Self-supervised Video Representation Learning with Cross-Stream Prototypical Contrasting.