Add init_position argument to UniformGenerator (facebook#2686)

Summary:
Pull Request resolved: facebook#2686

`SobolGenerator` uses `init_position` to ensure that when the model is reconstructed, it resumes candidate generation from where it was left (rather than starting from the beginning of the sequence). Without this, when `deduplicate=False`, the model would generate the same points it has already generated, which would lead to different candidate generation behaviors depending on how often the model was reconstructed. This is undesirable as we want the model to resume generation rather than repeating from scratch.

Prior to this diff, `UniformGenerator` did not have `init_position`, so had this exact issue. We fix it here.

Reviewed By: Balandat

Differential Revision: D61622058

ax/models/random/base.py

@@ -48,6 +48,10 @@ class RandomModel(Model):
             of the model will not return the same point twice. This flag
             is used in rejection sampling.
         seed: An optional seed value for scrambling.
+        init_position: The initial state of the generator. This is the number
+            of samples to fast-forward before generating new samples.
+            Used to ensure that the re-loaded generator will continue generating
+            from the same sequence rather than starting from scratch.
         generated_points: A set of previously generated points to use
             for deduplication. These should be provided in the raw transformed
             space the model operates in.
@@ -59,6 +63,7 @@ def __init__(
         self,
         deduplicate: bool = True,
         seed: Optional[int] = None,
+        init_position: int = 0,
         generated_points: Optional[np.ndarray] = None,
         fallback_to_sample_polytope: bool = False,
     ) -> None:
@@ -69,6 +74,7 @@ def __init__(
             if seed is not None
             else checked_cast(int, torch.randint(high=100_000, size=(1,)).item())
         )
+        self.init_position = init_position
         # Used for deduplication.
         self.generated_points = generated_points
         self.fallback_to_sample_polytope = fallback_to_sample_polytope
@@ -180,7 +186,13 @@ def gen(
     @copy_doc(Model._get_state)
     def _get_state(self) -> dict[str, Any]:
         state = super()._get_state()
-        state.update({"seed": self.seed, "generated_points": self.generated_points})
+        state.update(
+            {
+                "seed": self.seed,
+                "init_position": self.init_position,
+                "generated_points": self.generated_points,
+            }
+        )
         return state
 
     def _gen_unconstrained(

ax/models/random/sobol.py

@@ -6,15 +6,13 @@
 
 # pyre-strict
 
-from typing import Any, Callable, Optional
+from typing import Callable, Optional
 
 import numpy as np
 import torch
-from ax.models.base import Model
 from ax.models.model_utils import tunable_feature_indices
 from ax.models.random.base import RandomModel
 from ax.models.types import TConfig
-from ax.utils.common.docutils import copy_doc
 from ax.utils.common.typeutils import not_none
 from torch.quasirandom import SobolEngine
 
@@ -26,17 +24,15 @@ class SobolGenerator(RandomModel):
     the fit or predict methods.
 
     Attributes:
-        init_position: The initial state of the Sobol generator.
-            Starts at 0 by default.
         scramble: If True, permutes the parameter values among
             the elements of the Sobol sequence. Default is True.
         See base `RandomModel` for a description of remaining attributes.
     """
 
     def __init__(
         self,
-        seed: Optional[int] = None,
         deduplicate: bool = True,
+        seed: Optional[int] = None,
         init_position: int = 0,
         scramble: bool = True,
         generated_points: Optional[np.ndarray] = None,
@@ -45,10 +41,10 @@ def __init__(
         super().__init__(
             deduplicate=deduplicate,
             seed=seed,
+            init_position=init_position,
             generated_points=generated_points,
             fallback_to_sample_polytope=fallback_to_sample_polytope,
         )
-        self.init_position = init_position
         self.scramble = scramble
         # Initialize engine on gen.
         self._engine: Optional[SobolEngine] = None
@@ -121,12 +117,6 @@ def gen(
             self.init_position = not_none(self.engine).num_generated
         return (points, weights)
 
-    @copy_doc(Model._get_state)
-    def _get_state(self) -> dict[str, Any]:
-        state = super()._get_state()
-        state.update({"init_position": self.init_position})
-        return state
-
     def _gen_samples(self, n: int, tunable_d: int) -> np.ndarray:
         """Generate n samples.
 

