How to use Azure AutoML from a DSS Notebook

DivyaThomas · May 2020

Azure Machine Learning can be used for any kind of machine learning, from classical ML to deep learning, supervised, and unsupervised learning. Whether you prefer to write Python or R code or zero-code/low-code options such as the designer, you can build, train, and track highly accurate machine learning and deep-learning models in an Azure Machine Learning Workspace. Users can apply automated ML when they want Azure Machine Learning to train and tune a model for them using a specified target metric.

It is currently possible to leverage Azure's AutoML capabilities from within a DSS python notebook. The following details the necessary configuration for Azure and DSS to make this integration possible, as well as a code example of how to create and deploy an Azure AutoML job from DSS.

Configuration

Azure side

Create a machine learning workspace.
Create a storage account:
- Choose the new “StorageV2”
- Once created, go to the storage page -> Containers -> Create a container (mine is container-of-dku). This will be the container used by DSS.

DSS side

Create the new connection to the storage that we just created:
- Access Key can be found on the portal UI when going to your storage page -> Access Key.
- Path restrictions -> Container -> container-of-dku
- Advanced -> HDFS interface -> ABFS
Create a python3 (mandatory) code-env with these requirements (according to here):
- matplotlib
- numpy
- cython
- urllib3
- scipy
- scikit-learn
- tensorflow
- xgboost
- azureml-sdk
- azureml-widgets
- azureml-explain-model
- pandas-ml
- azureml-defaults
- azureml-dataprep[pandas]
- azureml-train-automl
- azureml-train
- azureml-widgets
- azureml-pipeline
- azureml-contrib-interpret
- pytorch-transformers==1.0.0
- spacy==2.1.8
- onnxruntime==1.0.0
- https://aka.ms/automl-resources/packages/en_core_web_sm-2.1.0.tar.gz
The autoML api we will use requires that the files are uncompressed csv, and they must have the columns name. This is not the default behavior when DSS creates a managed azure dataset, so there are some format config we need to modify manually.
- To do that, go to azure dataset on the flow you intend to use as the input for your model. Then select Settings -> Previews and input the following values:
  - File compression: None
  - Choose Parse next line as column headers
- Return to the parent recipe of the dataset and rerun it
To check before starting, go to your storage page -> Containers -> YOUR_CONTAINER -> dataiku -> YOUR_PROJECT -> YOUR_DATASET. Inside that folder you should see that your files are in the format “out-sX.csv”. Download the first file and check that it contains the column names.

Train an autoML model

Create a Python Code recipe with the desired training data, stored in Azure, as your recipe input
Open up your recipe in a notebook and set the code-env to the previously created AzureML environment

Connect to the workspace

from azureml.core import Workspace 
# THIS INFORMATION CAN BE FOUND ON THE AZURE WORKSPACE UI
config = {
     "subscription_id": "XXXXXXXXXXX",
     "resource_group": "resource-of-dku",
     "workspace_name": "playground-of-dku"
}
subscription_id = config.get('subscription_id')
resource_group = config.get('resource_group')
workspace_name = config.get('workspace_name')

ws = Workspace(subscription_id = subscription_id, resource_group = resource_group, workspace_name = workspace_name)

Create an Experiment

from azureml.core.experiment import Experiment
experiment_name = 'automl-experiment-of-dku' 

experiment = Experiment(ws, experiment_name)

Create a Compute Cluster

from azureml.core.compute import AmlCompute
from azureml.core.compute import ComputeTarget

# Choose a name for your cluster. 
amlcompute_cluster_name = "cluster-of-dku"
found = False
# Check if this compute target already exists in the workspace.
cts = ws.compute_targets
if amlcompute_cluster_name in cts and cts[amlcompute_cluster_name].type == 'AmlCompute':
    found = True
    print('Found existing compute target.')
    compute_target = cts[amlcompute_cluster_name]

if not found:
    print('Creating a new compute target...')
    provisioning_config = AmlCompute.provisioning_configuration(vm_size = "Standard_D4_v2",  max_nodes = 10)
    compute_target = ComputeTarget.create(ws, amlcompute_cluster_name, provisioning_config)

print('Checking cluster status...')
compute_target.wait_for_completion(show_output = True, min_node_count = None, timeout_in_minutes = 20)

Creation of the cluster will take some time so a waiting time of a few minutes is normal.

