Test

PiperOrigin-RevId: 705520724

vizier/_src/algorithms/designers/gp_ucb_pe.py

@@ -322,6 +322,10 @@ def score_with_aux(
     is_missing = (
         features.continuous.is_missing[0] | features.categorical.is_missing[0]
     )
+    logging.info(
+        'self.predictive.predictives.precomputed_solve_on_observation.ndim: %s',
+        self.predictive.predictives.precomputed_solve_on_observation.ndim,
+    )
     gprm_threshold = self.predictive.predict(features)
     threshold = _compute_ucb_threshold(
         gprm_threshold, is_missing, self.ucb_coefficient
@@ -457,6 +461,9 @@ class VizierGPUCBPEBandit(vza.Designer):
   Attributes:
     problem: Must be a flat study with a single metric.
     acquisition_optimizer:
+    gp_model_class: The GP model class, which must implement a `build_model`
+      class method that takes `ModelInput` and returns a
+      `StochasticProcessModel`.
     metadata_ns: Metadata namespace that this designer writes to.
     use_trust_region: Uses trust region.
     ard_optimizer: An optimizer object, which should return a batch of
@@ -475,6 +482,10 @@ class VizierGPUCBPEBandit(vza.Designer):
       kw_only=True,
       factory=lambda: VizierGPUCBPEBandit.default_acquisition_optimizer_factory,
   )
+  _gp_model_class: sp.ModelCoroutine[tfd.GaussianProcess] = attr.field(
+      kw_only=True,
+      factory=lambda: tuned_gp_models.VizierGaussianProcess,
+  )
   _metadata_ns: str = attr.field(
       default='google_gp_ucb_pe_bandit', kw_only=True
   )
@@ -611,7 +622,7 @@ def _build_gp_model_and_optimize_parameters(
       `data.labels`. If `data.features` is empty, the returned parameters are
       initial values picked by the GP model.
     """
-    coroutine = tuned_gp_models.VizierGaussianProcess.build_model(
+    coroutine = self._gp_model_class.build_model(  # pytype: disable=attribute-error
         data.features
     ).coroutine
     model = sp.CoroutineWithData(coroutine, data)

vizier/_src/jax/stochastic_process_model.py

@@ -854,7 +854,11 @@ def predict_with_aux(
         self.predictives.precomputed_solve_on_observation.ndim > 1
     )
     expand_x = len(xs.continuous.shape) == 3 and has_batched_hparams
+    logging.info(
+        'has_batched_hparams: %s, expand_x: %s', has_batched_hparams, expand_x
+    )
     dist = self.predictives._predict(xs, expand_batch_dim=expand_x)  # pylint: disable=protected-access
+    logging.info('dist.batch_shape: %s', dist.batch_shape)
     if has_batched_hparams:
       return (
           tfd.MixtureSameFamily(