Training Shallow and Thin Networks for Acceleration via Knowledge Distillation with Conditional Adversarial Networks