Hourglass-AVSR: Down-Up Sampling-based Computational Efficiency Model for Audio-Visual Speech Recognition.