Towards Target-Driven Visual Navigation in Indoor Scenes via Generative Imitation Learning