High-load face recognition on the photos of the social network users. We are no longer surprised by face recognition unless we have:
- 330 millions of user accounts;
- 20 millions of user photos uploaded every day;
- max 0.2 seconds for one photo processing (spoiler: we've managed to do it even faster);
- limited hardware.
In this talk we'll look at:
- the pipeline for building users vectors and searching for users on the uploaded photos;
- neural network training: building dataset > training > building dataset > cook until ready;
- the facial detector on the neural networks cascade and its optimization;
- building rescaled user vector on GPU;
- hardware and optimizations, running in the cloud, fault tolerance.