Unable to specify custom role for Sagemaker

[nasxisbest]

Issue Description

Unable to use Sagemaker role when provisioning Sagemaker resources. Here's how I am used to do it through AWS Console:

Found this in Sagemaker Python library docs:

Tried leveraging with:

backend_args = {
    "autogluon_sagemaker_estimator_kwargs": {
        "role": "arn:aws:iam::176253007097:role/service-role/SageMaker-winnie-ml",
    }
}

cloud_predictor.fit(
    job_name="insights-autogluon-refund-ml",
    predictor_init_args=predictor_init_args,
    predictor_fit_args=predictor_fit_args,
    backend_kwargs=backend_args,
)

But got this error:

TypeError: autogluon.cloud.utils.ag_sagemaker.AutoGluonSagemakerEstimator() got multiple values for keyword argument 'role'

[nasxisbest]

Will file the PR to patch this soon.

[nasxisbest]

Here you go boyz: #183

[suzhoum]

@nasxisbest Thanks for the contribution! AutoGluon Cloud makes use of sagemaker.get_execution_role() at the backend to assume the AWS role in your environment. May I understand what is your use case for specifying the role ARN?

[nasxisbest]

Sure, thanks for asking!

There are only specific enterprise accounts in AWS, that we're allowed to login to in our company. Those accounts are created per team.

For additional access like sagemaker:* we'd need to create a new role and assume it during the resource creation. This is what this dropdown in UI is for:

• 
• ... or same role argument in in Sagemaker Python library.

That's why during the use of Autogluon-Cloud, I am passing role into Sagemaker backend. Like this:

cloud_predictor.fit(
    ...
    backend_kwargs={
        "autogluon_sagemaker_estimator_kwargs": {
            "role": "arn:aws:iam::xxxxxxxxxxxx:role/service-role/SageMaker-winnie-ml",
        }
    },
)

But this doesn't work due to the bug within the code that has been fixed with this PR.

[suzhoum]

Got it. Thanks for the context! I'll follow up in the PR.