feature: add RandomSeed to support reproducible HPO

src/sagemaker/session.py

@@ -2146,6 +2146,7 @@ def tune(  # noqa: C901
         use_spot_instances=False,
         checkpoint_s3_uri=None,
         checkpoint_local_path=None,
+        random_seed=None,
     ):
         """Create an Amazon SageMaker hyperparameter tuning job.
 
@@ -2226,6 +2227,9 @@ def tune(  # noqa: C901
                 started. If the path is unset then SageMaker assumes the
                 checkpoints will be provided under `/opt/ml/checkpoints/`.
                 (default: ``None``).
+            random_seed (int): An initial value used to initialize a pseudo-random number generator.
+                Setting a random seed will make the hyperparameter tuning search strategies to
+                produce more consistent configurations for the same tuning job. (default: ``None``).
         """
 
         tune_request = {
@@ -2238,6 +2242,7 @@ def tune(  # noqa: C901
                 objective_metric_name=objective_metric_name,
                 parameter_ranges=parameter_ranges,
                 early_stopping_type=early_stopping_type,
+                random_seed=random_seed,
                 strategy_config=strategy_config,
             ),
             "TrainingJobDefinition": self._map_training_config(
@@ -2394,6 +2399,7 @@ def _map_tuning_config(
         objective_type=None,
         objective_metric_name=None,
         parameter_ranges=None,
+        random_seed=None,
         strategy_config=None,
     ):
         """Construct tuning job configuration dictionary.
@@ -2412,6 +2418,9 @@ def _map_tuning_config(
             objective_metric_name (str): Name of the metric for evaluating training jobs.
             parameter_ranges (dict): Dictionary of parameter ranges. These parameter ranges can
                 be one of three types: Continuous, Integer, or Categorical.
+            random_seed (int): An initial value used to initialize a pseudo-random number generator.
+                Setting a random seed will make the hyperparameter tuning search strategies to
+                produce more consistent configurations for the same tuning job.
             strategy_config (dict): A configuration for the hyperparameter tuning job optimisation
                 strategy.
 
@@ -2430,6 +2439,9 @@ def _map_tuning_config(
             "TrainingJobEarlyStoppingType": early_stopping_type,
         }
 
+        if random_seed is not None:
+            tuning_config["RandomSeed"] = random_seed
+
         tuning_objective = cls._map_tuning_objective(objective_type, objective_metric_name)
         if tuning_objective is not None:
             tuning_config["HyperParameterTuningJobObjective"] = tuning_objective

src/sagemaker/tuner.py

@@ -413,6 +413,7 @@ def __init__(
         strategy_config: Optional[StrategyConfig] = None,
         early_stopping_type: Union[str, PipelineVariable] = "Off",
         estimator_name: Optional[str] = None,
+        random_seed: Optional[int] = None,
     ):
         """Creates a ``HyperparameterTuner`` instance.
 
@@ -470,6 +471,9 @@ def __init__(
             estimator_name (str): A unique name to identify an estimator within the
                 hyperparameter tuning job, when more than one estimator is used with
                 the same tuning job (default: None).
+            random_seed (int): An initial value used to initialize a pseudo-random number generator.
+                Setting a random seed will make the hyperparameter tuning search strategies to
+                produce more consistent configurations for the same tuning job.
         """
         if hyperparameter_ranges is None or len(hyperparameter_ranges) == 0:
             raise ValueError("Need to specify hyperparameter ranges")
@@ -516,6 +520,7 @@ def __init__(
         self.latest_tuning_job = None
         self.warm_start_config = warm_start_config
         self.early_stopping_type = early_stopping_type
+        self.random_seed = random_seed
 
     def _prepare_for_tuning(self, job_name=None, include_cls_metadata=False):
         """Prepare the tuner instance for tuning (fit)."""
@@ -1222,6 +1227,9 @@ def _prepare_init_params_from_job_description(cls, job_details):
             "base_tuning_job_name": base_from_name(job_details["HyperParameterTuningJobName"]),
         }
 
+        if "RandomSeed" in tuning_config:
+            params["random_seed"] = tuning_config["RandomSeed"]
+
         if "HyperParameterTuningJobObjective" in tuning_config:
             params["objective_metric_name"] = tuning_config["HyperParameterTuningJobObjective"][
                 "MetricName"
@@ -1483,6 +1491,7 @@ def _create_warm_start_tuner(self, additional_parents, warm_start_type, estimato
                     warm_start_type=warm_start_type, parents=all_parents
                 ),
                 early_stopping_type=self.early_stopping_type,
+                random_seed=self.random_seed,
             )
 
         if len(self.estimator_dict) > 1:
@@ -1508,6 +1517,7 @@ def _create_warm_start_tuner(self, additional_parents, warm_start_type, estimato
             max_parallel_jobs=self.max_parallel_jobs,
             warm_start_config=WarmStartConfig(warm_start_type=warm_start_type, parents=all_parents),
             early_stopping_type=self.early_stopping_type,
+            random_seed=self.random_seed,
         )
 
     @classmethod
@@ -1526,6 +1536,7 @@ def create(
         tags=None,
         warm_start_config=None,
         early_stopping_type="Off",
+        random_seed=None,
     ):
         """Factory method to create a ``HyperparameterTuner`` instance.
 
