Remove custom optimizer_argparse case for qKG (facebook#2997)

Summary:

Context: It appears that using `qKnowledgeGradient` with MBM doesn't work, since [this line](https://github.com/facebook/Ax/blob/535af4edff70cbf20a49c676377f5c8945560d03/ax/models/torch/botorch_modular/acquisition.py#L339) passes the argument `optimizer` to `_argparse_kg`, which errors [here](https://github.com/facebook/Ax/blob/535af4edff70cbf20a49c676377f5c8945560d03/ax/models/torch/botorch_modular/optimizer_argparse.py#L169) because it has now received the argument "optimizer" twice.

We don't really need the `optimizer_argparse` special case for qKG anymore. This existed for two reasons:
- in order to construct initial conditions, which can be handled by `optimize_acqf`, and
- to ensure that the optimizer is `optimize_acqf`, because others are not supported

This diff:
* Modifies the `optimize_argparse` case for qKG to do nothing except update the optimizer to `optimize_acqf` and then call the base case

Implementation notes:
* Isn't it nonintuitive to set the optimizer then override it? Yes, a little, but the user can't choose the optimizer, so we're not overriding a user-specified choice. Also, lots of arguments to the `optimizer_argparse` functions get ignored. The "right" thing might be to put the choice of optimizer inside a dispatcher so that it can depend on the acquisition class, but that would be a bigger change.
* Do we really need this dispatcher anymore, if it is doing so little? Maybe. Third parties may wish to use it to accommodate acquisition functions that are not in Ax. On the other hand, this dispatcher is currently not doing much of anything.
* Do the `optimize_argparse` functions still need to support so many arguments, given that some of them seemed to just be there for constructing initial conditions? Probably not; I mean to look into that in a follow-up.

Reviewed By: saitcakmak

Differential Revision: D65227420

ax/benchmark/tests/test_benchmark.py

@@ -42,6 +42,7 @@
 )
 from ax.utils.testing.core_stubs import get_experiment
 from ax.utils.testing.mock import mock_botorch_optimize
+from botorch.acquisition.knowledge_gradient import qKnowledgeGradient
 from botorch.acquisition.logei import qLogNoisyExpectedImprovement
 from botorch.acquisition.multi_objective.logei import (
     qLogNoisyExpectedHypervolumeImprovement,
@@ -323,6 +324,17 @@ def test_replication_mbm(self) -> None:
                 mnist_problem,
                 "MBM::SingleTaskGP_qLogNEI",
             ),
+            (
+                get_sobol_botorch_modular_acquisition(
+                    model_cls=SingleTaskGP,
+                    acquisition_cls=qKnowledgeGradient,
+                    distribute_replications=False,
+                ),
+                get_single_objective_benchmark_problem(
+                    observe_noise_sd=False, num_trials=6
+                ),
+                "MBM::SingleTaskGP_qKnowledgeGradient",
+            ),
         ]:
             with self.subTest(method=method, problem=problem):
                 res = benchmark_replication(problem=problem, method=method, seed=0)

ax/models/torch/botorch_modular/optimizer_argparse.py

@@ -10,13 +10,10 @@
 
 from typing import Any, TypeVar, Union
 
-import torch
 from ax.exceptions.core import UnsupportedError
-from ax.utils.common.constants import Keys
 from ax.utils.common.typeutils import _argparse_type_encoder
 from botorch.acquisition.acquisition import AcquisitionFunction
 from botorch.acquisition.knowledge_gradient import qKnowledgeGradient
-from botorch.optim.initializers import gen_one_shot_kg_initial_conditions
 from botorch.utils.dispatcher import Dispatcher
 
