Search code examples
amazon-sagemakeraws-pipeline

AWS SageMaker Pipeline Issue - Pipeline variables do not support __str__ operation

I'm trying to build SageMaker Pipeline based on Tensorflow framework. I have only Training, Evaluating steps, and Register model. On the evaluation step I declared MetricsSource for ModelMetrics and received an error.

Code is below:

pipeline_model = PipelineModel(
    models=[tf_model],
    role=role, 
    sagemaker_session=sagemaker_session
)

eval_res = step_evaluate_model.arguments['ProcessingOutputConfig']['Outputs'][0]['S3Output']['S3Uri']
evaluation_s3_uri = f'{eval_res}/evaluation.json'

model_statistics=MetricsSource(
        s3_uri=evaluation_s3_uri,
        content_type='application/json')

model_metrics = ModelMetrics(model_statistics=model_statistics)

step_register_pipeline_model = pipeline_model.register(
    content_types=['application/json'],
    response_types=['application/json'],
    inference_instances=['ml.m4.xlarge','ml.c5.2xlarge'],
    transform_instances=['ml.c5.2xlarge'],
    model_package_group_name=model_package_group_name,
    model_metrics=model_metrics,
    approval_status=model_approval_status.default_value,
)

Error:

TypeError                                 Traceback (most recent call last)
Input In [17], in <cell line: 17>()
     14 model_metrics = ModelMetrics(model_statistics=model_statistics)
     15 # print('\n',pipeline_model)
---> 17 step_register_pipeline_model = pipeline_model.register(
     18     content_types=['application/json'],
     19     response_types=['application/json'],
     20     inference_instances=['ml.m4.xlarge','ml.c5.2xlarge'],
     21     transform_instances=['ml.c5.2xlarge'],
     22     model_package_group_name=model_package_group_name,
     23     model_metrics=model_metrics,
     24     approval_status=model_approval_status.default_value,
     25 )
TypeError: Pipeline variables do not support __str__ operation. Please use `.to_string()` to convert it to string type in execution timeor use `.expr` to translate it to Json for display purpose in Python SDK.

Could you please help me to solve it? I'd appreciate for any idea. Thanks

Solution

  • They way to create model and register model on Pipelines has changed slightly with the introduction of ModelStep, also the instantiation of session_pipeline is needed. Similarly, ModelStep will be used for registering the model . Reference: https://github.com/aws/sagemaker-python-sdk/pull/3076 Examples : https://sagemaker.readthedocs.io/en/stable/workflows/pipelines/sagemaker.workflow.pipelines.html?highlight=ModelStep#sagemaker.workflow.model_step.ModelStep