Do changes

Author: Scienfitz

baybe/surrogates/gaussian_process/core.py

@@ -40,11 +40,11 @@
 from baybe.surrogates.gaussian_process.presets import (
     GaussianProcessPreset,
 )
-from baybe.surrogates.gaussian_process.presets.baybe import (
-    BayBEFitCriterionFactory,
-    BayBEKernelFactory,
-    BayBELikelihoodFactory,
-    BayBEMeanFactory,
+from baybe.surrogates.gaussian_process.presets.hvarfner import (
+    FIT_CRITERION_FACTORY,
+    KERNEL_FACTORY,
+    LIKELIHOOD_FACTORY,
+    MEAN_FACTORY,
 )
 from baybe.utils.boolean import strtobool
 from baybe.utils.conversion import to_string
@@ -115,7 +115,7 @@ def _mark_custom_kernel(
     value: Kernel | KernelFactoryProtocol, self: GaussianProcessSurrogate
 ) -> Kernel | KernelFactoryProtocol:
     """Mark the surrogate as using a custom kernel (for deprecation purposes)."""
-    if type(value) is not BayBEKernelFactory:
+    if type(value) is not type(KERNEL_FACTORY):
         self._custom_kernel = True
 
     return value
@@ -152,7 +152,7 @@ class GaussianProcessSurrogate(Surrogate):
             Converter(_mark_custom_kernel, takes_self=True),  # type: ignore[call-overload]
             partial(to_component_factory, component_type=GPComponentType.KERNEL),
         ),
-        factory=BayBEKernelFactory,
+        default=KERNEL_FACTORY,
         validator=is_callable(),
     )
     """The factory used to create the kernel for the Gaussian process.
@@ -165,7 +165,7 @@ class GaussianProcessSurrogate(Surrogate):
 
     mean_factory: MeanFactoryProtocol = field(
         alias="mean_or_factory",
-        factory=BayBEMeanFactory,
+        default=MEAN_FACTORY,
         converter=partial(to_component_factory, component_type=GPComponentType.MEAN),  # type: ignore[misc]
         validator=is_callable(),
     )
@@ -178,7 +178,7 @@ class GaussianProcessSurrogate(Surrogate):
 
     likelihood_factory: LikelihoodFactoryProtocol = field(
         alias="likelihood_or_factory",
-        factory=BayBELikelihoodFactory,
+        default=LIKELIHOOD_FACTORY,
         converter=partial(  # type: ignore[misc]
             to_component_factory, component_type=GPComponentType.LIKELIHOOD
         ),
@@ -193,7 +193,7 @@ class GaussianProcessSurrogate(Surrogate):
 
     fit_criterion_factory: FitCriterionFactoryProtocol = field(
         alias="fit_criterion_or_factory",
-        factory=BayBEFitCriterionFactory,
+        default=FIT_CRITERION_FACTORY,
         converter=partial(  # type: ignore[misc]
             to_component_factory, component_type=GPComponentType.CRITERION
         ),

baybe/surrogates/gaussian_process/presets/hvarfner.py

@@ -47,10 +47,14 @@ class HvarfnerKernelFactory(_PureKernelFactory):
     def _make(
         self, searchspace: SearchSpace, objective: Objective, measurements: pd.DataFrame
     ) -> Kernel | GPyTorchKernel:
+        import math
+
         from botorch.models.kernels.positive_index import PositiveIndexKernel
         from botorch.models.utils.gpytorch_modules import (
             get_covar_module_with_dim_scaled_prior,
         )
+        from gpytorch.kernels import ScaleKernel as GPyTorchScaleKernel
+        from gpytorch.priors import LogNormalPrior
 
         parameter_names = self.get_parameter_names(searchspace)
 
@@ -70,16 +74,27 @@ def _make(
             ard_num_dims=ard_num_dims, active_dims=active_dims
         )
 
+        # Wrap in ScaleKernel with Hvarfner-consistent LogNormal outputscale prior
+        outputscale_prior = LogNormalPrior(
+            loc=math.sqrt(2) + math.log(ard_num_dims) * 0.5,
+            scale=math.sqrt(3),
+        )
+        scaled_kernel = GPyTorchScaleKernel(
+            base_kernel,
+            outputscale_prior=outputscale_prior,
+        )
+        scaled_kernel.outputscale = outputscale_prior.mode.detach().clone()
+
         # Single-task case
         if (task_idx := searchspace.task_idx) is None:
-            return base_kernel
+            return scaled_kernel
 
         index_kernel = PositiveIndexKernel(
             num_tasks=searchspace.n_tasks,
             rank=searchspace.n_tasks,
             active_dims=[task_idx],
         )
-        return ICMKernelFactory(base_kernel, index_kernel)(
+        return ICMKernelFactory(scaled_kernel, index_kernel)(
             searchspace, objective, measurements
         )
 
@@ -113,7 +128,6 @@ class HvarfnerLikelihoodFactory(LikelihoodFactoryProtocol):
     def __call__(
         self, searchspace: SearchSpace, objective: Objective, measurements: pd.DataFrame
     ) -> GPyTorchLikelihood:
-
         if searchspace.n_tasks == 1:
             from botorch.models.utils.gpytorch_modules import (
                 get_gaussian_likelihood_with_lognormal_prior,