F8 2017 Facebook has open sourced Caffe2, the toolbox of deep-learning software its own developers use to train AI models and build apps.
Caffe2 appears to be product driven, and is geared towards deploying machine-learning systems into smartphone applications and onto large-scale clusters. It differs from PyTorch, another software framework from Facebook that is more research oriented as it allows programmers to experiment with different neural network architectures more easily.
Facebook has been working with Nvidia to integrate Caffe2 into the graphics chip giant's deep-learning developer libraries so the framework can take advantage of hardware acceleration on Nv's GPUs. We're told, for example, Caffe2 is nippy on Facebook’s Big Basin OpenCompute AI servers that pack 64 Nvidia Tesla P100 GPUs. Intel, Microsoft and Amazon have also stepped up to make sure Caffe2 is optimized for their systems and services.
Meanwhile, Qualcomm’s Neural Processing Engine (NPE) software development kit supports Caffe2 and Google’s TensorFlow. This library glues software to the neural network math unit built into its top-end Snapdragon 835 system-on-chip, which is appearing in smartphones, tablets and notebooks this year. Code using the engine will therefore gain a performance boost via the on-chip acceleration. The NPE dev kit will be available in July.
It’s a sensible move, considering Qualcomm designs chips used in millions and millions of Android devices, and Facebook is investing heavily in augmented reality, a technology that will manipulate photos and video taken with smartphones. That's going to require some machine-learning processing to identify objects and meddle with them – hence the marriage between Caffe2, Qualcomm, and Facebook.
And, hey, if that means developers using Facebook's tech to produce apps to rival SnapChat – the image-fiddling toy that's cool with the kids and in the way of Mark Zuckerberg's world domination plans – so much the better.
“Augmented reality is going to help us mix the digital and physical in all new ways and that's going to make our physical reality better," Facebook CEO Zuckerberg told his social network's F8 conference in Silicon Valley on Tuesday.
Caffe2 doesn't just add cartoon doodles over images. It's hoping to be more general purpose, allowing developers to create chatbots, hook up IoT devices, use machine translation and speech, and image classification algorithms for medical applications.
Caffe, a predecessor of Caffe2, was developed by Yangqing Jia while he was a PhD student at the University of California, Berkeley. Jia is now a research scientist and leads Facebook’s efforts in building a general platform for its AI applications. ®