Deploy and retrain a model (10–15%) Flashcards
What things does MLflow track?
Everything, divided into parameters, metrics, and artifacts
If I want mlflow on azure what package should I install?
pip install mlflow azureml-mlflow
How to use mlflow on a local device?
Use the MLflow tracking URI from the overview of the workspace, then do mlflow.set_tracking_uri = “MLFLOW-TRACKING-URI”
How to start a run in mlflow?
It acts like a wrapper, so import it then do
mlflow.set_experiment(experiment_name=<experiment>)</experiment>
How to use autolog in MLflow?
from xgboost import XGBClassifier
with mlflow.start_run():
mlflow.xgboost.autolog()
model = XGBClassifier(use_label_encoder=False, eval_metric="logloss") model.fit(X_train, y_train, eval_set=[(X_test, y_test)], verbose=False)
What are some common function for custom logging in mlflow?
log_param(): single key-value parameter, log_metric(): single key-value metric, log_artifact(): logs a file, like an image, log_model(): logs a model as an MLflow model
How to use custom logging in MLflow?
from xgboost import XGBClassifier
from sklearn.metrics import accuracy_score
with mlflow.start_run():
model = XGBClassifier(use_label_encoder=False, eval_metric=”logloss”)
model.fit(X_train, y_train, eval_set=[(X_test, y_test)], verbose=False)
y_pred = model.predict(X_test)
accuracy = accuracy_score(y_test, y_pred)
mlflow.log_metric("accuracy", accuracy)
What is an endpoint?
An HTTPS endpoint to which you can send data and which will return a response
What are the two types of online endpoints within Azure Machine learning?
There are managed online endpoints and Kubernetes online endpoints
What are managed online endpoints?
Azure Machine Learning manages all the underlying infrastructure
What are kubernetes online endpoints?
Users manage the kubernetes cluster which provides the necessary infrastructure
What four things do you need to deploy a model to a managed online endpoint?
Model assets, scoring script, environment, compute configuration
What is automatically generated when you deploy MLFlow models to an online endpoint?
The scoring script and environment are automatically generated
What is blue/green deployment?
It’s just a/b testing. Set 90% of the traffic for the proven endpoint and 10% for the new one and see how it performs.
What do you use to create an online endpoint?
You use ManagedOnlineEndpoint with the name and auth_mode parameters
What do you need to deploy an MLflow model?
You either need local model files or a registered model in AML. You also need the instance_type and the instance_count
How can you test an endpoint in the AML?
Go to endpoints and then test. Use your data and get the test result.
How can you test an endpoint in Python SDK?
Send your data through ml_client.online_endpoints.invoke() with the name of the endpoint, the name of the deployment, and the json to be sent.
What do you need to create a deployment on the endpoint?
Use ManagedOnlineDeployment with a name, the name of the endpoint, the model, the compute instance, and the number of instances
What functions do you need to create a scoring script?
run() and init()
What is special about a batch endpoint?
You can trigger it from Azure Synapse analytics or Azure Databricks. You can also integrate it with an existing pipeline
What class is used to create a batch endpoint?
BatchEndpoint(name, description)
How do you deploy an MLflow model to a batch endpoint?
Use BatchDeployment with extra specification parameters such as instance_count, max_concurrency_per_instance, mini_batch_size, output_action, output_file_name
How do you invoke a batch endpoint?
You need an input() specifiying the path and asset type, then use ml_client.batch_endpoints.invoke()
