Support for grid search algorithm in Optuna Suggestion Service

Signed-off-by: Yuki Iwai <yuki.iwai.tz@gmail.com>

manifests/v1beta1/components/controller/katib-config.yaml

@@ -31,7 +31,7 @@ data:
         "image": "docker.io/kubeflowkatib/suggestion-hyperopt:latest"
       },
       "grid": {
-        "image": "docker.io/kubeflowkatib/suggestion-chocolate:latest"
+        "image": "docker.io/kubeflowkatib/suggestion-optuna:latest"
       },
       "hyperband": {
         "image": "docker.io/kubeflowkatib/suggestion-hyperband:latest"

pkg/suggestion/v1beta1/internal/search_space.py

@@ -13,11 +13,12 @@
 # limitations under the License.
 
 import logging
+import numpy as np
 
 from pkg.apis.manager.v1beta1.python import api_pb2 as api
+from pkg.suggestion.v1beta1.internal.constant import INTEGER, DOUBLE, CATEGORICAL, DISCRETE
 import pkg.suggestion.v1beta1.internal.constant as constant
 
-
 logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger(__name__)
 
@@ -36,15 +37,38 @@ def convert(experiment):
             search_space.goal = constant.MIN_GOAL
         for p in experiment.spec.parameter_specs.parameters:
             search_space.params.append(
-                HyperParameterSearchSpace.convertParameter(p))
+                HyperParameterSearchSpace.convert_parameter(p))
         return search_space
 
+    @staticmethod
+    def convert_to_combinations(search_space):
+        combinations = {}
+
+        for parameter in search_space.params:
+            if parameter.type == INTEGER:
+                combinations[parameter.name] = range(int(parameter.min), int(parameter.max)+1, int(parameter.step))
+            elif parameter.type == DOUBLE:
+                if parameter.step == "" or parameter.step is None:
+                    raise Exception(
+                        "Param {} step is nil; For discrete search space, all parameters must include step".
+                        format(parameter.name)
+                    )
+                double_list = np.arange(float(parameter.min), float(parameter.max)+float(parameter.step),
+                                        float(parameter.step))
+                if double_list[-1] > float(parameter.max):
+                    double_list = double_list[:-1]
+                combinations[parameter.name] = double_list
+            elif parameter.type == CATEGORICAL or parameter.type == DISCRETE:
+                combinations[parameter.name] = parameter.list
+
+        return combinations
+
     def __str__(self):
         return "HyperParameterSearchSpace(goal: {}, ".format(self.goal) + \
             "params: {})".format(", ".join([element.__str__() for element in self.params]))
 
     @staticmethod
-    def convertParameter(p):
+    def convert_parameter(p):
         if p.parameter_type == api.INT:
             # Default value for INT parameter step is 1
             step = 1

pkg/suggestion/v1beta1/optuna/base_service.py

@@ -17,6 +17,7 @@
 
 from pkg.suggestion.v1beta1.internal.constant import INTEGER, DOUBLE, CATEGORICAL, DISCRETE, MAX_GOAL
 from pkg.suggestion.v1beta1.internal.trial import Assignment
+from pkg.suggestion.v1beta1.internal.search_space import HyperParameterSearchSpace
 
 
 class BaseOptunaService(object):
@@ -48,6 +49,10 @@ def _create_sampler(self):
         elif self.algorithm_name == "random":
             return optuna.samplers.RandomSampler(**self.algorithm_config)
 
+        elif self.algorithm_name == "grid":
+            combinations = HyperParameterSearchSpace.convert_to_combinations(self.search_space)
+            return optuna.samplers.GridSampler(combinations, **self.algorithm_config)
+
     def get_suggestions(self, trials, current_request_number):
         if len(trials) != 0:
             self._tell(trials)

pkg/suggestion/v1beta1/optuna/service.py

@@ -15,10 +15,10 @@
 import threading
 import grpc
 import logging
+import itertools
 
 from pkg.apis.manager.v1beta1.python import api_pb2
 from pkg.apis.manager.v1beta1.python import api_pb2_grpc
-from pkg.suggestion.v1beta1.internal.constant import INTEGER, DOUBLE
 from pkg.suggestion.v1beta1.internal.search_space import HyperParameterSearchSpace
 from pkg.suggestion.v1beta1.internal.trial import Trial, Assignment
 from pkg.suggestion.v1beta1.optuna.base_service import BaseOptunaService
@@ -55,7 +55,7 @@ def GetSuggestions(self, request, context):
 
     def ValidateAlgorithmSettings(self, request, context):
         is_valid, message = OptimizerConfiguration.validate_algorithm_spec(
-            request.experiment.spec.algorithm)
+            request.experiment)
         if not is_valid:
             context.set_code(grpc.StatusCode.INVALID_ARGUMENT)
             context.set_details(message)
@@ -86,6 +86,9 @@ class OptimizerConfiguration(object):
         "random": {
             "seed": lambda x: int(x),
         },
+        "grid": {
+            "seed": lambda x: int(x),
+        }
     }
 
     @classmethod
@@ -110,7 +113,8 @@ def convert_algorithm_spec(cls, algorithm_spec):
         return algorithm_spec.algorithm_name, config
 
     @classmethod
-    def validate_algorithm_spec(cls, algorithm_spec):
+    def validate_algorithm_spec(cls, experiment):
+        algorithm_spec = experiment.spec.algorithm
         algorithm_name = algorithm_spec.algorithm_name
         algorithm_settings = algorithm_spec.algorithm_settings
 
@@ -120,6 +124,10 @@ def validate_algorithm_spec(cls, algorithm_spec):
             return cls._validate_cmaes_setting(algorithm_settings)
         elif algorithm_name == "random":
             return cls._validate_random_setting(algorithm_settings)
+        elif algorithm_name == "grid":
+            return cls._validate_grid_setting(experiment)
+        else:
+            return False, "unknown algorithm name {}".format(algorithm_name)
 
     @classmethod
     def _validate_tpe_setting(cls, algorithm_spec):
@@ -178,3 +186,34 @@ def _validate_random_setting(cls, algorithm_settings):
                                                                                        exception=e)
 
         return True, ""
+
+    @classmethod
+    def _validate_grid_setting(cls, experiment):
+        algorithm_settings = experiment.spec.algorithm.algorithm_settings
+        search_space = HyperParameterSearchSpace.convert(experiment)
+
+        for s in algorithm_settings:
+            try:
+                if s.name == "random_state":
+                    if not int(s.value) >= 0:
+                        return False, ""
+                else:
+                    return False, "unknown setting {} for algorithm grid".format(s.name)
+
+            except Exception as e:
+                return False, "failed to validate {name}({value}): {exception}".format(name=s.name, value=s.value,
+                                                                                       exception=e)
+
+        try:
+            combinations = HyperParameterSearchSpace.convert_to_combinations(search_space)
+            num_combinations = len(list(itertools.product(*combinations.values())))
+            max_trial_count = experiment.spec.max_trial_count
+            if max_trial_count > num_combinations:
+                return False, "Max Trial Count: {max_trial} > all possible search combinations: {combinations}".\
+                    format(max_trial=max_trial_count, combinations=num_combinations)
+
+        except Exception as e:
+            return False, "failed to validate parameters({parameters}): {exception}".\
+                format(parameters=search_space.params, exception=e)
+
+        return True, ""