AML: Clarify what DockerConfiguration does in `train-on-amlcompute.ipynb`

[hamelsmu]

@keijik @cody-dkdc @gregce

In this notebook we are setting a docker configuration, and later passing that to the docker_runtime parameter

from azureml.core import Environment
from azureml.core.runconfig import DockerConfiguration
from azureml.core.conda_dependencies import CondaDependencies

myenv = Environment("myenv")
myenv.python.conda_dependencies = CondaDependencies.create(conda_packages=['scikit-learn', 'packaging'])
# Enable Docker
docker_config = DockerConfiguration(use_docker=True)

....
from azureml.core import ScriptRunConfig
src = ScriptRunConfig(source_directory=project_folder, 
                      script='train.py', 
                      compute_target=cpu_cluster, 
                      environment=myenv,
                      docker_runtime_config=docker_config)

run = experiment.submit(config=src)

Since we are using conda dependencies and are not specifying a base image, what is Docker doing exactly then?

The example works just fine if we omit the docker_config completely. So what is this example trying to show? What is the difference between passing this config and not passing it in practice?

[saachigopal]

I believe this is due to docker being enabled by default now, @vizhur to follow up

[vizhur]

that parameter is not needed if executed on AzureML compute. It matters for local run only if user wants to execute inside the docker.

If you not specifying docker image in the environment the default cpu image will be used. Image version is pinned with every client version