Pipeline training step's custom output

Question

A pipeline train step saves a custom json file in the output path, set in the estimator's  `output_path` param, as seen below:
```
estimator = TensorFlow(
        entry_point=code_entry,
        source_dir=code_dir,
        output_path='s3://some-bucket/results/',
        [... other params...]
    )
```
It seems there is no step property to access custom output files, unlike we have for model artificats: `step.properties.ModelArtifacts.S3ModelArtifacts`.

Also, unlike [`ProcessingStep`](https://sagemaker.readthedocs.io/en/stable/workflows/pipelines/sagemaker.workflow.pipelines.html#sagemaker.workflow.steps.ProcessingStep), there is no `outputs` argument on  [`TrainingStep`](https://sagemaker.readthedocs.io/en/stable/workflows/pipelines/sagemaker.workflow.pipelines.html#sagemaker.workflow.steps.TrainingStep) that would allow other steps to access the S3 URI from the arguments e.g., something similar to this:

`step.arguments["ProcessingOutputConfig"]["Outputs"][0]["S3Output"]["S3Uri"]`.
 
How can one access the custom outputs of a training step, stored in the `output.tar.gz`?

AWS-User-5210721 · Answer

There should be a way to achieve this. The properties attribute just mimics the response of Describe* API call. In DescribeTrainingJob API, we have OutputDataConfig field as in https://docs.aws.amazon.com/sagemaker/latest/APIReference/API_DescribeTrainingJob.html#API_DescribeTrainingJob_ResponseSyntax

So should be something like step.properties.OutputDataConfig.S3OutputPath

So should be something like `step.properties.OutputDataConfig.S3OutputPath`

Jady · Answer

Hi,

You will need to provide the output path in the [EstimatorBase](https://sagemaker.readthedocs.io/en/stable/api/training/estimators.html#sagemaker.estimator.EstimatorBase) of the training step.

Examples

```
   estimator_example = Estimator(
        base_job_name="example",
        role=role_arn,
        instance_count=1,
        output_path=f"s3://example-bucket/my-model",
        environment={"region": region.name, "scope": scope_parameter},
        output_kms_key=kms_key_arn
    )

training_step = TrainingStep(
        name="ExampleModelTraining",
        estimator=estimator_example,
        inputs={
            "training_data": TrainingInput(
                s3_data=transform_step.properties.ProcessingOutputConfig.Outputs[
                    transform_training_data_output_name
                ].S3Output.S3Uri
            ),
            "training_target": TrainingInput(
                s3_data=transform_step.properties.ProcessingOutputConfig.Outputs[
                    transform_training_target_output_name
                ].S3Output.S3Uri
            ),
        },
    )
```

Thanks,

Pipeline training step's custom output

Add your answer

Relevant content