Effective Training and Efficient Inference of Deep Neural Networks for Visual Understanding