Define the datastore for ML purpose

from azureml.core import Datastore
import dataiku

client = dataiku.api_client()

my_dss_azure_connection_name = 'azure_blob_dku' # CHANGE THIS TO YOUR CONNECTION
azure_connection = client.get_connection(my_dss_azure_connection_name)
azure_connection_info = azure_connection.get_info().get('params', {})
datastore_name='datastore_of_dku' # CHANGE THIS
container_name = azure_connection_info.get('chcontainer')
account_name = azure_connection_info.get('storageAccount')
account_key = azure_connection_info.get('accessKey')
default_managed_folder = azure_connection_info.get('defaultManagedContainer')

blob_datastore = Datastore.register_azure_blob_container(workspace=ws, datastore_name=datastore_name, container_name=container_name, account_name=account_name, account_key=account_key)

Represent your azure dataset as a TabularDataset, so that it can be used with the autoML api

from azureml.core.dataset import Dataset

dataset = dataiku.Dataset('simple_table_azure') # CHANGE THIS INPUT DATASET NAME
project_key = dataset.get_config().get('projectKey')
raw_path = dataset.get_config().get('params').get('path')
path = raw_path.replace('${projectKey}', project_key)

my_train_dataset = Dataset.Tabular.from_delimited_files(path=[(blob_datastore, path)], separator='\t')

You can test if the file format has been configured correctly by doing the following:

my_train_dataset.take(3).to_pandas_dataframe()

It should return a clean dataframe.

If everything looks good, register your TabularDataset, the name and description are not that important, so you can just keep them as in the example if you want.

my_train_dataset = my_train_dataset.register(workspace=ws, name='simple_blob_train_dataset', description='Simple training data', create_new_version=True)

Configure the AutoML task

import logging
from azureml.train.automl import AutoMLConfig
target_column  = 'Churn' #CHANGE THIS TO THE NAME OF YOUR TARGET COLUMN
automl_settings = {
    "n_cross_validations": 2,
    "primary_metric": 'average_precision_score_weighted',
    "whitelist_models": ['RandomForest'],
    "enable_early_stopping": True,
    "max_concurrent_iterations": 8, 
    "max_cores_per_iteration": -1,
    "experiment_timeout_hours" : 0.25,
    "iteration_timeout_minutes": 1, 
    "verbosity": logging.INFO,
}
automl_config = AutoMLConfig(task = 'classification', debug_log = 'automl_errors.log', compute_target = compute_target, training_data = my_train_dataset, label_column_name = target_column, **automl_settings)

Submit the task

remote_run = experiment.submit(automl_config, show_output = True)

And that’s it, now just sit back, and wait until the job completes. It will output a summary of its training test and best model.

Deploy the model

There is no easy way to do this by code, so users will need to create 1) a python scoring file and 2) a yaml configuration file.

Meanwhile, in the UI, it just takes 2 click like specified here, so we recommend users to deploy via the UI.

Users can choose to either deploy the model on Azure Kubernetes Service (AKS) or Azure Container Instance (ACI).

To test the endpoint, you can try this code snippet in a notebook:

# URL for the web service
scoring_uri = "http://f9ed9cf6-3541-4f8e-aed3-525d8e731a38.eastus2.azurecontainer.io/score"
# If the service is authenticated, set the key or token
key = '<your key or token>'

scoring_data = [[11.01, 717.2, 75.56]]
data = {"data": scoring_data}
# Convert to JSON string
input_data = json.dumps(data)

# Set the content type
headers = {'Content-Type': 'application/json'}
# If authentication is enabled, set the authorization header
#headers['Authorization'] = f'Bearer {key}'

# Make the request and display the response
resp = requests.post(scoring_uri, input_data, headers=headers)