The two main phases of AI development — researching your models and deploying them to production — usually require separate frameworks. You might perform your research and make your improvements using Python-based PyTorch and lots of computational power, then use Caffe2 to deploy a lightweight model to production that will run on lower-power devices like smartphones. Each iteration between the frameworks involves steps, tools and time.
The Qualcomm® Snapdragon™ mobile platform is announcing support of PyTorch 1.0, the latest version of Facebook’s open source AI framework. PyTorch 1.0 is designed to accelerate the research-to-production cycle and help you speed up AI development. It combines the production-oriented features of Caffe2 and ONNX with the research-focused design of previous versions of PyTorch.
PyTorch 1.0 — A competitive advantage in AI development
AI developers will be able create prototypes rapidly and optimize performance without having to move their projects between frameworks — a competitive advantage in a fast-evolving area.
Available now in preview-release, PyTorch 1.0 allows you to quickly prototype using Python or C/C++ to develop and improve your machine learning model. Then, using the same framework, you can execute the model natively, distributed across multiple nodes. PyTorch 1.0 scales from portable mobile devices to high-tier datacenter servers.
The hybrid front end in PyTorch 1.0 offers easy transition between imperative and declarative execution modes. Instead of needing one framework for training/research and another framework for production, you can easily go from research to production in a single framework.
The technology in PyTorch 1.0 is already at work in Facebook products and services, such as performing billions of text translations every day. And the CycleGAN project from the Berkeley AI Research Lab at University of California Berkeley has used PyTorch for machine learning that translates impressionist paintings into photos, transfigures objects between zebras and horses, and modifies landscapes into summer and winter.
Ongoing support from Qualcomm Developer Network
We at Qualcomm Developer Network know that developers and the software community are vital to the consumer adoption of AI. To help you tap into the heterogeneous computing power of the Snapdragon mobile platform, we continue to build support for as many frameworks as possible into our deep learning tools, such as the Qualcomm Neural Processing SDK.
We have long supported the Caffe2 and Tensorflow frameworks and we were one of the first mobile silicon manufacturers to support the ONNX initiative in October 2017. To continue our relationship with Facebook and keep the AI ecosystem moving forward, we will support PyTorch 1.0 because we see it as a fast, clear path that AI engineers and scientists can take from research to production.
See for yourself
Facebook is hosting the first PyTorch Developer Conference in San Francisco. We’ll be on hand as Facebook brings together top AI developers from academia, enterprise, and the developer community to explore where PyTorch 1.0 can take us.
Have a look at PyTorch 1.0 to see how it fits in with your AI development. It’s designed to let you spend more time in your research and production phases and less time plodding back and forth between them.