Facebook Inc. today headlined its annual F8 developers conference in San Jose, California, with an update of PyTorch, its popular artificial intelligence software framework used by developers to accelerate the deployment of AI-based applications.

The PyTorch platform has rapidly become a big hit with AI developers since its launch back in 2017, as it enables them to not only research and build AI models for their apps, but also move those apps into production thanks to its integration with leading public cloud platforms.

It was first built by Facebook’s AI research group as a machine learning library of functions for the programming language Python. It’s primarily designed for use with deep learning, which is a branch of machine learning that attempts to emulate the way the human brain functions, and has led to major breakthroughs in areas such as language translation, image and voice recognition.

In a blog post announcing the release of PyTorch 1.1, Facebook engineers Joe Spisak, Soumith Chintala, Dmytro Dzhulgakov, Lin Qiao and Greg Chanan talked about how the company has used the platform to deploy its translation and natural language processing services at scale. Those services now carry out about 6 billion translations per day in applications such as Facebook Messenger, the engineers said.

“PyTorch’s unified framework has allowed us to iterate our ML systems more quickly,” Facebook’s engineers wrote.

It’s not only Facebook that’s using PyTorch, though. In fact, the platform has seen adoption grow on a scale that few other open-source technologies can match. And this adoption is being driven by some of the biggest companies around.

“We’ve see a big increase in the use of PyTorch in production applications — inside Facebook and at other companies,” Srinivas Narayanan, head of Facebook AI applied research, told SiliconANGLE.

Microsoft Corp., for example, uses PyTorch as the foundation for developing many of its own machine learning models, which are then deployed across its ONNX Runtime framework to power its Cognitive Services. Airbnb Inc. uses it to design and build its conversational AI tools for customer services, and the Toyota Research Institute is developing new safety systems for autonomous vehicles with it.

PyTorch 1.1 gets usability, performance boost

With such rapid adoption, it’s only natural for Facebook to step things up with regards to PyTorch’s development, and the latest release comes with several new features, including better visualization tools to improve usability and others aimed at boosting its performance.

These additions include support for TensorBoard, which is a web application suite originally designed for the rival TensorFlow AI framework built by Google LLC, which developers can use to inspect and analyze AI model training runs. There’s also an updated just-in-time compilation tool that transforms code into instructions that can be sent directly to a processor, plus new application programming interfaces that add support for custom recurrent neural networks.

On the performance side, Facebook has added a new Distributed Training feature that enables developers to split workloads across multiple graphics processing units. This last feature is especially useful for developers running models where different parameters are used in each iteration.

More interesting than Facebook’s new additions, perhaps, are some of the new projects and tools being added to the PyTorch ecosystem by other members of the community. Although Facebook originally developed PyTorch, the framework is actually open source, which means lots of other companies are also contributing to its success.

Those contributors include Google, which has created a new hosted JupyterLab service on its public cloud platform for use with PyTorch. The AI Platform Notebooks service provides a simpler user interface for developers working with PyTorch and allows them to easily fire up virtual machines to run their AI models. AI Platform Notebooks integrates with Google services such as BigQuery, Cloud Dataproc and AI Factory, which means developers can run their entire experiments without leaving JupyterLab.

Other new features created by third parties include BoTorch, which is a research framework that enables Bayesian optimization to help identify the best models from multiple versions, and Ax, which is used to manage adaptive experiments in machine learning model training to optimize them for various applications and infrastructures.

PyTorch is just one of many open-source AI frameworks, which include the TensorFlow framework developed by Google Inc., MXNet championed by Amazon Web Services Inc. and the CNTK framework developed by Microsoft Research. Each, however, has its advantages, and Narayanan said PyTorch is aimed more specifically at making it as easy and fast as possible for developers to express an idea in software.

Constellation Research Inc. analyst Holger Mueller told SiliconANGLE the new features help to establish PyTorch as one of the most viable platforms for AI development.

“One has to notice the addition of TensorBoard, which is an incline toward the overwhelming success of Google’s TensorFlow,” Mueller said. “Overall, enterprises want to see competition on AI platforms to power their next-generation applications, and Facebook pushing PyTorch is a key next step.”

The full list of PyTorch’s new features and tools is available in Facebook’s blog post.

In addition to the new features, Facebook is also working with several educational firms in order to make PyTorch even more accessible. These include popular online training sites Udacity and Fast.ai, as well as universities such as Stanford NLP and UC Berkeley Computer Vision. There are also several massive open online courses or MOOCs available for developers wishing to learn more about PyTorch.

Although there’s a notion that everyone will benefit from anything that improves machine learning, Facebook’s promotion of PyTorch also could help Facebook attract new, scarce AI talent that already will be trained using its internally developed framework.

With reporting from Robert Hof

Image: Facebook