGP(gp2))
             # test non-default kernel
-            gp1 = SingleTaskGP(train_X, train_Y1, covar_module=RBFKernel())
-            gp2 = SingleTaskGP(train_X, train_Y2, covar_module=RBFKernel())
+            gp1 = SingleTaskGP(train_X, train_Y1, covar_module=MaternKernel())
+            gp2 = SingleTaskGP(train_X, train_Y2, covar_module=MaternKernel())
             list_gp = ModelListGP(gp1, gp2)
             batch_gp = model_list_to_batched(list_gp)
-            self.assertEqual(type(batch_gp.covar_module), RBFKernel)
+            self.assertEqual(type(batch_gp.covar_module), MaternKernel)
             # test error when component GPs have different kernel types
-            gp1 = SingleTaskGP(train_X, train_Y1, covar_module=RBFKernel())
-            gp2 = SingleTaskGP(train_X, train_Y2)
+            # added types for both default and non-default kernels for clarity
+            gp1 = SingleTaskGP(train_X, train_Y1, covar_module=MaternKernel())
+            gp2 = SingleTaskGP(train_X, train_Y2, covar_module=RBFKernel())
             list_gp = ModelListGP(gp1, gp2)
             with self.assertRaises(UnsupportedError):
                 model_list_to_batched(list_gp)
@@ -244,6 +251,23 @@ def test_model_list_to_batched(self):
             with self.assertRaises(UnsupportedError):
                 model_list_to_batched(list_gp)
 
+    def test_model_list_to_batched_with_legacy_prior(self) -> None:
+        train_X = torch.rand(10, 2, device=self.device, dtype=torch.double)
+        # and test the old prior for completeness & test coverage
+        gp1_gamma = SingleTaskGP(
+            train_X,
+            train_X.sum(dim=-1, keepdim=True),
+            covar_module=get_matern_kernel_with_gamma_prior(train_X.shape[-1]),
+        )
+        gp2_gamma = SingleTaskGP(
+            train_X,
+            train_X.sum(dim=-1, keepdim=True),
+            covar_module=get_matern_kernel_with_gamma_prior(train_X.shape[-1]),
+        )
+        gp1_gamma.covar_module.base_kernel.lengthscale_prior.rate.fill_(1.0)
+        with self.assertRaises(UnsupportedError):
+            model_list_to_batched(ModelListGP(gp1_gamma, gp2_gamma))
+
     def test_model_list_to_batched_with_different_prior(self) -> None:
         # The goal is to test priors that didn't have their parameters
         # recorded in the state dict prior to GPyTorch #2551.

test/models/test_deterministic.py

@@ -172,7 +172,7 @@ def test_FixedSingleSampleModel(self):
         post = model.posterior(test_X)
         original_output = post.mean + post.variance.sqrt() * w
         fss_output = fss_model(test_X)
-        self.assertTrue(torch.equal(original_output, fss_output))
+        self.assertAllClose(original_output, fss_output)
 
         self.assertTrue(hasattr(fss_model, "num_outputs"))
 

test/models/test_gp_regression.py

@@ -23,7 +23,7 @@
 from botorch.utils.sampling import manual_seed
 from botorch.utils.test_helpers import get_pvar_expected
 from botorch.utils.testing import _get_random_data, BotorchTestCase
-from gpytorch.kernels import MaternKernel, RBFKernel, ScaleKernel
+from gpytorch.kernels import RBFKernel
 from gpytorch.likelihoods import (
     _GaussianLikelihoodBase,
     FixedNoiseGaussianLikelihood,
@@ -33,7 +33,7 @@
 from gpytorch.means import ConstantMean, ZeroMean
 from gpytorch.mlls.exact_marginal_log_likelihood import ExactMarginalLogLikelihood
 from gpytorch.mlls.noise_model_added_loss_term import NoiseModelAddedLossTerm
-from gpytorch.priors import GammaPrior
+from gpytorch.priors import LogNormalPrior
 
 
 class TestGPRegressionBase(BotorchTestCase):
@@ -96,10 +96,10 @@ def test_gp(self, double_only: bool = False):
 