 T = TypeVar("T")
@@ -146,44 +143,36 @@ def _argparse_base(
 @optimizer_argparse.register(qKnowledgeGradient)
 def _argparse_kg(
     acqf: qKnowledgeGradient,
-    q: int,
-    bounds: torch.Tensor,
+    *,
+    optimizer: str = "optimize_acqf",
+    sequential: bool = True,
     num_restarts: int = NUM_RESTARTS,
     raw_samples: int = RAW_SAMPLES,
-    frac_random: float = 0.1,
+    init_batch_limit: int = INIT_BATCH_LIMIT,
+    batch_limit: int = BATCH_LIMIT,
     optimizer_options: dict[str, Any] | None = None,
-    **kwargs: Any,
+    **ignore: Any,
 ) -> dict[str, Any]:
     """
     Argument constructor for optimization with qKG, differing from the
-    base case in that it computes and returns initial conditions.
-
-    To do so, it requires specifying additional arguments `q` and `bounds` and
-    allows for specifying `frac_random`.
+    base case in that it errors if the optimizer is not `optimize_acqf`.
     """
-    base_options = _argparse_base(
-        acqf,
-        num_restarts=num_restarts,
-        raw_samples=raw_samples,
-        optimizer_options=optimizer_options,
+    if optimizer != "optimize_acqf":
+        raise RuntimeError(
+            "Ax is attempting to use a discrete or mixed optimizer, "
+            f"`{optimizer}`, but this is not compatible with "
+            "`qKnowledgeGradient` or its subclasses. To address this, please "
+            "either use a different acquisition class or make parameters "
+            "continuous using the transform "
+            "`ax.modelbridge.registry.Cont_X_trans`."
+        )
+    return _argparse_base(
+        acqf=acqf,
         optimizer="optimize_acqf",
-        **kwargs,
-    )
-
-    initial_conditions = gen_one_shot_kg_initial_conditions(
-        acq_function=acqf,
-        bounds=bounds,
-        q=q,
+        sequential=sequential,
         num_restarts=num_restarts,
         raw_samples=raw_samples,
-        options={
-            Keys.FRAC_RANDOM: frac_random,
-            Keys.NUM_INNER_RESTARTS: num_restarts,
-            Keys.RAW_INNER_SAMPLES: raw_samples,
-        },
+        init_batch_limit=init_batch_limit,
+        batch_limit=batch_limit,
+        optimizer_options=optimizer_options,
     )
-
-    return {
-        **base_options,
-        Keys.BATCH_INIT_CONDITIONS: initial_conditions,
-    }

ax/models/torch/tests/test_optimizer_argparse.py

@@ -8,12 +8,10 @@
 
 from __future__ import annotations
 
-from importlib import reload
 from itertools import product
 from unittest.mock import patch
 
 from ax.exceptions.core import UnsupportedError
-from ax.models.torch.botorch_modular import optimizer_argparse as Argparse
 from ax.models.torch.botorch_modular.optimizer_argparse import (
     _argparse_base,
     BATCH_LIMIT,
@@ -23,7 +21,6 @@
     optimizer_argparse,
     RAW_SAMPLES,
 )
-from ax.utils.common.constants import Keys
 from ax.utils.common.testutils import TestCase
 from botorch.acquisition.acquisition import AcquisitionFunction
 from botorch.acquisition.analytic import LogExpectedImprovement
@@ -182,23 +179,23 @@ def test_optimizer_options(self) -> None:
                     func(None, sequential=False, optimizer=optimizer)
 
     def test_kg(self) -> None:
-        with patch(
-            "botorch.optim.initializers.gen_one_shot_kg_initial_conditions"
-        ) as mock_gen_initial_conditions:
-            mock_gen_initial_conditions.return_value = "TEST"
-            reload(Argparse)
-
-            user_options = {"foo": "bar", "num_restarts": 114}
-            generic_options = _argparse_base(
-                None, optimizer_options=user_options, optimizer="optimize_acqf"
-            )
-            for acqf in (qKnowledgeGradient, qMultiFidelityKnowledgeGradient):
-                with self.subTest(acqf=acqf):
-                    options = optimizer_argparse(
-                        acqf,
-                        q=None,
-                        bounds=None,
-                        optimizer_options=user_options,
-                    )
-                    self.assertEqual(options.pop(Keys.BATCH_INIT_CONDITIONS), "TEST")
-                    self.assertEqual(options, generic_options)
+        user_options = {"foo": "bar", "num_restarts": 114}
+        generic_options = _argparse_base(
+            None, optimizer_options=user_options, optimizer="optimize_acqf"
+        )
+        for acqf in (qKnowledgeGradient, qMultiFidelityKnowledgeGradient):
+            with self.subTest(acqf=acqf):
+                options = optimizer_argparse(
+                    acqf,
+                    q=None,
+                    bounds=None,
+                    optimizer_options=user_options,
+                )
+                self.assertEqual(options, generic_options)
+
+                with self.assertRaisesRegex(
+                    RuntimeError,
+                    "Ax is attempting to use a discrete or mixed optimizer, "
+                    "`optimize_acqf_mixed`, ",
+                ):
+                    optimizer_argparse(acqf, optimizer="optimize_acqf_mixed")