The new open-source ONNX format will help developers move between AI frameworks with greater ease
Apart from powering smart speakers and powering intelligent services, artificial intelligence (AI) also has a knack of bringing technology giants together.
On Sept. 7, Microsoft and Facebook announced Open Neural Network Exchange (ONNX), a format for deep learning models that encourages interoperability between AI frameworks. ONNX was do-developed by the two companies as an open-source project and are now encouraging developers to help extend the technology by visiting the project’s page on GitHub and downloading its code and related documentation.
“ONNX provides a shared model representation for interoperability and innovation in the AI framework ecosystem,” wrote Eric Boyd, corporate vice president of AI Data and Infrastructure at Microsoft, in a blog post. “Cognitive Toolkit, Caffe2, and PyTorch will all be supporting ONNX.”
AI framework interoperability
Currently, many AI frameworks offer similar capabilities, but each framework uses a different format for the computation graphs that depict neural networks, explained Boyd.
The ONNX standard will allow developers to move between frameworks with greater ease, encountering less friction as they settle on the frameworks that offer them the ideal functionality and optimizations for their applications.
There are perks for hardware vendors as well, Boyd said.
By targeting ONNX, hardware makers can effect multiple frameworks with their neural network optimizations in one fell swoop. “Frequently optimizations need to be integrated separately into each framework which can be a time-consuming process. The ONNX representation makes it easier for optimizations to reach more developers,” explained Boyd.
At Facebook, ONNX is helping to narrow the gap between the social media giant’s research and its commercialization efforts, pushing AI innovations through the company’s product pipeline faster, explained Joaquin Quinonero Candela, director of Applied Machine Learning, at Facebook, in a separate blog post.
“We can now export models for many common neural networks from PyTorch and deploy them on Caffe2. This is the first step in enabling us to rapidly move our latest research developments into production,” stated the Facebook executive. “Over the coming months, we will be enhancing ONNX and releasing improvements to Caffe2 and PyTorch that enable them to interoperate more deeply.”
Caffe2, an Facebook open source project, is a framework for modular and scalable deep learning systems. PyTorch is a Python-based framework that uses GPUs (graphics processing units) to accelerate deep learning workloads.
Cognitive Toolkit, Microsoft’s open-source AI software toolset, will support ONNX in an upcoming release, Boyd said. Cognitive Toolkit version 2.0 (formerly Computational Network Toolkit or CNTK) was released on June 1 and includes a deep learning system based on image and speech recognition technologies developed by Microsoft Research. With Facebook’s help, the software maker also plans to contribute tools and reference implementations, among other resources, said Boyd.
Facebook isn’t the only company forming AI alliances with Microsoft.
Amazon and Microsoft announced on Aug. 30 that they were opening up the lines of AI communication by allowing their respective virtual assistant platforms, Alexa and Cortana, to talk to one another. Industry watchers expect the integration to give the AI and voice-enabled commerce markets at boost.
Originally published on eWeek