Data Concepts in Azure Machine Learning

Azure Machine Learning Data Concepts

URI

• A Uniform Resource Identifier (URI) represents a storage location on a local computer, Azure storage, or a publicly available http(s) location.

• URIs can be used as inputs or outputs to an Azure Machine Learning job and can be mapped to the compute target filesystem in different modes: read-only mount, read-write mount, download, or upload.

• URIs use identity-based authentication to connect to storage services, with options for Azure Active Directory ID or Managed Identity.

Data types

• Azure Machine Learning supports three data types: File, Folder, and Table.

• File: References a single file and can have any format.

• Folder: References a single folder and is useful for deep-learning tasks with various file types such as images, text, audio, and video.

• Table: References a data table and is suitable for a complex schema with frequent changes or large tabular data subsets.

Data runtime capability

• Azure Machine Learning uses its own data runtime for mounts, uploads, downloads, and materialization of tabular data into pandas/spark.

• The data runtime is built with Rust language for high speed and efficiency.

• It has no dependencies on other technologies, allowing for quick installation on compute targets.



• It supports multi-process data loading and pre-fetching to enhance GPU utilization in deep-learning operations.

• Provides seamless authentication to cloud storage.

Datastore

• An Azure Machine Learning datastore is a reference to an existing Azure storage account.

• It provides a common API for interacting with different storage types (Blob/Files/ADLS) and facilitates team operations.

• Datastore creation and use offer easier discovery of useful datastores and secure connection information for credential-based access.

• Authentication methods include credential-based (service principal/SAS/key) and identity-based (Azure Active Directory or managed identity).

Data asset

• An Azure Machine Learning data asset allows users to create a reference to frequently used data sources with a friendly name.

• Data asset creation includes metadata and a reference to the data source location without incurring extra storage costs or risking data source integrity.

• Data assets can be created from Azure Machine Learning datastores, Azure Storage, public URLs, or local files.

Data splits & cross-validation (Python)

Data Splits

• In Azure Automated Machine Learning, the recommended approach is to randomly split the data into training and evaluation sets based on rows.

• The AutoMLConfig object represents the configuration for submitting an automated ML experiment in Azure Machine Learning, containing parameters and training data for the experiment run.

Provide validation data

• Provide a separate validation set by specifying the validation data in your machine learning process to assess the model’s performance on unseen data during training.

Provide validation set size

• Control the size of the validation set by specifying the desired percentage or number of samples to be allocated for validation to fine-tuning the model and evaluating its generalization ability.

K-fold cross-validation

• Dividing the data into K subsets, or “folds,” and using each fold as a validation set while training on the remaining data to provide a robust evaluation by averaging the results across multiple iterations.

Monte Carlo cross-validation

• A technique where multiple random training and validation splits are generated to mitigate bias in the model evaluation caused by a particular split.

Specify custom cross-validation data folds

• Specify custom cross-validation data folds using CV split columns in the model configuration, giving you control over the data divisions for validation.

Metric calculation for cross-validation in machine learning

• Calculates metrics on each validation fold and aggregates them for comprehensive model performance evaluation, ensuring reliable assessment.

References:

Data concepts in Azure Machine Learning – Azure Machine Learning

Data splits and cross-validation in automated machine learning – Azure Machine Learning

azureml.train.automl.automlconfig.AutoMLConfig class – Azure Machine Learning Python

Secure data access in the cloud v1 – Azure Machine Learning

Evaluate AutoML experiment results – Azure Machine Learning

Azure Machine Learning – ML as a Service | Microsoft Azure