             # test init
             self.assertIsInstance(model.mean_module, ConstantMean)
-            self.assertIsInstance(model.covar_module, ScaleKernel)
-            matern_kernel = model.covar_module.base_kernel
-            self.assertIsInstance(matern_kernel, MaternKernel)
-            self.assertIsInstance(matern_kernel.lengthscale_prior, GammaPrior)
+            self.assertIsInstance(model.covar_module, RBFKernel)
+            rbf_kernel = model.covar_module
+            self.assertIsInstance(rbf_kernel, RBFKernel)
+            self.assertIsInstance(rbf_kernel.lengthscale_prior, LogNormalPrior)
             if use_octf:
                 self.assertIsInstance(model.outcome_transform, Standardize)
             if use_intf:

test/models/test_model_list_gp_regression.py

@@ -25,7 +25,7 @@
 from botorch.sampling.normal import IIDNormalSampler
 from botorch.utils.testing import _get_random_data, BotorchTestCase
 from gpytorch.distributions import MultitaskMultivariateNormal, MultivariateNormal
-from gpytorch.kernels import MaternKernel, ScaleKernel
+from gpytorch.kernels import RBFKernel
 from gpytorch.likelihoods import LikelihoodList
 from gpytorch.likelihoods.gaussian_likelihood import (
     FixedNoiseGaussianLikelihood,
@@ -34,7 +34,7 @@
 from gpytorch.means import ConstantMean
 from gpytorch.mlls import SumMarginalLogLikelihood
 from gpytorch.mlls.exact_marginal_log_likelihood import ExactMarginalLogLikelihood
-from gpytorch.priors import GammaPrior
+from gpytorch.priors import LogNormalPrior
 from torch import Tensor
 
 
@@ -104,10 +104,8 @@ def _base_test_ModelListGP(
         self.assertEqual(model.num_outputs, 2)
         for m in model.models:
             self.assertIsInstance(m.mean_module, ConstantMean)
-            self.assertIsInstance(m.covar_module, ScaleKernel)
-            matern_kernel = m.covar_module.base_kernel
-            self.assertIsInstance(matern_kernel, MaternKernel)
-            self.assertIsInstance(matern_kernel.lengthscale_prior, GammaPrior)
+            self.assertIsInstance(m.covar_module, RBFKernel)
+            self.assertIsInstance(m.covar_module.lengthscale_prior, LogNormalPrior)
             if outcome_transform != "None":
                 self.assertIsInstance(
                     m.outcome_transform, (Log, Standardize, ChainedOutcomeTransform)

test/optim/test_fit.py

@@ -29,9 +29,9 @@ def setUp(self) -> None:
         self.mlls = {}
         with torch.random.fork_rng():
             torch.manual_seed(0)
-            train_X = torch.linspace(0, 1, 10).unsqueeze(-1)
-            train_Y = torch.sin((2 * math.pi) * train_X)
-            train_Y = train_Y + 0.1 * torch.randn_like(train_Y)
+            train_X = torch.linspace(0, 1, 30).unsqueeze(-1)
+            train_Y = torch.sin((6 * math.pi) * train_X)
+            train_Y = train_Y + 0.01 * torch.randn_like(train_Y)
 
         model = SingleTaskGP(
             train_X=train_X,

test/optim/utils/test_model_utils.py

@@ -6,6 +6,7 @@
 
 from __future__ import annotations
 
+import itertools
 import re
 import warnings
 from copy import deepcopy
@@ -16,6 +17,10 @@
 import torch
 from botorch import settings
 from botorch.models import SingleTaskGP
+from botorch.models.utils.gpytorch_modules import (
+    get_covar_module_with_dim_scaled_prior,
+    get_matern_kernel_with_gamma_prior,
+)
 from botorch.optim.utils import (
     get_data_loader,
     get_name_filter,
@@ -158,10 +163,18 @@ def test__get_name_filter(self) -> None:
 
