Azure の PyTorch

クラウドにおける簡単かつシームレスな PyTorch エクスペリエンス

PyTorch is an open-source deep-learning framework that provides a seamless path from research to production. Many AI innovations are developed on PyTorch and quickly adopted by the industry. Microsoft uses PyTorch internally and also actively contributes to development and maintenance of the PyTorch ecosystem.

PyTorch and Azure are better together, making it easy to develop, train, and deploy machine learning models. Get started with PyTorch in Azure Machine Learning, a fully managed service that helps you build and deploy models faster; Azure Data Science Virtual Machines, pre-configured for managing your own infrastructure; and Azure Functions, a serverless hosting environment.

Why use PyTorch on Azure

Production-ready

Improve your team's productivity with Azure Machine Learning. Train models reliably at scale using built-in PyTorch environment. Deploy to production with DevOps for PyTorch.

Accelerated performance

Decrease your time to market with powerful GPU hardware, production-grade software accelerator (ONNX Runtime), and latest innovative scaling techniques (DeepSpeed) in Azure.

Robust ecosystem

Achieve more with the rich ecosystem of tools and models in PyTorch actively supported by Microsoft.

Astrazeneca

"Scientists have to comb through massive amounts of data to deliver life-changing medicines. Microsoft and PyTorch are helping global biopharmaceutical company AstraZeneca to accelerate its drug discovery research."

Microsoft is also an active contributor to an ecosystem of PyTorch open-source projects:

PyTorch on Windows

Microsoft maintains PyTorch builds for Windows so users can enjoy well-tested and stable builds, simple and reliable installation, quick-starts and tutorials, and high performance and support for more advanced features such as distributed GPU training.

torch.onnx

Microsoft maintains the ONNX export functionality in core PyTorch. PyTorch has built-in support for ONNX that enables deploying models with higher performance on multiple platforms.

PyTorch Profiler (release with PyTorch 1.8)

Microsoft developed the PyTorch Profiler in collaboration with Facebook to enable users to easily find and remove performance bottlenecks in the models.

Microsoft is also an active contributor to an ecosystem of PyTorch open-source projects:

ONNX Runtime: runtime for accelerated inferencing and training of PyTorch models, supporting Windows, Mac, Linux, Android, and iOS and optimized for a variety of hardware accelerators.

DeepSpeed: a library of algorithms for training of next-generation large models, including state-of-the-art model-parallel training algorithms and other optimizations for distributed training.

Hummingbird: a library that compiles traditional models like Scikit-Learn or LightGBM into PyTorch tensor computation for faster inference.

FastFormers: a framework that provides hyper-efficient inferencing of large transformer models by applying fine-tuned recipe of knowledge distillation, pruning, quantization, and other techniques.

PyTorch 開発用に Azure を使用する 3 つの方法

Accelerate your workflow with Azure Machine Learning

Build, train, and deploy PyTorch models with ease. Enable productivity and collaboration using machine learning workspaces and familiar tools, such as Jupyter Notebooks and Visual Studio Code. Azure Machine Learning removes the heavy lifting of end-to-end machine learning workflows while also handling housekeeping tasks such as data preparation and experiment tracking, which cuts time to production from weeks to hours.

Develop with preconfigured Azure Data Science Virtual Machines

Go straight to development with custom Windows or Linux virtual machines specially configured for machine learning workloads. Data Science Virtual Machines come installed with PyTorch as well as the necessary GPU drivers and a comprehensive suite of other popular data science tools. Get a frictionless development experience out of the box and the ability to work with all Azure hardware configurations including GPUs.

Go serverless with Azure Functions

Deploy lightweight models with the most efficiency by utilizing low-maintenance serverless hosting provided by Azure Functions.

Get Started with PyTorch on the AI Show

Learn the basics of PyTorch, including how to build and deploy a model and how to connect to the strong community of users.

DeepSpeed—PyTorch Developer Day

In this talk, Yuxiong He, a partner research manager at Microsoft, presented DeepSpeed, an open-source deep-learning training optimization library compatible with PyTorch.

PyTorch on Windows

Maxim Lukiyanov, a product manager for the Azure AI platform, described improvements within the Windows platform support made in PyTorch version 1.7.

PyTorch, the PyTorch logo and any related marks are trademarks of Facebook, Inc.

Accelerate your PyTorch project in the cloud with Azure