ML.NET CI/CD MLOps Azure DevOps Sample

Sample ML.NET application that trains, evaluates and deploys a model using Azure DevOps. The model uses reviews from Yelp to train an binary classification model to analyze sentiment. The trained model is stored in an Azure Storage Blob container and used to make predictions over HTTP by an Azure Function.

Prerequisites

• .NET SDK 3.1
• Azure Storage Account
• Azure Functions Core Tools

The Application

• Domain: F# .NET Core Class Library containing schema of input and output data used by the model.
• TrainConsole: C# .NET Core Console Application. Trains a binary classification model to analyze sentiment.
• TestModel: C# XUnit Test Project. Used to evaluate the model.
• AzFnUploadService: C# Azure Functions HTTP Triggered serverless application that uploads the trained model to a validation / staging Azure Storage Blob container and the evaluated model to a production Azure Storage Blob container.
• AzFnSentimentAnalysisService: C# Azure Functions HTTP Triggered serverless application that uses the production model to make predictions.
• azure-pipelines.yml: Contains Azure DevOps Build/Test/Deploy job definitions.

Data

The data used in this application comes from the Sentiment Labeled Sentences UCI repository.

Run the applications

Use the following steps to run the applications

TrainConsole

1. Navigate to TrainConsole project directory.
2. Use .NET CLI and provide the URI of the dataset used for training as an argument. The URI can be hosted anywhere. For this application, it was hosted in an Azure Storage Blob container and anonymous read access was provided for the blob containing the data.

dotnet run <DATASET-URI>

TestModel

1. Navigate to TestModel project directory.
2. Use the .NET CLI to run the following command:

dotnet test

AzFnUploadService

1. Make sure you have in Azure Blob Storage, containers called validation and production.
2. Navigate to the AzFnUploadService project directory.
3. Create a local.settings.json file with the following content. Replace the AZURE-BLOB-STORAGE-CONNECTION-STRING with the connection string of the blob account where you'll be saving your models to.

{
  "IsEncrypted": false,
  "Values": {
    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
    "FUNCTIONS_WORKER_RUNTIME": "dotnet",
    "AzModelStorageConnectionString": "<AZURE-BLOB-STORAGE-CONNECTION-STRING>"
  }
}

1. Use the Azure Functions Core Tools in the terminal to start the application:

func host start

1. In another terminal, navigate to the TrainConsole project directory.
2. Use curl in the terminal to upload the model. Make sure to replace ENDPOINT with Validation or Production depending which Azure Storage Blob Container you want to upload the model to.

curl --location --request POST http://localhost:7071/api/Upload<ENDPOINT> \
--form '=@model.zip'

You should receive the following response if the model was uploaded successfully:

Uploaded Successfully

AzFnSentimentAnalysisService

1. Navigate to the AzFnSentimentAnalysisService project directory.
2. Create a local.settings.json file with the following content. Replace the PRODUCTION-MODEL-URI with the URI of the blob containing the production model. (Validation is also valid. This URI can point to any serialized version of the model)

{
  "IsEncrypted": false,
  "Values": {
    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
    "FUNCTIONS_WORKER_RUNTIME": "dotnet",
    "MODEL_URI": "<PRODUCTION-MODEL-URI>"
  }
}

1. Use the Azure Functions Core Tools in the terminal to start the application:

func host start

1. In another terminal, use curl to classify the sentiment of a comment.

curl --location --request POST 'http://localhost:7071/api/AnalyzeSentiment' \
--header 'Content-Type: application/json' \
--data-raw '{
	"Comment": "The steak was bad!"
}'

The response should look as follows:

Negative

Additional Resources

• ML.NET Documentation
• ML.NET YouTube Playlist
• ML.NET Samples
• ML.NET GitHub Repo
• Azure DevOps
• Azure DevOps Documentation
• Azure Functions Documentation
• .NET Documentation