PyTorch 是一个开放源代码深度学习框架,可加速从研究到生产的路径。Microsoft 的数据科学家使用 PyTorch 作为主要框架,以开发在 Microsoft 365、必应、Xbox 等中启用新体验的模型。作为 PyTorch 生态系统的主要参与者,Microsoft 最近所做的贡献包括 PyTorch Profiler。
Azure 上的 PyTorch - 相辅相成
生产就绪
使用 Azure 机器学习中的内置 PyTorch 环境可靠、大规模地训练和部署模型,以确保通过适用于 PyTorch 的 Azure 容器完全支持最新的 PyTorch 版本。
增强的生态系统
通过丰富的 PyTorch 工具和功能生态系统(包括 PyTorch Profiler)实现更多目标。
受到各种规模公司的信赖
Yuji Fukaya,AI 咨询组 AI 转型中心经理,Information Services International-Dentsu"Other deep learning frameworks and cloud services are out there, but we think Azure, Azure Machine Learning, and PyTorch are the best choices because they enhance accuracy, efficiency, scalability, and speed of development."

Jeremy Jancsary,Nuance 高级首席研究科学家"The new enterprise-level offering by Microsoft closes an important gap. Serving PyTorch models in production can be a challenge. The direct involvement of Microsoft lets us deploy new versions of PyTorch to Azure with confidence."

Alexander Vaagan,Inmeta (Crayon 的一部分)首席数据科学家"I would recommend the Azure environment to other developers. It's user-friendly, easy to develop with, and very importantly, it follows best practices for AI and machine learning work."

Pablo Castellanos Garcia,Wayve 工程部副总裁"Running PyTorch on Azure gives us the best platform to build our embodied intelligence. It's easy for our engineers to run the experiments they need, all at once, at petabyte scale."

Solliance CEO 和 Baseline CTO Zoiner Tejada"With Azure AI and PyTorch, we combined focused applications of AI with journalistic processes and financial intelligence, yielding a solution that is unique in the market and valuable for cryptocurrency investors."

Tom Chmielenski,Bentley 首席 MLOps 工程师"We use Azure Machine Learning and PyTorch in our new framework to develop and move AI models into production faster, in a repeatable process that allows data scientists to work both on-premises and in Azure."

Microsoft 是 PyTorch 开源项目生态系统的积极参与者

PyTorch Profiler
PyTorch Profiler 是一种开放源代码工具,可帮助你了解模型中各种 PyTorch 操作的硬件资源(如时间和内存)消耗情况,并解决性能瓶颈。这使模型的执行速度更快、开销更小。

PyTorch 上的 ONNX 运行时
随着深度学习模型不断增大,减少训练时间在财务和环境方面都成为了一个问题。ONNX 运行时通过一行代码更改加速了 PyTorch 转换器模型的大规模分布式训练。结合 DeepSpeed 进一步提高 PyTorch 的训练速度。

Windows 上的 PyTorch
Microsoft 会维护 Windows 的 PyTorch 版本,以便你的团队能够体验经过良好测试且稳定的版本、简单可靠的安装、快速入门和教程,以及对高性能和更高级功能(如分布式 GPU 训练)的支持。

PyTorch Foundation
随着 PyTorch 对 AI 研究和生产的重要性越来越高,Mark Zuckerberg 和 Linux Foundation 共同宣布 PyTorch 将过渡为 Linux Foundation,以支持社区继续发展,并为其提供一个在未来几年蓬勃发展的家园。为了对 PyTorch 的未来发展做出贡献,Microsoft 作为管理委员会成员加入了 PyTorch Foundation,以领导 AI/ML 的民主化和协作。请阅读 Meta 的博客文章,以详细了解 PyTorch Foundation并探索最新的 PyTorch 功能。
ONNX 运行时: 用于加快推理和训练 PyTorch 模型的运行时,支持 Windows、Mac、Linux、Android 和 iOS,并针对各种硬件加速器进行了优化。
DeepSpeed: 用于训练下一代大型模型的算法库,其中包括最先进的模型并行训练算法和分布式训练的其他优化项。
Hummingbird: 可将 Scikit-Learn 或 LightGBM 等传统模型编译为 PyTorch 张量计算,以加快推理速度的库。
利用 Azure 进行 PyTorch 开发的两种方式
使用 Azure 机器学习加快工作流速度
使用适用于 PyTorch 的 Azure 容器轻松生成、训练和部署 PyTorch 模型。它与用于试验管理和完整机器学习生命周期支持的 Azure 机器学习深度集成。Azure 机器学习消除了端到端机器学习工作流的繁重工作,同时可处理数据准备和试验跟踪等管家任务,这会将生产时间从数周缩短到数小时。
使用适用于 PyTorch 的 Azure Data Science Virtual Machine 进行开发
适用于 PyTorch 的 Data Science Virtual Machine 预装了最新的 PyTorch 版本并已经过验证,可降低安装成本并加快价值实现。这些包含有各种优化功能,如 ONNX 运行时、DeepSpeed 和 PySpark,可获得现成的无摩擦开发体验,并能够与包括 GPU 在内的所有 Azure 硬件配置一起工作。
详细了解 Azure 上的 PyTorch
阅读有关 PyTorch 的博客
使用 Azure 在云中提升 PyTorch 项目的完成速度
