Train AI models in Azure Batch AI

  • Artículo

Batch AI is a managed service that enables data scientists and AI researchers to train AI and other machine learning models on clusters of Azure virtual machines, including VMs with GPU support. You describe the requirements of your job, where to find the inputs and store the outputs, and Batch AI handles the rest. Learn more about Azure Batch AI

It's integrated with Visual Studio Tools for AI so you can dynamically scale out training models in Azure. Once you've installed Visual Studio Tools for AI, it's easy to create a new Python project using pre-made recipes in the Azure Machine Learning Sample Gallery.

  1. Launch Visual Studio. Open the Server Explorer by opening the AI Tools menu and choosing Select Cluster

    Cluster chooser

  2. Expand AI Tools. Any Batch AI resources you have will be auto-detected and appear in the Server Explorer.

    Screenshot of the expanded folder tree for AI Tools in Server Explorer, showing expanded subfolders for Azure Batch AI and Azure Machine Learning.

  3. Select View > Team Explorer... to open the Team Explorer window in which you can connect to GitHub or Azure DevOps, or clone a repository.

    Team explorer window showing Azure DevOps, GitHub, and cloning a repository

  4. In the URL field under Local Git Repositories, enter https://github.com/Microsoft/samples-for-ai, enter a folder for the cloned files, and select Clone.

    Tip

    The folder you specify in Team Explorer is the specific folder to receive the cloned files. Unlike the git clone command, creating a clone in Team Explorer does not automatically create a subfolder with the name of the repository.

  5. When cloning is complete, click File > Open Solution > Project / Solution

    Screenshot showing part of the Server Explorer File menu with the Open command selected and Project/Solution selected on the context menu.

  6. Open samples-for-ai\TensorFlowExamples\TensorFlowExamples.sln in the directory you cloned the repository

    Screenshot showing the solution file TensorflowExamples.sln listed in the contents of the TensorflowExamples folder in the samples-for-ai repository.

  7. Set MNIST project as the Startup Project

    Screenshot showing Set as StartUp Project selected on the context menu for the MNIST project in Solution Explorer.

  8. Right-click MNIST project, Submit Job

    Screenshot showing Submit Job selected on the context menu for the MNIST project in Solution Explorer.

  9. Select your Azure Batch AI cluster, then click Import. Select the AzureBatchAI_TF_MNIST.json file to quickly populate some default values like which Docker Image to use. Then click Submit

    Screenshot of the Submit Job dialog with values populated for Use cluster, Startup script, Job name, Image name, StdOutErr path prefix, and CLI parameters.