Learning Versatile Filters for Efficient Convolutional Neural Networks