TrainingStep fails when Estimator object has 'debugger_hook_config' as False

[jennakwon06]

Hello,

When debugger_hook_config=False is specified in an estimator object, TrainingStep will fail with the following error:

AttributeError                            Traceback (most recent call last)
<ipython-input-9-ce223140d492> in <module>
     20     state_id="samplestate",
     21     estimator=estimator,
---> 22     job_name="samplejob"
     23 )

~/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/stepfunctions/steps/sagemaker.py in __init__(self, state_id, estimator, job_name, data, hyperparameters, mini_batch_size, experiment_config, wait_for_completion, tags, **kwargs)
     69 
     70         if estimator.debugger_hook_config != None:
---> 71             parameters['DebugHookConfig'] = estimator.debugger_hook_config._to_request_dict()
     72 
     73         if estimator.rules != None:

AttributeError: 'bool' object has no attribute '_to_request_dict'

As per this doc: https://sagemaker.readthedocs.io/en/stable/amazon_sagemaker_debugger.html#default-behavior-and-opting-out , debugger_hook_config=False is required to disable the hook iniialization.

Code sample:

estimator = TensorFlow(
    entry_point="training.py",
    role="arn:aws:iam::747911269416:role/service-role/AmazonSageMaker-ExecutionRole-20190313T101302",
    train_instance_count=1,
    train_instance_type="ml.m5d.large",
    output_path="s3://jkkwon-miami-us-west-2",
    code_location="s3://jkkwon-miami-us-west-2",
    train_volume_size=1024,
    metric_definitions=[
        {"Name": "train:loss", "Regex": "Train Loss: (.*?);"},
        {"Name": "test:loss", "Regex": "Test Average loss: (.*?),"},
        {"Name": "test:accuracy", "Regex": "Test Accuracy: (.*?)%;"},
    ],
    enable_sagemaker_metrics=True,
    debugger_hook_config=False
)


training_step = TrainingStep(
    state_id="samplestate",
    estimator=estimator,
    job_name="samplejob"
)

[vaib-amz]

Hi @jennakwon06,

Thank you for raising this. We have merged fix #83 for this. This should be available with the next release of the SDK.

Thanks,
Vaib

[jennakwon06]

Hello,

When is the next release containing this scheduled?
We are needing this fix to be downloadable from PyPi.

[shivlaks]

#83 was included in v1.1.2, which was released almost a year ago (Sept 23, 2020)