Reduce Computational Complexity for Convolutional Layers by Skipping Zeros