Should to_cmd_args pass complex types through json.dumps instead of str?

I am currently hunting down a bug that I could track back into this in sagemaker-training-toolkit.

In my case, I am passing a list as a hyperparameter type, just like:

estimator = Estimator(
    image_uri=docker_img,
    role=role,
    instance_type=instance_type,
    instance_count=1,
    output_path=output_s3,
    hyperparameters={
        "str_param": "foobar",
        "list_param": ["foo", "bar"]
    }
)

Background

To my understanding, sagemaker now uses json.dumps to convert all the values to string and again dumps the resulting dictionary to hyperparamters.json
In sage-maker-toolkit the transformation is reversedhere: https://github.com/aws/sagemaker-training-toolkit/blob/master/src/sagemaker_training/environment.py#L210, leaving us with the original dictionary and values that contain complex objects such as lists.

Issue

When I use env.to_cmd_args() to convert my hyperparameters to command line args, complex values such as lists are not converted back to a json string but simply passed through the str() function. This makes it difficult to reverse the encoding.

    hyperparameters={
        "str_param": "foobar",
        "list_param": ["foo", "bar"]
    }

Will end up in following command line args:
["--str_param", "foobar", "--list_param", "['foo', 'bar']"]. Note that the value for --list_param is a list converted to a string with the str function, not json.dumps.

sagemaker-training-toolkit/src/sagemaker_training/mapping.py

Line 115 in bb469b9

return str(obj)

A proposed solution

If json.dumps was used instead of str, (e.g. return obj if instanceof(obj, str) else json.dumps(obj)) the transformation could be easily reversed with a command line parser:

parser.add_argument("--list_param", type=json.loads)

This seems to be a misunderstanding. As far as I understand now there is no json.dumps logic when passing the hyperparameters to the training instance, so the issue seems invalid.