 class TestSampleAllPriors(BotorchTestCase):
     def test_sample_all_priors(self):
-        for dtype in (torch.float, torch.double):
+        for dtype, covar_module in itertools.product(
+            (torch.float, torch.double),
+            (
+                get_covar_module_with_dim_scaled_prior(ard_num_dims=5),
+                get_matern_kernel_with_gamma_prior(ard_num_dims=5),
+            ),
+        ):
             train_X = torch.rand(3, 5, device=self.device, dtype=dtype)
             train_Y = torch.rand(3, 1, device=self.device, dtype=dtype)
-            model = SingleTaskGP(train_X=train_X, train_Y=train_Y)
+            model = SingleTaskGP(
+                train_X=train_X, train_Y=train_Y, covar_module=covar_module
+            )
             mll = ExactMarginalLogLikelihood(model.likelihood, model)
             mll.to(device=self.device, dtype=dtype)
             original_state_dict = dict(deepcopy(mll.model.state_dict()))
@@ -173,7 +186,10 @@ def test_sample_all_priors(self):
                 != original_state_dict["likelihood.noise_covar.raw_noise"]
             )
             # check that lengthscales are all different
-            ls = model.covar_module.base_kernel.raw_lengthscale.view(-1).tolist()
+            if isinstance(model.covar_module, ScaleKernel):
+                ls = model.covar_module.base_kernel.raw_lengthscale.view(-1).tolist()
+            else:
+                ls = model.covar_module.raw_lengthscale.view(-1).tolist()
             self.assertTrue(all(ls[0] != ls[i]) for i in range(1, len(ls)))
 
             # change one of the priors to a dummy prior that does not support sampling

test/test_fit.py

@@ -27,7 +27,7 @@
 from botorch.settings import debug
 from botorch.utils.context_managers import module_rollback_ctx, TensorCheckpoint
 from botorch.utils.testing import BotorchTestCase
-from gpytorch.kernels import MaternKernel
+from gpytorch.kernels import RBFKernel
 from gpytorch.mlls import ExactMarginalLogLikelihood, VariationalELBO
 from linear_operator.utils.errors import NotPSDError
 
@@ -136,8 +136,7 @@ def setUp(self, suppress_input_warnings: bool = True) -> None:
                     input_transform=Normalize(d=1),
                     outcome_transform=Standardize(m=output_dim),
                 )
-                self.assertIsInstance(model.covar_module.base_kernel, MaternKernel)
-                model.covar_module.base_kernel.nu = 2.5
+                self.assertIsInstance(model.covar_module, RBFKernel)
 
                 mll = ExactMarginalLogLikelihood(model.likelihood, model)
                 for dtype in (torch.float32, torch.float64):

test_community/acquisition/test_multi_source.py

@@ -6,8 +6,11 @@
 
 import torch
 from botorch.exceptions import UnsupportedError
+from botorch.models.utils.gpytorch_modules import (
+    get_gaussian_likelihood_with_gamma_prior,
+    get_matern_kernel_with_gamma_prior,
+)
 from botorch.utils.testing import BotorchTestCase, MockModel, MockPosterior
-
 from botorch_community.acquisition.augmented_multisource import (
     AugmentedUpperConfidenceBound,
 )
@@ -21,11 +24,17 @@ def _get_mock_agp(self, batch_shape, dtype):
         rep_shape = batch_shape + torch.Size([1, 1])
         train_X = train_X.repeat(rep_shape)
         train_Y = train_Y.repeat(rep_shape)
+        covar_module = get_matern_kernel_with_gamma_prior(
+            ard_num_dims=train_X.shape[-1] - 1,
+        )
         model_kwargs = {
             "train_X": train_X,
             "train_Y": train_Y,
+            "covar_module": covar_module,
+            "likelihood": get_gaussian_likelihood_with_gamma_prior(),
         }
         model = SingleTaskAugmentedGP(**model_kwargs)
+
         return model
 
     def test_upper_confidence_bound(self):