To appear as a paper in ICLR. We design a class of neural networks to be input-convex, which achieves excellent learning plus control performances on tasks such as Mujoco locomotion and building energy management.