Facebook wants to make sure the open-source PyTorch machine-learning framework supports the needs of developers who want to use its AI models in production systems, not just research projects, it plans to announce Wednesday at F8.
Mark Zuckerberg’s surveillance factory plans to release PyTorch 1.0 Wednesday, which was designed to incorporate some of the production-oriented qualities of the other open-source AI frameworks developed in part by Facebook, Caffe2 and ONNX. Over the next few months, Facebook plans to merge the PyTorch and Caffe2 code bases, and it also announced that cloud heavyweight Amazon Web Services and Microsoft Azure plan to support PyTorch 1.0 in their cloud AI services.
PyTorch is known for its powerful research capabilities, as developers used the project to experiment with the machine-learning models best suited for their applications. In a field like artificial intelligence that is still fairly underdeveloped despite years of hype, it’s hard to know at first which machine-learning models will work best for certain applications, and landing on the right model as soon as possible prevents a lot of wasted effort.
But PyTorch wasn’t nearly as suitable for production-grade use of AI, such as Facebook’s own use of neural networks, for which Caffe2 is the engine. ONNX was designed to help developers bridge that gap, but it still required more manual labor that developers would have liked.
That means PyTorch is getting an upgrade, and the 1.0 version will allow developers to share code more easily between research and production environments. The new version will be in beta mode for several months, Facebook said.
