change: remove instance_pools parameter from tuner

this change is to remove the concept of instance_pools from multi-
algorithm hyperparameter tuning as an input configuration.

Author: Alex

src/sagemaker/session.py

@@ -788,7 +788,6 @@ def _map_tuning_config(
         objective_type=None,
         objective_metric_name=None,
         parameter_ranges=None,
-        training_instance_pools=None,
     ):
         """
         Construct tuning job configuration dictionary.
@@ -807,10 +806,6 @@ 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.
-            training_instance_pools (dict[str, int]): Dictionary to specify how many ML instances
-                of different types to be reserved before starting the hyperparameter tuning job.
-                The keys are the ML instance types, and the values are the numbers of the
-                instances to be reserved.
 
         Returns:
             A dictionary of tuning job configuration. For format details, please refer to
@@ -834,12 +829,6 @@ def _map_tuning_config(
         if parameter_ranges is not None:
             tuning_config["ParameterRanges"] = parameter_ranges
 
-        if training_instance_pools is not None:
-            tuning_config["TrainingJobInstancePools"] = [
-                {"InstanceType": instance_type, "PoolSize": training_instance_pools[instance_type]}
-                for instance_type in sorted(training_instance_pools.keys())
-            ]
-
         return tuning_config
 
     @classmethod

src/sagemaker/tuner.py

@@ -206,7 +206,6 @@ def __init__(
         warm_start_config=None,
         early_stopping_type="Off",
         estimator_name=None,
-        training_instance_pools=None,
     ):
         """Initialize a ``HyperparameterTuner``. It takes an estimator to obtain
         configuration information for training jobs that are created as the
@@ -260,19 +259,6 @@ 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).
-            training_instance_pools (dict[str, str]): Dictionary to specify how many
-                ML instances of different types to be reserved before starting the
-                hyperparameter tuning job (default: None). The keys are the ML instance
-                types, and the values are the numbers of the instances to be reserved.
-                The following example specifies two instance pools. One pool has
-                10 ml.m4.4xlarge instances, while the other has 2 ml.p2.8xlarge
-                instances.
-                >>> {
-                >>>     'ml.m4.4xlarge': 10,
-                >>>     'ml.p2.8xlarge': 2
-                >>> }
-                For more details, see
-                https://botocore.readthedocs.io/en/latest/reference/services/sagemaker.html#SageMaker.Client.create_hyper_parameter_tuning_job
         """
         if hyperparameter_ranges is None or len(hyperparameter_ranges) == 0:
             raise ValueError("Need to specify hyperparameter ranges")
@@ -313,7 +299,6 @@ def __init__(
         self.latest_tuning_job = None
         self.warm_start_config = warm_start_config
         self.early_stopping_type = early_stopping_type
-        self.training_instance_pools = training_instance_pools
 
     def _prepare_for_tuning(self, job_name=None, include_cls_metadata=False):
         """Prepare the tuner instance for tuning (fit)"""
@@ -952,12 +937,6 @@ def _prepare_init_params_from_job_description(cls, job_details):
                 tuning_config["ParameterRanges"]
             )
 
-        if "TrainingJobInstancePools" in tuning_config:
-            params["training_instance_pools"] = {
-                instance_pool["InstanceType"]: instance_pool["PoolSize"]
-                for instance_pool in tuning_config["TrainingJobInstancePools"]
-            }
-
         if "TrainingJobDefinition" in job_details:
             params["metric_definitions"] = job_details["TrainingJobDefinition"][
                 "AlgorithmSpecification"
@@ -1230,7 +1209,6 @@ def _create_warm_start_tuner(self, additional_parents, warm_start_type, estimato
             max_jobs=self.max_jobs,
             max_parallel_jobs=self.max_parallel_jobs,
             warm_start_config=WarmStartConfig(warm_start_type=warm_start_type, parents=all_parents),
-            training_instance_pools=self.training_instance_pools,
             early_stopping_type=self.early_stopping_type,
         )
 
