ParameterBoolean is not serialziable.

[Guillem96]

sagemaker-python-sdk/src/sagemaker/workflow/parameters.py

Line 142 in 012c4e8

ParameterBoolean = partial(Parameter, parameter_type=ParameterTypeEnum.BOOLEAN)

When using sagemaker.worflow.parameters.ParameterBoolean as an estimator hyperparameter the pipeline fails since it is not serializable.

train_data = ParameterString(name="TrainData", 
                                               default_value="s3://...")

use_fast_tokenizer = ParameterBoolean(name="UseFastTokenzier", default_value=True)

hyperparameters = {
    # ...
    "use_fast_tokenizer": use_fast_tokenizer,
}

hf_estimator = HuggingFace(
    entry_point='train.py',
    source_dir='src',
    instance_type="...",
    instance_count=1,
    role="...",
    transformers_version='4.4',
    pytorch_version='1.6',
    py_version='py36',
    hyperparameters=hyperparameters
)

training_step = TrainingStep(name="Train",
                             estimator=hf_estimator,
                             inputs={"train": TrainingInput(s3_data=train_data)})

pipeline = Pipeline(
    name="name",
    parameters=[
        train_data,
        use_fast_tokenizer,
    ],
    steps=[training_step],
)

pipeline.upsert(role_arn=role)

The error:

TypeError: Object of type 'Parameter' is not JSON serializable

[melzareix]

Setting it using estimator.set_hyperparameters should work

[Guillem96]

Why? Does it make sense that one approach works while the other not?

[dpatro]

@Guillem96 I was trying to reproduce the issue using the script you have provided. I encountered the following error

botocore.exceptions.ClientError: An error occurred (ValidationException) when calling the CreatePipeline operation: Cannot assign property reference [Parameters.UseFastTokenzier] to argument of type [String].

I was able to get past it by making the following change:

hyperparameters = {
    "use_fast_tokenizer": "True",
}

Reason:

sagemaker-python-sdk/src/sagemaker/huggingface/estimator.py

Lines 81 to 86 in 012c4e8

	hyperparameters (dict): Hyperparameters that will be used for
	training (default: None). The hyperparameters are made
	accessible as a dict[str, str] to the training code on
	SageMaker. For convenience, this accepts other types for keys
	and values, but ``str()`` will be called to convert them before
	training.

After that, the script proceeded to completion without failure.

Can you confirm that you are using the latest version of the SDK? It might be possible that there was a bug in some earlier version but it is fixed.

[dpatro]

Update to my previous comment. If anyone is trying to reference to a pipeline parameter from a training step, the previous method does not work as it hard codes the parameter "use_fast_tokenizer" to True for the training step.

The original problem was because of type mismatch between the hyperparameters values and the defined pipeline parameters values. The hyperparameters expect the values to be Strings. So, by making the following update, the origninal code snippet works fine.

# original - is incompatible with the estimator's hyperparamters values
ParameterBoolean(name="UseFastTokenzier", default_value=True)

# updated - works
ParameterString(name="UseFastTokenzier", default_value="True")

[houman-moallemi]

It also fails for environment variables in the estimator:

is_coco = ParameterBoolean(name="IsCoCo", default_value=False)

Estimator(
    ...
    environment={
        'IS_COCO': is_coco,
    }
)