Optimizing the Latent Space of Generative Networks