@@ -1241,7 +1219,6 @@ def create(
         objective_metric_name_dict,
         hyperparameter_ranges_dict,
         metric_definitions_dict=None,
-        training_instance_pools=None,
         base_tuning_job_name=None,
         strategy="Bayesian",
         objective_type="Maximize",
@@ -1284,10 +1261,6 @@ def create(
                 name of the metric, and 'Regex' for the regular expression used to extract the
                 metric from the logs. This should be defined only for hyperparameter tuning jobs
                 that don't use an Amazon algorithm.
-            training_instance_pools (dict[str, str]): Dictionary to specify how many ML instances
-                of different types to be reserved before starting the hyperparameter tuning job.
-                The keys are the ML instance types, and the values are the numbers of the instances
-                to be reserved.
             base_tuning_job_name (str): Prefix for the hyperparameter tuning job name when the
                 :meth:`~sagemaker.tuner.HyperparameterTuner.fit` method launches. If not specified,
                 a default job name is generated, based on the training image name and current
@@ -1344,7 +1317,6 @@ def create(
             max_jobs=max_jobs,
             max_parallel_jobs=max_parallel_jobs,
             tags=tags,
-            training_instance_pools=training_instance_pools,
             warm_start_config=warm_start_config,
             early_stopping_type=early_stopping_type,
         )
@@ -1484,9 +1456,6 @@ def start_new(cls, tuner, inputs):
         if parameter_ranges is not None:
             tuning_config["parameter_ranges"] = parameter_ranges
 
-        if tuner.training_instance_pools is not None:
-            tuning_config["training_instance_pools"] = tuner.training_instance_pools
-
         tuner_args = {
             "job_name": tuner._current_job_name,
             "tuning_config": tuning_config,

src/sagemaker/workflow/airflow.py

@@ -357,15 +357,6 @@ def _extract_tuning_job_config(tuner):
     if parameter_ranges:
         tuning_job_config["ParameterRanges"] = parameter_ranges
 
-    if tuner.training_instance_pools:
-        tuning_job_config["TrainingJobInstancePools"] = [
-            {
-                "InstanceType": instance_type,
-                "PoolSize": tuner.training_instance_pools[instance_type],
-            }
-            for instance_type in sorted(tuner.training_instance_pools.keys())
-        ]
-
     return tuning_job_config
 
 

tests/integ/test_tuner_multi_algo.py

@@ -68,11 +68,6 @@ def data_set():
     return data_set
 
 
-@pytest.fixture(scope="module")
-def training_instance_pools(cpu_instance_type):
-    return {cpu_instance_type: 2}
-
-
 @pytest.fixture(scope="function")
 def estimator_fm(sagemaker_session, cpu_instance_type):
     fm_image = get_image_uri(
@@ -114,12 +109,7 @@ def estimator_knn(sagemaker_session, cpu_instance_type):
 
 @pytest.mark.canary_quick
 def test_multi_estimator_tuning(
-    sagemaker_session,
-    estimator_fm,
-    estimator_knn,
-    training_instance_pools,
-    data_set,
-    cpu_instance_type,
+    sagemaker_session, estimator_fm, estimator_knn, data_set, cpu_instance_type
 ):
     tuner = HyperparameterTuner.create(
         base_tuning_job_name=BASE_TUNING_JOB_NAME,
@@ -136,7 +126,6 @@ def test_multi_estimator_tuning(
         objective_type=OBJECTIVE_TYPE,
         max_jobs=MAX_JOBS,
         max_parallel_jobs=MAX_PARALLEL_JOBS,
-        training_instance_pools=training_instance_pools,
         tags=TAGS,
     )
 

tests/unit/test_airflow.py

@@ -671,7 +671,6 @@ def test_multi_estimator_tuning_config(sagemaker_session):
         max_parallel_jobs="{{ max_parallel_job }}",
         tags=[{"{{ key }}": "{{ value }}"}],
         base_tuning_job_name="{{ base_job_name }}",
-        training_instance_pools={"ml.m4.xlarge": 4, "ml.c4.2xlarge": 10},
     )
 
     data = {
@@ -690,10 +689,6 @@ def test_multi_estimator_tuning_config(sagemaker_session):
                 "MaxParallelTrainingJobs": "{{ max_parallel_job }}",
             },
             "TrainingJobEarlyStoppingType": "Off",
-            "TrainingJobInstancePools": [
-                {"InstanceType": "ml.c4.2xlarge", "PoolSize": 10},
-                {"InstanceType": "ml.m4.xlarge", "PoolSize": 4},
-            ],
         },
         "TrainingJobDefinitions": [
             {

tests/unit/test_session.py

@@ -807,10 +807,6 @@ def test_train_pack_to_request(sagemaker_session):
         "Strategy": "Bayesian",
         "ResourceLimits": {"MaxNumberOfTrainingJobs": 100, "MaxParallelTrainingJobs": 5},
         "TrainingJobEarlyStoppingType": "Off",
-        "TrainingJobInstancePools": [
-            {"InstanceType": "ml.m4.4xlarge", "PoolSize": 3},
-            {"InstanceType": "ml.p2.xlarge", "PoolSize": 1},
-        ],
     },
     "TrainingJobDefinitions": [
         {
@@ -1001,12 +997,7 @@ def assert_create_tuning_job_request(**kwrags):
     )
     sagemaker_session.create_tuning_job(
         job_name="dummy-tuning-1",
-        tuning_config={
-            "strategy": "Bayesian",
-            "max_jobs": 100,
-            "max_parallel_jobs": 5,
-            "training_instance_pools": {"ml.m4.4xlarge": 3, "ml.p2.xlarge": 1},
-        },
+        tuning_config={"strategy": "Bayesian", "max_jobs": 100, "max_parallel_jobs": 5},
         training_config_list=[
             {
                 "static_hyperparameters": STATIC_HPs,

