Azure Machine Learning Documentation

How to Consume Azure Machine Learning Models

This guide provides detailed instructions on how to consume and interact with machine learning models deployed in Azure Machine Learning. We will cover various methods for sending data to your deployed models and receiving predictions.

Introduction

Once your machine learning model is trained and deployed, the next crucial step is to integrate it into your applications or workflows. Azure Machine Learning offers flexible ways to consume your deployed models, enabling real-time inference or batch scoring.

Methods for Consuming Models

You can consume your Azure ML models through several primary mechanisms:

• REST Endpoints: The most common method, allowing any application that can make HTTP requests to interact with your model.
• SDKs: Utilize Azure Machine Learning SDKs (Python, .NET, etc.) for programmatic access and deeper integration.
• Batch Endpoints: For scoring large datasets asynchronously.

1. Consuming via REST Endpoints

Azure Machine Learning deploys models as web services, accessible via REST APIs. This is a language-agnostic approach.

Prerequisites

• A deployed model in Azure Machine Learning.
• The scoring URI and authentication key for your endpoint.

Steps

{
  "data": [
    [1.0, 2.5, 3.1],
    [4.2, 5.0, 6.7]
  ]
}
                    

curl -X POST \
  https://your-endpoint-name.your-region.azureml.ms/score \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer YOUR_PRIMARY_KEY' \
  -d '{
        "data": [
          [1.0, 2.5, 3.1],
          [4.2, 5.0, 6.7]
        ]
      }'
                    

{
  "predictions": [0, 1]
}
                    

2. Consuming via Azure Machine Learning SDK (Python)

The Python SDK provides a more integrated experience for interacting with deployed models.

Prerequisites

• Azure ML SDK installed (pip install azureml-core).
• An authenticated Workspace object.
• The name or ID of your deployed model endpoint.

Steps

from azureml.core.workspace import Workspace
from azureml.core.model import InferenceConfig, Model
from azureml.core.webservice import AciWebservice, Webservice

# Load workspace
ws = Workspace.from_config() # Assumes config.json is present or specified

# Get the deployed service
service_name = 'your-deployed-service-name'
service = Webservice(name=service_name, workspace=ws)
                    

input_data = {"data": [[1.0, 2.5, 3.1], [4.2, 5.0, 6.7]]}
                    

# For real-time inference
result = service.run(input_data=input_data)
print(result)
                    

3. Consuming via Batch Endpoints

Batch endpoints are designed for scoring large volumes of data asynchronously. They are ideal for scenarios where you don't need immediate, real-time predictions.

Key Concepts

• Input Data: Typically stored in Azure Blob Storage or ADLS Gen2.
• Output: Predictions are written to a specified output location.
• Asynchronous: You submit a job and poll for its completion.

Steps (Conceptual)

1. Deploy your model to a Batch Endpoint.
2. Prepare your input data and upload it to a supported storage location.
3. Submit a batch inference job using the Azure ML SDK or CLI, specifying input and output locations.
4. Monitor the job status until completion.
5. Retrieve predictions from the output location.

Refer to the specific Azure ML documentation for Batch Endpoints for detailed command-line or SDK examples.

Best Practices

• Error Handling: Implement robust error handling in your consuming application to gracefully manage API failures or unexpected responses.
• Authentication: Securely manage your authentication keys. Consider using Azure Key Vault.
• Data Validation: Validate input data before sending it to the model to prevent errors and ensure consistency.
• Monitoring: Monitor your deployed endpoints for performance and usage patterns.