MSDN Community

Deploying AI Models with .NET

This article guides you through the process of deploying your trained AI models using the .NET ecosystem. We'll cover various strategies and best practices to ensure your AI solutions are robust, scalable, and accessible.

Understanding Deployment Scenarios

The way you deploy your AI model often depends on its purpose and the environment it will run in. Common scenarios include:

• Web Services (APIs): Exposing your model as a RESTful API for on-demand predictions.
• Edge Devices: Deploying models directly onto IoT devices or mobile applications for real-time processing.
• Batch Processing: Running predictions on large datasets offline.
• Cloud-Native Applications: Integrating models into microservices architectures running on platforms like Azure Kubernetes Service (AKS).

Key Technologies and Frameworks in .NET

The .NET ecosystem offers a rich set of tools for AI model deployment:

• ML.NET: Microsoft's open-source, cross-platform machine learning framework for .NET developers. It provides APIs for model training and consumption.
• ASP.NET Core: A high-performance, cross-platform framework for building modern, cloud-based, internet-connected applications. Ideal for creating AI model APIs.
• Azure Machine Learning: A cloud-based service that enables developers to build, train, and deploy machine learning models.
• ONNX Runtime: A high-performance inference engine that allows you to run models trained in various frameworks (like TensorFlow, PyTorch) on .NET applications.

Step-by-Step Deployment to a Web Service

Let's walk through deploying a simple ML.NET model as an ASP.NET Core Web API.

1. Train and Save Your Model

Assuming you have already trained a model using ML.NET and saved it to a file (e.g., model.zip).

2. Create an ASP.NET Core Web API Project

Use the .NET CLI to create a new project:

dotnet new webapi -n AiModelApi
cd AiModelApi

3. Add ML.NET NuGet Package

Install the necessary ML.NET package:

dotnet add package Microsoft.ML

4. Define Input and Output Schemas

Create classes that represent the input data and predicted output of your model. For example:

// Models/InputData.cs
public class InputData
{
    [LoadColumn(0)]
    public float Feature1 { get; set; }
    [LoadColumn(1)]
    public float Feature2 { get; set; }
}

// Models/Prediction.cs
public class Prediction
{
    [ColumnName("Score")]
    public float PredictedValue { get; set; }
}

5. Load the Model and Create an Endpoint

Modify your Controllers/PredictionController.cs:

using Microsoft.AspNetCore.Mvc;
using Microsoft.ML;
using Microsoft.ML.Data;
using AiModelApi.Models; // Assuming your models are in this namespace
using System;
using System.IO;

namespace AiModelApi.Controllers
{
    [ApiController]
    [Route("[controller]")]
    public class PredictionController : ControllerBase
    {
        private readonly MLContext _mlContext;
        private readonly PredictionEngine _predictionEngine;

        public PredictionController()
        {
            _mlContext = new MLContext();
            var modelPath = Path.Combine(Directory.GetCurrentDirectory(), "model.zip"); // Path to your saved model

            if (!System.Exists(modelPath))
            {
                throw new FileNotFoundException($"Model file not found at {modelPath}");
            }

            ITransformer loadedModel = _mlContext.Model.Load(modelPath, out var modelSchema);
            _predictionEngine = _mlContext.Model.CreatePredictionEngine(loadedModel);
        }

        [HttpPost]
        public IActionResult Predict([FromBody] InputData input)
        {
            if (input == null)
            {
                return BadRequest("Invalid input data.");
            }

            try
            {
                var prediction = _predictionEngine.Predict(input);
                return Ok(prediction);
            }
            catch (Exception ex)
            {
                return StatusCode(500, $"An error occurred during prediction: {ex.Message}");
            }
        }
    }
}

6. Place Your Model File

Copy your trained model.zip file into the root directory of your API project (where the .csproj file is). Ensure it's included in the build output by checking your project file for something like:

<ItemGroup>
  <Content Include="model.zip" CopyToOutputDirectory="PreserveNewest" />
</ItemGroup>

7. Run and Test Your API

Run the API using the .NET CLI:

dotnet run

You can then use tools like Postman or cURL to send POST requests to http://localhost:5000/prediction with a JSON body matching your InputData schema.