What is Azure Machine Learning?
A cloud service called Azure Machine Learning is used to manage and expedite the machine learning (ML) project lifecycle. It is useful for managing machine learning operations (MLOps) for training and deploying models in the daily workflows of ML experts, data scientists, and engineers.
You can utilize a model that has been constructed using an open-source platform like PyTorch, TensorFlow, or scikit-learn, or you can design one in machine learning. You can monitor, retrain, and redeploy models with the use of MLOps tools.
Azure Machine Learning is for whom?
Machine learning enables people and groups putting MLOps into practice inside their company to deploy ML models in a safe, auditable production setting.
Tools can help data scientists and machine learning engineers speed up and automate their daily tasks. Tools for incorporating models into apps or services are available to application developers. Platform developers can create sophisticated ML toolset with a wide range of tools supported by resilient Azure Resource Manager APIs.
Role-based access control for infrastructure and well-known security are available to businesses using the Microsoft Azure cloud. A project can be set up to restrict access to specific operations and protected data.
Features
Utilize important features for the entire machine learning lifecycle.
Preparing data
Data preparation on Apache Spark clusters within Azure Machine Learning may be iterated quickly and is compatible with Microsoft Fabric.
The feature store
By making features discoverable and reusable across workspaces, you may increase the agility with which you ship your models.
Infrastructure for AI
Benefit from specially created AI infrastructure that combines the newest GPUs with InfiniBand networking.
Machine learning that is automated
Quickly develop precise machine learning models for problems like natural language processing, classification, regression, and vision.
Conscientious AI
Create interpretable AI solutions that are accountable. Use disparity measures to evaluate the model’s fairness and reduce unfairness.
Catalog of models
Use the model catalog to find, optimize, and implement foundation models from Hugging Face, Microsoft, OpenAI, Meta, Cohere, and more.
Quick flow
Create, build, test, and implement language model processes in a timely manner.
Endpoint management
Log metrics, carry out safe model rollouts, and operationalize model deployment and scoring.
Azure Machine Learning services
Your needs-compatible cross-platform tools
Anyone on an ML team can utilize their favorite tools. Run quick experiments, hyperparameter-tune, develop pipelines, or manage conclusions using familiar interfaces:
- Azure Machine Learning Studio
- Python SDK (v2)
- Azure CLI(v2)
- Azure Resource Manager REST APIs
Sharing and finding files, resources, and analytics for your projects on the Machine Learning studio UI lets you refine the model and collaborate with others throughout the development cycle.
Azure Machine Learning Studio
Machine Learning Studio provides many authoring options based on project type and familiarity with machine learning, without the need for installation.
- Use managed Jupyter Notebook servers integrated inside the studio to write and run code. Open the notebooks in VS Code, online, or on your PC.
- Visualise run metrics to optimize trials.
- Azure Machine Learning designer: Train and deploy ML models without coding. Drag and drop datasets and components to build ML pipelines.
- Learn how to automate ML experiments with an easy-to-use interface.
- Machine Learning data labeling: Coordinate image and text labeling tasks efficiently.
Using LLMs and Generative AI
Microsoft Azure Machine Learning helps you construct Generative AI applications using Large Language Models. The solution streamlines AI application development with a model portfolio, fast flow, and tools.
Azure Machine Learning Studio and Azure AI Studio support LLMs. This information will help you choose a studio.
Model catalog
Azure Machine Learning studio’s model catalog lets you find and use many models for Generative AI applications. The model catalog includes hundreds of models from Azure OpenAI service, Mistral, Meta, Cohere, Nvidia, Hugging Face, and Microsoft-trained models. Microsoft’s Product Terms define Non-Microsoft Products as models from other sources, which are subject to their terms.
Prompt flow
Azure Machine Learning quick flow simplifies the creation of AI applications using Large Language Models. Prompt flow streamlines AI application prototyping, experimentation, iterating, and deployment.
Enterprise security and readiness
Security is added to ML projects by Azure.
Integrations for security include:
- Network security groups for Azure Virtual Networks.
