Facebook Inc. and Microsoft Corp. have teamed up to create an open-source tool that makes it easier for artificial intelligence developers to transfer their projects to different frameworks without having to start over from scratch.

The Open Neural Network Exchange, or ONNX, is intended to be a standardized format that will allow deep learning models trained on one framework to be transferred to another framework with minimal extra work.

“When developing learning models, engineers and researchers have many AI frameworks to choose from,” explained Joaquin Quinonero Candela, director of applied machine learning at Facebook. “At the outset of a project, developers have to choose features and commit to a framework. Many times, the features chosen when experimenting during research and development are different than the features desired for shipping to production.”

Candela said that there is a gap between these operating modes, and the goal behind ONNX is to bridge that gap. Eric Boyd, vice president of AI data and infrastructure at Microsoft, added that each framework offers different benefits, from faster training time to better mobile implementation, and ONNX will allow developers to get the best of both worlds without having to resort to complicated workarounds.

ONNX allows developers to transfer their deep learning models between several popular frameworks, including Cognitive Toolkit, Caffe2 and PyTorch. Candela also invited the developer community to “join the effort and support ONNX in their ecosystem.”

ONNX uses a few different methods to achieve its flexibility. Candela explained that for some frameworks, such as Caffe2, the process was more or less the same as implementing a translator, which translates the syntax for one framework into the syntax for another. However, the process for PyTorch was a little more complicated and required a “tracer” that essentially recorded the execution of a program as it runs and then emulates the same function in the new framework.

According to Candela, the Facebook AI Research team is already using ONNX for some of their implementations, which allows the company to roll out new advancements across multiple projects at the same time.

The source code of ONNX is available on Github.

Photo: Massachusetts General Hospital and Draper Labs [Public domain], via Wikimedia Commons