@@ -1586,6 +1597,9 @@ def create(
                 Can be either 'Auto' or 'Off' (default: 'Off'). If set to 'Off', early stopping
                 will not be attempted. If set to 'Auto', early stopping of some training jobs may
                 happen, but is not guaranteed to.
+            random_seed (int): An initial value used to initialize a pseudo-random number generator.
+                Setting a random seed will make the hyperparameter tuning search strategies to
+                produce more consistent configurations for the same tuning job.
 
         Returns:
             sagemaker.tuner.HyperparameterTuner: a new ``HyperparameterTuner`` object that can
@@ -1624,6 +1638,7 @@ def create(
             tags=tags,
             warm_start_config=warm_start_config,
             early_stopping_type=early_stopping_type,
+            random_seed=random_seed,
         )
 
         for estimator_name in estimator_names[1:]:
@@ -1775,6 +1790,9 @@ def _get_tuner_args(cls, tuner, inputs):
             "early_stopping_type": tuner.early_stopping_type,
         }
 
+        if tuner.random_seed is not None:
+            tuning_config["random_seed"] = tuner.random_seed
+
         if tuner.strategy_config is not None:
             tuning_config["strategy_config"] = tuner.strategy_config.to_input_req()
 

tests/unit/test_session.py

@@ -897,6 +897,7 @@ def test_train_pack_to_request(sagemaker_session):
         "ResourceLimits": {"MaxNumberOfTrainingJobs": 100, "MaxParallelTrainingJobs": 5},
         "ParameterRanges": SAMPLE_PARAM_RANGES,
         "TrainingJobEarlyStoppingType": "Off",
+        "RandomSeed": 0,
     },
     "TrainingJobDefinition": {
         "StaticHyperParameters": STATIC_HPs,
@@ -989,6 +990,7 @@ def assert_create_tuning_job_request(**kwrags):
     sagemaker_session.tune(
         job_name="dummy-tuning-1",
         strategy="Bayesian",
+        random_seed=0,
         objective_type="Maximize",
         objective_metric_name="val-score",
         max_jobs=100,
@@ -1080,6 +1082,7 @@ def assert_create_tuning_job_request(**kwrags):
             "max_jobs": 100,
             "max_parallel_jobs": 5,
             "parameter_ranges": SAMPLE_PARAM_RANGES,
+            "random_seed": 0,
         },
         training_config={
             "static_hyperparameters": STATIC_HPs,
@@ -1170,6 +1173,7 @@ def assert_create_tuning_job_request(**kwrags):
     sagemaker_session.tune(
         job_name="dummy-tuning-1",
         strategy="Bayesian",
+        random_seed=0,
         objective_type="Maximize",
         objective_metric_name="val-score",
         max_jobs=100,
@@ -1246,6 +1250,7 @@ def assert_create_tuning_job_request(**kwrags):
     sagemaker_session.tune(
         job_name="dummy-tuning-1",
         strategy="Bayesian",
+        random_seed=0,
         objective_type="Maximize",
         objective_metric_name="val-score",
         max_jobs=100,
@@ -1289,6 +1294,7 @@ def assert_create_tuning_job_request(**kwargs):
     sagemaker_session.tune(
         job_name="dummy-tuning-1",
         strategy="Bayesian",
+        random_seed=0,
         objective_type="Maximize",
         objective_metric_name="val-score",
         max_jobs=100,

tests/unit/test_tuner.py

@@ -545,6 +545,7 @@ def test_attach_tuning_job_with_estimator_from_hyperparameters(sagemaker_session
     assert tuner.strategy == "Bayesian"
     assert tuner.objective_type == "Minimize"
     assert tuner.early_stopping_type == "Off"
+    assert tuner.random_seed == 0
 
     assert isinstance(tuner.estimator, PCA)
     assert tuner.estimator.role == ROLE

tests/unit/tuner_test_utils.py

@@ -112,6 +112,7 @@
             ],
         },
         "TrainingJobEarlyStoppingType": "Off",
+        "RandomSeed": 0,
     },
     "HyperParameterTuningJobName": JOB_NAME,
     "TrainingJobDefinition": {