How to set a seed for hyperparameter tuning jobs in AWS SageMaker for reproducibility?

[idanmoradarthas]

Hello,

I have a hyper-parameter tuning step in my sagemaker pipeline object as following:

def get_pipelineopt_step_hyper(role, instance_type, instance_count, base_job_prefix, sagemaker_session,
                               step_network_config, data_input_bucket, cache_bucket, branch_name,
                               train_start_date, train_end_date, validation_start_date, validation_end_date,
                               search_type, search_iterations, n_jobs, seed, logger_level,
                               execution_id, volume_kms_key=None):
    parameters = [
        "--train_start_date", train_start_date,
        "--train_end_date", train_end_date,
        "--validation_start_date", validation_start_date,
        "--validation_end_date", validation_end_date,
        "--pipeline_steps", steps
        "--intervention_adjustment", intervention_adjustment,
        "--logger_level", logger_level
    ]

    input_path =...
  
    output_path = ...

    estimator = Estimator(
        image_uri=TRAIN_IMAGE_URI,
        role=role,
        instance_count=instance_count,
        instance_type=instance_type,
        volume_size=200,
        base_job_name=f"{base_job_prefix}/mlp-pipeopt",
        entry_point=str(Path(BASE_DIR).joinpath("pipline_extract_metric_param.py")),
        dependencies=[str(Path(BASE_DIR).joinpath('utils')), str(Path(BASE_DIR).joinpath('features'))],
        sagemaker_session=sagemaker_session,
        output_path=output_path,
        subnets=step_network_config.subnets,
        security_group_ids=step_network_config.security_group_ids,
        enable_network_isolation=step_network_config.enable_network_isolation,
        volume_kms_key=volume_kms_key
    )

    hyperparameter_ranges = {
        "learningWindow__windowsize": IntegerParameter(1, 15),
        "SMOTE__k_neighbors": IntegerParameter(3, 17),
        "SMOTE__sampling_strategy": ContinuousParameter(1 / 5, 4 / 5, scaling_type="Linear"),
        "SampleUsers__lower_bound_hash": CategoricalParameter([0]),
        "SampleUsers__upper_bound_hash": CategoricalParameter([15, 20, 30, 40, 50, 70, 90, 100])
    }

    objective_metric_name = "Harmonic mean score"
    metric_definitions = [
        {
            "Name": objective_metric_name,
            "Regex": r"FDR#, FDR\$ Harmonic mean score: (\d+(.|e\-)\d+)"
        }
    ]

    tuner = HyperparameterTuner(
        estimator=estimator,
        objective_metric_name=objective_metric_name,
        metric_definitions=metric_definitions,
        objective_type="Maximize",
        hyperparameter_ranges=hyperparameter_ranges,
        strategy=search_type,
        max_jobs=search_iterations,
        max_parallel_jobs=n_jobs
    )

    step_tuning = TuningStep(
        name="PipelineOpt",
        tuner=tuner,
        inputs={
            "train": TrainingInput(
                s3_data=Join('/', [input_path, "train", Join("_", [train_start_date, train_end_date]), "train.csv"]),
                content_type="text/csv"
            ),
            "validation": TrainingInput(
                s3_data=Join('/',
                             [input_path, "validation", Join("_", [validation_start_date, validation_end_date]),
                              "validation.csv"]),
                content_type="text/csv"
            )
        },
        job_arguments=parameters
    )

    return step_tuning

How can I set a seed for hyperparameter tuning jobs in AWS SageMaker for reproducibility?

[idanmoradarthas]

Fixed in v2.125.0