Sagemaker how to pass tuning step's best hyperparameter into another estimator?

Question

I'm running a Sagemaker pipeline with 2 steps, tuning and then training. The purpose is the get the best hyperparameter with tuning, and then use those hyperparameters in the next training step. I am aware that I can use HyperparameterTuningJobAnalytics to retrieve the tuning job specs after the tuning. However, I want to be able to use the hyperparameters like dependency and pass them directly to next trainingStep's estimator, see code below: hyperparameters=step_tuning.properties.BestTrainingJob.TunedHyperParameters, But this doesn't work with this error msg: AttributeError: 'PropertiesMap' object has no attribute 'update'

tf_estimator_final = TensorFlow(entry_point='./train.py',
                          role=role,
                          sagemaker_session=sagemaker_session,
                          code_location=code_location,
                          instance_count=1,
                          instance_type="ml.p3.16xlarge",
                          framework_version='2.4',
                          py_version="py37",
                          base_job_name=base_job_name,
                          output_path=model_path, # if output_path not specified,
hyperparameters=step_tuning.properties.BestTrainingJob.TunedHyperParameters,
                          model_dir="/opt/ml/model",
                          script_mode=True
                          )

step_train = TrainingStep(
    name=base_job_name,
    estimator=tf_estimator_final,
    inputs={
        "train": TrainingInput(
            s3_data=train_s3
        )
    },
    depends_on = [step_tuning]
)

pipeline = Pipeline(
    name=jobname,
    steps=[
        step_tuning,
        step_train
    ],
    sagemaker_session=sagemaker_session
)

json.loads(pipeline.definition())

Any suggestions?

Answer 1

This can't be done in SageMaker Pipelines at the moment.