ax/models/random/uniform.py

@@ -10,7 +10,6 @@
 
 import numpy as np
 from ax.models.random.base import RandomModel
-from scipy.stats import uniform
 
 
 class UniformGenerator(RandomModel):
@@ -26,16 +25,21 @@ def __init__(
         self,
         deduplicate: bool = True,
         seed: Optional[int] = None,
+        init_position: int = 0,
         generated_points: Optional[np.ndarray] = None,
         fallback_to_sample_polytope: bool = False,
     ) -> None:
         super().__init__(
             deduplicate=deduplicate,
             seed=seed,
+            init_position=init_position,
             generated_points=generated_points,
             fallback_to_sample_polytope=fallback_to_sample_polytope,
         )
         self._rs = np.random.RandomState(seed=self.seed)
+        if self.init_position > 0:
+            # Fast-forward the random state by generating & discarding samples.
+            self._rs.uniform(size=(self.init_position))
 
     def _gen_samples(self, n: int, tunable_d: int) -> np.ndarray:
         """Generate samples from the scipy uniform distribution.
@@ -48,4 +52,5 @@ def _gen_samples(self, n: int, tunable_d: int) -> np.ndarray:
             samples: An (n x d) array of random points.
 
         """
-        return uniform.rvs(size=(n, tunable_d), random_state=self._rs)  # pyre-ignore
+        self.init_position += n * tunable_d
+        return self._rs.uniform(size=(n, tunable_d))

ax/models/tests/test_uniform.py

@@ -16,36 +16,34 @@
 class UniformGeneratorTest(TestCase):
     def setUp(self) -> None:
         super().setUp()
-        self.tunable_param_bounds = (0, 1)
-        self.fixed_param_bounds = (1, 100)
+        self.tunable_param_bounds = (0.0, 1.0)
+        self.fixed_param_bounds = (1.0, 100.0)
+        self.seed = 0
+        self.expected_points = np.array(
+            [
+                [0.5488135, 0.71518937, 0.60276338],
+                [0.54488318, 0.4236548, 0.64589411],
+                [0.43758721, 0.891773, 0.96366276],
+            ]
+        )
 
-    # pyre-fixme[3]: Return type must be annotated.
-    # pyre-fixme[2]: Parameter must be annotated.
-    def _create_bounds(self, n_tunable, n_fixed):
+    def _create_bounds(self, n_tunable: int, n_fixed: int) -> list[tuple[float, float]]:
         tunable_bounds = [self.tunable_param_bounds] * n_tunable
         fixed_bounds = [self.fixed_param_bounds] * n_fixed
         return tunable_bounds + fixed_bounds
 
-    def test_UniformGeneratorAllTunable(self) -> None:
-        generator = UniformGenerator(seed=0)
+    def test_with_all_tunable(self) -> None:
+        generator = UniformGenerator(seed=self.seed)
         bounds = self._create_bounds(n_tunable=3, n_fixed=0)
         generated_points, weights = generator.gen(
             n=3, bounds=bounds, rounding_func=lambda x: x
         )
-
-        expected_points = np.array(
-            [
-                [0.5488135, 0.71518937, 0.60276338],
-                [0.54488318, 0.4236548, 0.64589411],
-                [0.43758721, 0.891773, 0.96366276],
-            ]
-        )
-        self.assertTrue(np.shape(expected_points) == np.shape(generated_points))
-        self.assertTrue(np.allclose(expected_points, generated_points))
+        self.assertTrue(np.shape(self.expected_points) == np.shape(generated_points))
+        self.assertTrue(np.allclose(self.expected_points, generated_points))
         self.assertTrue(np.all(weights == 1.0))
 
-    def test_UniformGeneratorFixedSpace(self) -> None:
-        generator = UniformGenerator(seed=0)
+    def test_with_fixed_space(self) -> None:
+        generator = UniformGenerator(seed=self.seed)
         bounds = self._create_bounds(n_tunable=0, n_fixed=2)
         n = 3
         with self.assertRaises(SearchSpaceExhausted):
@@ -55,7 +53,7 @@ def test_UniformGeneratorFixedSpace(self) -> None:
                 fixed_features={0: 1, 1: 2},
                 rounding_func=lambda x: x,
             )
-        generator = UniformGenerator(seed=0, deduplicate=False)
+        generator = UniformGenerator(seed=self.seed, deduplicate=False)
         generated_points, _ = generator.gen(
             n=3,
             bounds=bounds,
@@ -66,57 +64,50 @@ def test_UniformGeneratorFixedSpace(self) -> None:
         self.assertTrue(np.shape(expected_points) == np.shape(generated_points))
         self.assertTrue(np.allclose(expected_points, generated_points))
 