- Azure Key Vault stores security secrets like storage account access.
- Virtual network-protected Azure Container Registry.
Azure integrations for full solutions
ML projects are supported by other Azure integrations. Among them:
- Azure Synapse Analytics allows Spark data processing and streaming.
- Azure Arc lets you run Azure services on Kubernetes.
- Azure SQL Database, Azure Blob Storage.
- Azure App Service for ML app deployment and management.
- Microsoft Purview lets you find and catalog company data.
Project workflow for machine learning
Models are usually part of a project with goals. Projects usually involve multiple people. Iterative development involves data, algorithms, and models.
Project lifecycle
Project lifecycles vary, but this diagram is typical.
Many users working toward a same goal can collaborate in a workspace. The studio user interface lets workspace users share experimentation results. Job types like environments and storage references can employ versioned assets.
User work can be automated in an ML pipeline and triggered by a schedule or HTTPS request when a project is operational.
The managed inferencing system abstracts infrastructure administration for real-time and batch model deployments.
Train models
Azure Machine Learning lets you run training scripts or construct models in the cloud. Customers commonly bring open-source framework-trained models to operationalize in the cloud.
Open and compatible
Data scientists can utilize Python models in Azure Machine Learning, such as:
- PyTorch
- TensorFlow
- scikit-learn
- XGBoost
- LightGBM
Other languages and frameworks are supported:
- R
- .NET
Automated feature and algorithm selection
Data scientists employ knowledge and intuition to choose the proper data feature and method for training in traditional ML, a repetitive, time-consuming procedure. Automation (AutoML) accelerates this. Use it with Machine Learning Studio UI or Python SDK.
Optimization of hyperparameters
Optimization and adjusting hyperparameters can be arduous. Machine Learning can automate this procedure for every parameterized command with minimal job description changes. The studio displays results.
Multiple-node distributed training
Multinode distributed training can boost deep learning and classical machine learning training efficiency. Azure Machine Learning computing clusters and serverless compute offer the newest GPUs.
Azure Machine Learning Kubernetes, compute clusters, and serverless compute support:
- PyTorch
- TensorFlow
- MPI
MPI distribution works for Horovod and bespoke multinode logic. Serverless Spark compute and Azure Synapse Analytics Spark clusters support Apache Spark.
Embarrassingly parallel training
Scaling an ML project may involve embarrassingly parallel model training. Forecasting demand sometimes involves training a model for many stores.
Deploy models
Use deployment to put a model into production. Azure Machine Learning managed endpoints encapsulate batch or real-time (online) model scoring infrastructure.
Real-time and batch scoring (inferencing)
Endpoints with data references are invoked in batch scoring or inferencing. The batch endpoint asynchronously processes data on computing clusters and stores it for analysis.
Online inference, or real-time scoring, includes contacting an endpoint with one or more model installations and receiving a result via HTTPS in near real time. Traffic can be split over many deployments to test new model versions by redirecting some traffic initially and increasing it after confidence is achieved.
Machine learning DevOps
Production-ready ML models are developed using DevOps, or MLOps. From training to deployment, a model must be auditable if not replicable.
ML model lifecycle
Integrations for MLOPs Machine Learning considers the entire model lifecycle. Auditors can trace the model lifecycle to a commit and environment.
Features that enable MLOps include:
- Git integration.
- Integration of MLflow.
- Machine-learning pipeline scheduling.
- Custom triggers in Azure Event Grid.
- Usability of CI/CD tools like GitHub Actions and Azure DevOps.
Machine Learning has monitoring and auditing features:
- Code snapshots, logs, and other job outputs.
- Asset-job relationship for containers, data, and compute resources.
The airflow-provider-azure-machinelearning package lets Apache Airflow users submit workflows to Azure Machine Learning.
Azure Machine Learning pricing
Pay only what you require; there are no up-front fees.
Utilize Azure Machine Learning for free. Only the underlying computational resources used for model training or inference are subject to charges. A wide variety of machine kinds are available for you to choose from, including general-purpose CPUs and specialist GPUs.
