TrainingStep's hyperparameters does not work on v2.2.0

[ohbuchim]

I cannot set TrainingStep's 'hyperparameters' parameter using ExecutionInput on v2.2.0. v2.1.0 works well.

When I used v2.1.0 this code worked, and I was able to set hyperparameters when I executed workflow.

training_step = steps.TrainingStep(
    "SageMaker Training Step",
    estimator=sklearn,
    data={"train": sagemaker.TrainingInput(execution_input["PreprocessingOutputDataTrain"], content_type="text/csv")},
    job_name=execution_input["TrainingJobName"],
    hyperparameters=execution_input["TrainingParameters"],
    wait_for_completion=True,
)

When I used v2.2.0 I needed to add get_schema_as_dict() to ExecutionInput. After that, 'hyperparameters' was not updated by ExecutionInput. It seems that the estimator's 'hyperparameters' setting only is set (ExecutionInput is ignored).

training_step = steps.TrainingStep(
    "SageMaker Training Step",
    estimator=sklearn,
    data={"train": sagemaker.TrainingInput(execution_input["PreprocessingOutputDataTrain"], content_type="text/csv")},
    job_name=execution_input["TrainingJobName"],
    hyperparameters=execution_input["TrainingParameters"].get_schema_as_dict(),
    wait_for_completion=True,
)

[shivlaks]

@ohbuchim - thanks for reporting the issue, it does seem to be a regression due to merging the hyperparameters set in the estimator along with the ones established in the constructor for TrainingStep

in v2.1.0

it would resolve to something like the following, but it would also get rid of any hyperparameters that were established in the estimator by doing so.

'HyperParameters.$': "$$.Execution.Input['TrainingParameters']"

does the workaround actually work though? I think it just discards the constructor parameter altogether.

To support placeholders, I'm thinking perhaps we should skip merging estimator established hyperparameters if a placeholder is provided. Note that this will result in us dropping any hyperparameters that were established in the estimator (as it was with v2.1.0.

what do you think?

[ohbuchim]

Hi, thanks for your comments. I expected estimator's 'hyperparameters' and ExecutionInput's 'hyperparameters' would be merged.

It works for me that estimator's 'hyperparameters' is set by default, and only the parameters set by ExecutionInput are updated. I mean, if the estimator sets hyperparameters={'A': a, 'B': b, 'C': c} and ExecutionInput sets hyperparameters={'B': bb, 'D': dd}, the final hyperparameters should be {'A': a, 'B': bb, 'C': c, 'D': dd}.

[wong-a]

Some clarification about the expected behaviour for this use case is being discussed in https://github.com/aws/aws-step-functions-data-science-sdk-python/pull/159#discussion_r705736120

[ca-nguyen]

Fixed the breaking change: If a Placeholder is provided as input to hyperparameters in TrainingStep constructor, the estimator hyperparameters are overwritten

The fix will be included in the next release

Created a new feature request (#163 ) to facilitate the merging of placeholder hyperparmeters with the estimator parameters

[wong-a]

facilitate the merging of placeholder hyperparmeters with the estimator parameters

Just to clarify, this is possible today. The hyperparameters argument in TrainingStep's constructor just needs to be a dict. TrainingStep does a merge of the hyperparameters dict and estimator's hyperparameters. The dict's values can be Placeholders to reference dynamic values too.

https://github.com/aws/aws-step-functions-data-science-sdk-python/blob/main/src/stepfunctions/steps/sagemaker.py#L72-L74

The hyperparameters arg will only overwrite the estimator's values if the keys conflict. If the top level value is a placeholder, hyperparameters=execution_input["TrainingParameters"] from @ohbuchim's example, the constructor arg will win, producing the following ASL:

"HyperParameters.$": "$$.Execution.Input['TrainingParameters']"

if the estimator sets hyperparameters={'A': a, 'B': b, 'C': c} and ExecutionInput sets hyperparameters={'B': bb, 'D': dd}

To merge the two you need to do the following:

training_step = TrainingStep(...,
  hyperparmeters={ 
    'B': execution_input['TrainingParameters']['B'],
    'D': execution_input['TrainingParameters']['D']
  }
)