-    def test_UniformGeneratorOnline(self) -> None:
+    def test_generating_one_by_one(self, init_position: int = 0) -> None:
         # Verify that the generator will return the expected arms if called
         # one at a time.
-        generator = UniformGenerator(seed=0)
+        generator = UniformGenerator(seed=self.seed, init_position=init_position)
         n_tunable = fixed_param_index = 3
         bounds = self._create_bounds(n_tunable=n_tunable, n_fixed=1)
 
-        n = 3
-        expected_points = np.array(
-            [
-                [0.5488135, 0.71518937, 0.60276338, 1],
-                [0.54488318, 0.4236548, 0.64589411, 1],
-                [0.43758721, 0.891773, 0.96366276, 1],
-            ]
-        )
-        for i in range(n):
+        for i in range(init_position, 3):
             generated_points, weights = generator.gen(
                 n=1,
                 bounds=bounds,
                 fixed_features={fixed_param_index: 1},
                 rounding_func=lambda x: x,
             )
             self.assertEqual(weights, [1])
-            self.assertTrue(np.allclose(generated_points, expected_points[i, :]))
+            self.assertTrue(
+                np.allclose(generated_points[..., :-1], self.expected_points[i, :])
+            )
+            self.assertEqual(generated_points[..., -1], 1)
+            self.assertEqual(generator.init_position, (i + 1) * n_tunable)
 
-    def test_UniformGeneratorReseed(self) -> None:
-        # Verify that the generator will return the expected arms if called
-        # one at a time.
-        generator = UniformGenerator(seed=0)
-        n_tunable = fixed_param_index = 3
-        bounds = self._create_bounds(n_tunable=n_tunable, n_fixed=1)
+    def test_with_init_position(self) -> None:
+        # These are multiples of 3 since there are 3 tunable parameters.
+        self.test_generating_one_by_one(init_position=3)
+        self.test_generating_one_by_one(init_position=6)
 
-        n = 3
-        expected_points = np.array(
-            [
-                [0.5488135, 0.71518937, 0.60276338, 1],
-                [0.54488318, 0.4236548, 0.64589411, 1],
-                [0.43758721, 0.891773, 0.96366276, 1],
-            ]
+    def test_with_reloaded_state(self) -> None:
+        # Check that a reloaded generator will produce the same samples.
+        org_generator = UniformGenerator()
+        bounds = self._create_bounds(n_tunable=3, n_fixed=0)
+        # Generate some to advance the state.
+        org_generator.gen(n=3, bounds=bounds, rounding_func=lambda x: x)
+        # Construct a new generator with the state.
+        new_generator = UniformGenerator(**org_generator._get_state())
+        # Compare the generated samples.
+        org_samples, _ = org_generator.gen(
+            n=3, bounds=bounds, rounding_func=lambda x: x
         )
-        for i in range(n):
-            generated_points, weights = generator.gen(
-                n=1,
-                bounds=bounds,
-                fixed_features={fixed_param_index: 1},
-                rounding_func=lambda x: x,
-            )
-            self.assertEqual(weights, [1])
-            self.assertTrue(np.allclose(generated_points, expected_points[i, :]))
+        new_samples, _ = new_generator.gen(
+            n=3, bounds=bounds, rounding_func=lambda x: x
+        )
+        self.assertTrue(np.allclose(org_samples, new_samples))
 
-    def test_UniformGeneratorWithOrderConstraints(self) -> None:
+    def test_with_order_constraints(self) -> None:
         # Enforce dim_0 <= dim_1 <= dim_2 <= dim_3.
         # Enforce both fixed and tunable constraints.
         generator = UniformGenerator(seed=0)
@@ -143,7 +134,7 @@ def test_UniformGeneratorWithOrderConstraints(self) -> None:
         self.assertTrue(np.shape(expected_points) == np.shape(generated_points))
         self.assertTrue(np.allclose(expected_points, generated_points))
 
-    def test_UniformGeneratorWithLinearConstraints(self) -> None:
+    def test_with_linear_constraints(self) -> None:
         # Enforce dim_0 <= dim_1 <= dim_2 <= dim_3.
         # Enforce both fixed and tunable constraints.
         generator = UniformGenerator(seed=0)
@@ -169,7 +160,7 @@ def test_UniformGeneratorWithLinearConstraints(self) -> None:
         self.assertTrue(np.shape(expected_points) == np.shape(generated_points))
         self.assertTrue(np.allclose(expected_points, generated_points))
 
-    def test_UniformGeneratorBadBounds(self) -> None:
+    def test_with_bad_bounds(self) -> None:
         generator = UniformGenerator()
         with self.assertRaises(ValueError):
             generated_points, weights = generator.gen(