Though Facebook’s focus on day 1 of its F8 conference centered on the company’s recent struggles and their relationship with the phrase “taking broader responsibility,” day 2 shifted most of the pizazz to the technical advances its giant team has made over the past year.
Today, the company announced PyTorch 1.0, a new iteration of the framework that merges Python-based PyTorch with Caffe2 allowing developers to move from research to production in a more frictionless way without having to deal with migration.
At Facebook, the company’s AI efforts are split between two teams, the Facebook AI Research group (FAIR) and the company’s Applied Machine Learning team (AML). The distinction ultimately boils down to one division researching AI with seemingly limitless computational resources at their disposal and the other looking to implement lightweight machine learning models more suited for consumers. In the past, the former mission has been better-suited for the research-optimized PyTorch while the more resource-efficient Caffe2 framework has made sense for the latter.
Moving between these frameworks through research and production phases has been a bit of a headache for Facebook. PyTorch 1.0 fuses the best of both worlds with immediate and graph execution modes that should continue its research-focused prowess while gaining some optimizations for production.
“Instead of needing a framework for training/research and a framework for production, now you only need to use PyTorch 1.0 and you can seamlessly go from research to production in one framework,” a Facebook spokesperson told TechCrunch.
Facebook collaborated with Microsoft last year to create the Open Neural Network Exchange (ONNX) which was designed to make movements between frameworks more simple. They’ve now integrated ONNX into PyTorch 1.0 so that the models can be interoperable with other frameworks and developers can “mix-and-match.”
The company also says that it will be unifying the PyTorch 0.4 and Caffe2 codebases over the next several months to create a unified framework that supports several features including efficient graph-mode execution with profiling, mobile deployment and extensive vendor integrations.
PyTorch 1.0 will be released in beta later this year. The company says that Microsoft is planning to support the framework in Azure and that Amazon will support PyTorch 1.0 across their cloud products.