[Tune] Nevergrad optimizer with extra parameters (#31015)

Some Nevergrad search algorithms have required inputs, such as `budget` for the `NgOpt` search algorithm, but it is not possible with the NevergradSearch class to pass these parameters down to the search algorithm. I would propose adding something like optimizer_kwargs to the NevergradSearch that get passed to the optimizer when instantiating it. Signed-off-by: yhna <[email protected]> Signed-off-by: YH <[email protected]> Signed-off-by: Younghwan Na <[email protected]> Signed-off-by: yhna940 <[email protected]> Co-authored-by: Justin Yu <[email protected]>
ray-project · Jan 23, 2023 · 33d4b14 · 33d4b14
1 parent ee23cc8
commit 33d4b14
Show file tree

Hide file tree

Showing 3 changed files with 26 additions and 6 deletions.
diff --git a/python/ray/tune/search/nevergrad/nevergrad_search.py b/python/ray/tune/search/nevergrad/nevergrad_search.py
@@ -50,8 +50,12 @@ class NevergradSearch(Searcher):
         $ pip install nevergrad
 
     Parameters:
-        optimizer: Optimizer provided
-            from Nevergrad. Alter
+        optimizer: Optimizer class provided from Nevergrad.
+            See here for available optimizers:
+            https://facebookresearch.github.io/nevergrad/optimizers_ref.html#optimizers
+            This can also be an instance of a `ConfiguredOptimizer`. See the
+            section on configured optimizers in the above link.
+        optimizer_kwargs: Kwargs passed in when instantiating the `optimizer`
         space: Nevergrad parametrization
             to be passed to optimizer on instantiation, or list of parameter
             names if you passed an optimizer object.
@@ -120,11 +124,11 @@ def __init__(
         optimizer: Optional[
             Union[Optimizer, Type[Optimizer], ConfiguredOptimizer]
         ] = None,
+        optimizer_kwargs: Optional[Dict] = None,
         space: Optional[Union[Dict, Parameter]] = None,
         metric: Optional[str] = None,
         mode: Optional[str] = None,
         points_to_evaluate: Optional[List[Dict]] = None,
-        **kwargs,
     ):
         assert (
             ng is not None
@@ -134,11 +138,12 @@ def __init__(
         if mode:
             assert mode in ["min", "max"], "`mode` must be 'min' or 'max'."
 
-        super(NevergradSearch, self).__init__(metric=metric, mode=mode, **kwargs)
+        super(NevergradSearch, self).__init__(metric=metric, mode=mode)
 
         self._space = None
         self._opt_factory = None
         self._nevergrad_opt = None
+        self._optimizer_kwargs = optimizer_kwargs or {}
 
         if points_to_evaluate is None:
             self._points_to_evaluate = None
@@ -166,6 +171,13 @@ def __init__(
                     "pass a list of parameter names or None as the `space` "
                     "parameter."
                 )
+            if self._optimizer_kwargs:
+                raise ValueError(
+                    "If you pass in optimizer kwargs, either pass "
+                    "an `Optimizer` subclass or an instance of "
+                    "`ConfiguredOptimizer`."
+                )
+
             self._parameters = space
             self._nevergrad_opt = optimizer
         elif (
@@ -187,7 +199,9 @@ def __init__(
 
     def _setup_nevergrad(self):
         if self._opt_factory:
-            self._nevergrad_opt = self._opt_factory(self._space)
+            self._nevergrad_opt = self._opt_factory(
+                self._space, **self._optimizer_kwargs
+            )
 
         # nevergrad.tell internally minimizes, so "max" => -1
         if self._mode == "max":

diff --git a/python/ray/tune/tests/test_searchers.py b/python/ray/tune/tests/test_searchers.py
@@ -272,6 +272,12 @@ def testNevergrad(self):
             )
         self.assertCorrectExperimentOutput(out)
 
+    def testNevergradWithRequiredOptimizerKwargs(self):
+        from ray.tune.search.nevergrad import NevergradSearch
+        import nevergrad as ng
+
+        NevergradSearch(optimizer=ng.optimizers.CM, optimizer_kwargs=dict(budget=16))
+
     def testOptuna(self):
         from ray.tune.search.optuna import OptunaSearch
         from optuna.samplers import RandomSampler

diff --git a/python/ray/tune/tests/test_tune_restore_warm_start.py b/python/ray/tune/tests/test_tune_restore_warm_start.py
@@ -264,7 +264,7 @@ def cost(space, reporter):
 
         search_alg = NevergradSearch(
             optimizer,
-            parameter_names,
+            space=parameter_names,
             metric="loss",
             mode="min",
         )