copybara-service · Sep 10, 2024 · Sep 10, 2024
diff --git a/vizier/_src/benchmarks/experimenters/normalizing_experimenter.py b/vizier/_src/benchmarks/experimenters/normalizing_experimenter.py
@@ -15,11 +15,12 @@
 from __future__ import annotations
 
 """Experimenter that normalizes the range of each metric."""
+
 import copy
 from typing import Dict, Sequence
 
 import numpy as np
-from vizier import pyvizier
+from vizier import pyvizier as vz
 from vizier._src.benchmarks.experimenters import experimenter
 from vizier.pyvizier import converters
 
@@ -42,25 +43,19 @@ def __init__(
     self._exptr = exptr
     self._problem_statement = exptr.problem_statement()
 
-    min_vals: list[float] = []
-    max_vals: list[float] = []
-    for pc in self._problem_statement.search_space.parameters:
-      if pc.type != pyvizier.ParameterType.DOUBLE:
-        raise ValueError(f'Only DOUBLE parameters can be normalized: {pc}.')
-      min_value, max_value = pc.bounds
-      min_vals.append(min_value)
-      max_vals.append(max_value)
-
-    grid = np.linspace(
-        np.asarray(min_vals), np.asarray(max_vals), num_normalization_samples
-    )
+    # Convert search space into hypercube, then sample.
     converter = converters.TrialToArrayConverter.from_study_config(
-        study_config=self._problem_statement, scale=False
+        study_config=self._problem_statement, scale=True, pad_oovs=False
     )
-    grid_parameters = converter.to_parameters(grid)
+    feature_dim = sum([o.num_dimensions for o in converter.output_specs])
+
+    normalized_samples = np.linspace(
+        np.zeros(feature_dim), np.ones(feature_dim), num_normalization_samples
+    )
+    sampled_params = converter.to_parameters(normalized_samples)
     metrics: Dict[str, list[float]] = {}
-    for parameters in grid_parameters:
-      trial = pyvizier.Trial(parameters=parameters)
+        for parameters in sampled_params:
+      trial = vz.Trial(parameters=parameters)
       exptr.evaluate([trial])
       measurement = trial.final_measurement
       for name, metric in (measurement.metrics if measurement else {}).items():
@@ -79,17 +74,17 @@ def __init__(
         )
       self._normalizations[name] = normalization
 
-  def problem_statement(self) -> pyvizier.ProblemStatement:
+  def problem_statement(self) -> vz.ProblemStatement:
     return copy.deepcopy(self._problem_statement)
 
-  def evaluate(self, suggestions: Sequence[pyvizier.Trial]):
+  def evaluate(self, suggestions: Sequence[vz.Trial]):
     self._exptr.evaluate(suggestions)
     for suggestion in suggestions:
       if suggestion.final_measurement is None:
         continue
-      normalized_metrics: Dict[str, pyvizier.Metric] = {}
+      normalized_metrics: Dict[str, vz.Metric] = {}
       for name, metric in suggestion.final_measurement.metrics.items():
-        normalized_metrics[name] = pyvizier.Metric(
+        normalized_metrics[name] = vz.Metric(
             value=metric.value / self._normalizations[name]
         )
       suggestion.final_measurement.metrics = normalized_metrics
@@ -99,3 +94,44 @@ def __repr__(self):
         f'NormalizingExperimenter with normalizations {self._normalizations} on'
         f' {self._exptr}'
     )
+
+
+class HyperCubeExperimenter(experimenter.Experimenter):
+  """Normalizes the search space into unit hypercube using converter."""
+
+  def __init__(self, exptr: experimenter.Experimenter):
+    self._exptr = exptr
+    original_problem = exptr.problem_statement()
+
+    converter = converters.TrialToArrayConverter.from_study_config(
+        study_config=original_problem, scale=True, pad_oovs=False
+    )
+    self._converter = converter
+    feature_dim = sum([o.num_dimensions for o in converter.output_specs])
+
+    new_space = vz.SearchSpace()  # Setup hypercube search space.
+    for i in range(feature_dim):
+      new_space.add(vz.ParameterConfig.factory(f'h{i}', bounds=(0.0, 1.0)))
+    self._problem_statement = copy.deepcopy(original_problem)
+    self._problem_statement.search_space = new_space
+
+    # For simply converting new hypercube parameters to numpy arrays.
+    self._eval_converter = converters.TrialToArrayConverter.from_study_config(
+        study_config=self._problem_statement, scale=False, pad_oovs=False
+    )
+
+  def problem_statement(self) -> vz.ProblemStatement:
+    return copy.deepcopy(self._problem_statement)
+
+  def evaluate(self, suggestions: Sequence[vz.Trial]):
+    hypercube_features = self._eval_converter.to_features(suggestions)
+    original_params = self._converter.to_parameters(hypercube_features)
+
+    orig_suggestions = copy.deepcopy(suggestions)
+    for param_dict, orig_suggestion in zip(original_params, orig_suggestions):
+      orig_suggestion.parameters = param_dict
+
+    self._exptr.evaluate(orig_suggestions)
+
+    for suggestion, orig_suggestion in zip(suggestions, orig_suggestions):
+      suggestion.final_measurement = orig_suggestion.final_measurement
diff --git a/vizier/_src/benchmarks/experimenters/normalizing_experimenter_test.py b/vizier/_src/benchmarks/experimenters/normalizing_experimenter_test.py
@@ -15,6 +15,7 @@
 from __future__ import annotations
 
