Urban planning refers to the efforts of designing land-use configurations given a region. However, to obtain effective urban plans, urban experts have to spend much time and effort analyzing sophisticated planning constraints based on domain knowledge and personal experiences. To alleviate the heavy burden of them and produce consistent urban plans, we want to ask that can AI accelerate the urban planning process, so that human planners only adjust generated configurations for specific needs? The recent advance of deep generative models provides a possible answer, which inspires us to automate urban planning from an adversarial learning perspective. However, three major challenges arise: 1) how to define a quantitative land-use configuration? 2) how to automate configuration planning? 3) how to evaluate the quality of a generated configuration? In this paper, we systematically address the three challenges. Specifically, 1) We define a land-use configuration as a longitude-latitude-channel tensor. 2) We formulate the automated urban planning problem into a task of deep generative learning. The objective is to generate a configuration tensor given the surrounding contexts of a target region. 3) We provide quantitative evaluation metrics and conduct extensive experiments to demonstrate the effectiveness of our framework.