Azure DP-100 Summary

Major Components of AML Workspaces

Workspace -> Top Level Resource for AML

• Compute Instances
• User Roles
• Compute Targets
• Experiments
• Pipelines
• Datasets
• Models (registered models)
• Deployment Endpoints

Creating an AML Workspace (opens new window)

Resources built as accompanying resource

• Azure Storage Account
• Azure Container Registry
• Azure Application Insights
• Azure Key Vault

Workspace Settings

• Access Controls
• Event Suscriptions (generate alerts or triggers based on events)
• Alerts & Diagnostic Settings

AML Studio

• Author
  • Notebooks
  • Automate ML
  • Designer
• Assets
  • Datasets
  • Experiments
  • Models
  • Endpoints
• Monitoring
  • Compute
  • Datastores
  • Data Labelling

IAM/RBAC (opens new window)

Users in the Azure Active Directory are assigned specific roles which grants access to resources via multippe ways (CLI/Portal/etc).

Experiments (opens new window)

A grouping of many runs

• Run: Singe execution of a training script
• Info Stored: Metadata about run, metrics, etc.

Run Configuration

• Used when we want to run a training experiment on a different compute targets.

Estimator Class

• Allows the creation of run configuration utilizing the AML Python SDK

Designer can only be run on Azure Machine Learning Compute Cluster

Data Objects

Pipelines

Independently executable workflow of a ML task (orchestration).

• Steps that don't need re-run are not run
• Each step can run in a separate compute target
• Dependencies are managed by the pipeline

Datastores

An abstraction over Azure Storage services.

Datasets

References to where the data lives (tabular, file)

Dataset Management

• Version and tracking
• Monitor (data drift)
• Open datasets

Data Drift

• Change in model input data that leads to the degradation of model performance
• Possible causes:
  • Upstream process change
  • Quality issues
  • Natural Causes
  • Covariate shift

Other Notes CSV files can expand up to 10x in a dataframe, and you want double the RAM of that (20GB Ram for a 1GB dataset).

Feature Selection (AML Exam Concepts)

• Pearson's correlation
  • Dependent and Independent Variables don't make any diff.
  • Linear data
• Mutual Info Score
  • Measure the reduction in uncertainty to predict parts of outcomes of a system
  • h(x) = -log(p(x))
• Kendall's Correlation Coefficient
  • A nonparametric analysis of the strength of a relationship between 2 variables
  • Variables are measured on an ordinal scale and data needs to have a monotonic relationship
  • Usually preferred over Spearman
• Chi-Squared Stat
  • Reveals how close expected values are to actual results
  • Used for categorical variables
• Fisher Score
  • Measures the variance between expected value and observed value
  • Determines if a features are independent
• Count-based Feature Selection

Inferencing Notes

Main Docs (opens new window)

• For inferencing in Prod you should use AKS (which supports GPU)
• AML compute instances for deploying real-time services do not provide GPU and should be mainly used for Batch inferencing models.
• Azure Container Instances are for testing/debugging and don't provide GPU (only low-scale CPU workloads).

Authentication

Authentication Docs (opens new window)

Consuming a Web Service (opens new window)

Auth Method	ACI	AKS
Key	Disabled by default	Enabled by default
Token	Not available	Disabled by default

List of Common Modules in AML

• Clean Missing Data (opens new window)
• Select Columns in Dataset (opens new window)
• Normalize Data (opens new window)
• Partition and Sample Data (opens new window)

List of Common/Functions Methods in the SDK

Package	Class	Example	Description
azureml.core	Workspace	Workspace.from_config()
azureml.core	Experiment	Experiment(workspace=ws, name='name')
azureml.core	Experiment	experiment.start_logging()
azureml.core	Datastore	Datastore.register_azure_blob_container(...)
azureml.core	Datastore	Datastore.get(workspace,datastore_name)
azureml.core	Dataset	Dataset.Tabular.from_delimited(path=datastore_path)
azureml.core	Dataset	my_dataset.register(ws,name,description)
azureml.core	Dataset	Dataset.get_by_name(ws,name)
azureml.train.estimator	Estimator	Estimator(source_directory,script_params,...)
azureml.core	Run	Run.get_context()
azureml.core	Run	run.log(),run.log_list(),run.log_row(),etc..
azureml.core	Run	run.get_details(), run.get_metrics(),run.get_file_names
azureml.widgets	RunDetails	RunDetails(run).show()
azureml.core.webservice	AciWebservice	aci_service.get_logs()

Other Commands