F8 Facebook announced Pytorch 1.0, an updated version of the popular AI framework Pytorch, that aims to make it easier for developers to use neural network systems in production.
On the second day of its developer conference F8 in San Jose, California, CTO Mike Schroepfer, introduced Pytorch 1.0, and said it combines Pytorch, Caffe 2, with Open Neural Network Exchange (ONNX).
Pytorch 1.0 will let developers use their tools of choice and run models on their cloud of choice at peak performance, Schroepfer said. Microsoft and Amazon are, apparently, planning to support Pytorch 1.0 for Azure and AWS.
It’s already deployed in some Facebook’s services such as its machine translation system. But Pytorch is not quite ready to be released to the public yet. The beta version is expected to be published later this year.
Facebook opens up Big Basin Volta plans to share the server wealthREAD MORE
Pytorch is commonly used by AI researchers to build neural networks easily in Python. Creating small models for research is pretty simple, but trying to implement them on scale for production is much trickier. The Pytorch code has to be converted to a graph mode representation using Caffe 2, another framework that uses a mixture of Python and C++, since executing Python is painfully slow.
“The migration from PyTorch to Caffe2 to ship to production used to be a manual process, time-intensive and overall error-prone,” according to a Facebook blog post today.
Pytorch 1.0 now includes a third step with ONNX. ONNX started off as a project between Facebook, Microsoft, Amazon, AMD, IBM, Huawei and Qualcomm. The goal was to develop a format that allows neural networks trained in one framework to be transferred to another for the inference stage.
It makes it easier to export trained Pytorch models to Caffe 2 for inference and deployment with large-scale servers or on mobile phones.
Alongside the Pytorch 1.0 announcement, a list of other machine learning tools are being open sourced. These include a PyTorch Language Library for language translation; ELF, an AI research platform for games; Glow, a new compiler to speed up models across different hardware; Tensor Comprehensions, a tool that generates GPU code from mathematical operations; and finally Detectron, an object detection model. ®