Add a wrapper to allow fairer comparisons between Vizier and other ra…

…ytune algorithms.

This wrapper ensures that all algorithms start with the same default parameters.

PiperOrigin-RevId: 635475355

vizier/_src/pythia/suggest_default.py

@@ -30,7 +30,7 @@
 _T = TypeVar('_T')
 
 
-def _get_default_parameters(search_space: vz.SearchSpace) -> vz.ParameterDict:
+def get_default_parameters(search_space: vz.SearchSpace) -> vz.ParameterDict:
   """Gets the default parameters for the given search space."""
   builder = vz.SequentialParameterBuilder(search_space)
 
@@ -94,7 +94,7 @@ def wrapper_fn(self: Policy, request: SuggestRequest) -> SuggestDecision:
     if request.max_trial_id > 0:
       return suggest_fn(self, request)
 
-    default_parameters = _get_default_parameters(
+    default_parameters = get_default_parameters(
         request.study_config.search_space
     )
     decision = SuggestDecision([vz.TrialSuggestion(default_parameters)])

vizier/_src/pythia/suggest_default_test.py

@@ -39,7 +39,7 @@ class GetDefaultParametersTest(absltest.TestCase):
   def test_double_user_default(self):
     ss = vz.SearchSpace()
     ss.root.add_float_param('x', 0.0, 1.0, default_value=0.2)
-    params = suggest_default._get_default_parameters(ss)
+    params = suggest_default.get_default_parameters(ss)
     self.assertEqual(params.get_value('x'), 0.2)
 
   @unittest.skip('TODO')
@@ -48,25 +48,25 @@ def test_double_logscale(self):
     ss.root.add_float_param(
         'x', np.exp(-2), np.exp(2), scale_type=vz.ScaleType.LOG
     )
-    params = suggest_default._get_default_parameters(ss)
+    params = suggest_default.get_default_parameters(ss)
     self.assertEqual(params.get_value('x'), 1.0)
 
   def test_double_fixed(self):
     ss = vz.SearchSpace()
     ss.root.add_float_param('x', 1.0, 1.0)
-    params = suggest_default._get_default_parameters(ss)
+    params = suggest_default.get_default_parameters(ss)
     self.assertEqual(params.get_value('x'), 1.0)
 
   def test_discrete(self):
     ss = vz.SearchSpace()
     ss.root.add_discrete_param('x', [1, 2, 3, 6])
-    params = suggest_default._get_default_parameters(ss)
+    params = suggest_default.get_default_parameters(ss)
     self.assertEqual(params.get_value('x'), 3)
 
   def test_categorical(self):
     ss = vz.SearchSpace()
     ss.root.add_categorical_param('x', ['a', 'b', 'c', 'd'])
-    params = suggest_default._get_default_parameters(ss)
+    params = suggest_default.get_default_parameters(ss)
     self.assertEqual(params.get_value('x'), 'c')