 from vizier import pyvizier
+from vizier._src.benchmarks.experimenters import combo_experimenter
 from vizier._src.benchmarks.experimenters import normalizing_experimenter
 from vizier._src.benchmarks.experimenters import numpy_experimenter
 from vizier._src.benchmarks.experimenters.synthetic import bbob
@@ -72,5 +73,46 @@ def testNormalizationApply(self, func):
     self.assertBetween(normalized_value, -10, 10)
 
 
+class HyperCubeExperimenterTest(parameterized.TestCase):
+
+  def testE2E(self):
+    num_boolean_params = 3
+
+    original_exptr = combo_experimenter.ContaminationExperimenter(
+        contamination_n_stages=num_boolean_params
+    )
+    hypercube_exptr = normalizing_experimenter.HyperCubeExperimenter(
+        original_exptr
+    )
+
+    problem = hypercube_exptr.problem_statement()
+    metric_name = problem.metric_information.item().name
+
+    # 3 Booleans -> 6 total coordinates after one-hots.
+    self.assertLen(problem.search_space.parameters, 2 * num_boolean_params)
+
+    t = pyvizier.Trial(
+        parameters={
+            'h0': 0.0,
+            'h1': 1.0,
+            'h2': 1.0,
+            'h3': 0.0,
+            'h4': 0.3,
+            'h5': 0.6,
+        }
+    )
+    hypercube_exptr.evaluate([t])
+    hypercube_value = t.final_measurement_or_die.metrics[metric_name].value
+
+    # Argmax over hypercube one-hots.
+    t = pyvizier.Trial(
+        parameters={'x_0': 'True', 'x_1': 'False', 'x_2': 'True'}
+    )
+    original_exptr.evaluate([t])
+    original_value = t.final_measurement_or_die.metrics[metric_name].value
+
+    self.assertEqual(hypercube_value, original_value)
+
+
 if __name__ == '__main__':
   absltest.main()
diff --git a/vizier/benchmarks/experimenters/__init__.py b/vizier/benchmarks/experimenters/__init__.py
@@ -34,6 +34,7 @@
 from vizier._src.benchmarks.experimenters.l1_categorical_experimenter import L1CategorialExperimenter
 from vizier._src.benchmarks.experimenters.multiobjective_experimenter import MultiObjectiveExperimenter
 from vizier._src.benchmarks.experimenters.noisy_experimenter import NoisyExperimenter
+from vizier._src.benchmarks.experimenters.normalizing_experimenter import HyperCubeExperimenter
 from vizier._src.benchmarks.experimenters.normalizing_experimenter import NormalizingExperimenter
 from vizier._src.benchmarks.experimenters.numpy_experimenter import MultiObjectiveNumpyExperimenter
 from vizier._src.benchmarks.experimenters.numpy_experimenter import NumpyExperimenter