tests/unit/test_tuner.py

@@ -376,7 +376,6 @@ def test_fit_multi_estimators(sagemaker_session):
     assert tune_kwargs["tuning_config"]["max_jobs"] == MAX_JOBS
     assert tune_kwargs["tuning_config"]["max_parallel_jobs"] == MAX_PARALLEL_JOBS
     assert tune_kwargs["tuning_config"]["early_stopping_type"] == EARLY_STOPPING_TYPE
-    assert tune_kwargs["tuning_config"]["training_instance_pools"] == TRAINING_INSTANCE_POOLS
 
     assert "tuning_objective" not in tune_kwargs["tuning_config"]
     assert "parameter_ranges" not in tune_kwargs["tuning_config"]
@@ -455,7 +454,6 @@ def _create_multi_estimator_tuner(sagemaker_session):
             ESTIMATOR_NAME_TWO: HYPERPARAMETER_RANGES_TWO,
         },
         metric_definitions_dict={ESTIMATOR_NAME: METRIC_DEFINITIONS},
-        training_instance_pools=TRAINING_INSTANCE_POOLS,
         strategy=STRATEGY,
         objective_type=OBJECTIVE_TYPE,
         max_jobs=MAX_JOBS,
@@ -613,8 +611,6 @@ def test_attach_tuning_job_with_multi_estimators(sagemaker_session):
     assert tuner.early_stopping_type == "Off"
     assert tuner.warm_start_config is None
 
-    assert tuner.training_instance_pools == TRAINING_INSTANCE_POOLS
-
     assert tuner.estimator is None
     assert tuner.objective_metric_name is None
     assert tuner._hyperparameter_ranges is None
@@ -1001,7 +997,6 @@ def test_create_tuner(estimator_dict, obj_metric_name_dict, param_ranges_dict, m
         objective_metric_name_dict=obj_metric_name_dict,
         hyperparameter_ranges_dict=param_ranges_dict,
         metric_definitions_dict=metric_def_dict,
-        training_instance_pools=TRAINING_INSTANCE_POOLS,
         strategy="Bayesian",
         objective_type="Minimize",
         max_jobs=MAX_JOBS,
@@ -1019,7 +1014,6 @@ def test_create_tuner(estimator_dict, obj_metric_name_dict, param_ranges_dict, m
     assert tuner.metric_definitions_dict == metric_def_dict
 
     assert tuner.base_tuning_job_name == BASE_JOB_NAME
-    assert tuner.training_instance_pools == TRAINING_INSTANCE_POOLS
     assert tuner.strategy == "Bayesian"
     assert tuner.objective_type == "Minimize"
     assert tuner.max_jobs == MAX_JOBS
@@ -1103,7 +1097,6 @@ def test_create_tuner_negative(
             objective_metric_name_dict=obj_metric_name_dict,
             hyperparameter_ranges_dict=param_ranges_dict,
             metric_definitions_dict=metric_def_dict,
-            training_instance_pools=TRAINING_INSTANCE_POOLS,
             strategy="Bayesian",
             objective_type="Minimize",
             max_jobs=MAX_JOBS,

tests/unit/tuner_test_utils.py

@@ -59,7 +59,6 @@
 
 METRIC_DEFINITIONS = "mock_metric_definitions"
 
-TRAINING_INSTANCE_POOLS = {"ml.m4.4xlarge": 5, "ml.p2.8xlarge": 2}
 MAX_JOBS = 10
 MAX_PARALLEL_JOBS = 5
 TAGS = [{"key1": "value1"}]
@@ -165,10 +164,6 @@
         "ResourceLimits": {"MaxParallelTrainingJobs": 2, "MaxNumberOfTrainingJobs": 4},
         "Strategy": "Bayesian",
         "TrainingJobEarlyStoppingType": "Off",
-        "TrainingJobInstancePools": [
-            {"InstanceType": instance_type, "PoolSize": pool_size}
-            for (instance_type, pool_size) in TRAINING_INSTANCE_POOLS.items()
-        ],
     },
     "HyperParameterTuningJobName": JOB_NAME,
     "TrainingJobDefinitions": [