ASCNet: Self-supervised Video Representation Learning with Appearance-Speed Consistency.