Traffic prediction is practically important to facilitate many real applications in urban areas such as relieving traffic congestion. Traditional traffic prediction models are mostly statistic based methods, and they cannot effectively capture the nonlinear, stochastic and time-varying characteristics of the urban transportation systems. Another limitation of these methods is that they usually focus on analyzing one or several roads or road segments, but are not capable to predict the traffic conditions of all the road segments in a large transportation network of a city as a whole. Therefore, in recent years, deep neural network based methods for forecasting the road network-scale traffic have been emphasized greatly. However, most existing deep neural network methods model the traffic data of a road network as "images" rather than graphs, and thus they suffer from the blurry prediction issue and do not perform well on the task of multi-step traffic prediction. In this paper, We propose a network-scale deep traffic prediction model called GCGAN by combining adversarial training and graph CNN. Specifically, we propose a Generative Adversarial Net based prediction framework to address the blurry prediction issue by introducing the adversarial training loss. To predict the traffic conditions in multiple future time intervals simultaneously, we design a sequence to sequence (Seq2Seq) based encoder-decoder model as the generator of GCGAN. To fully capture the spatial correlations among the road segments of a transportation network, we propose to apply a graph convolution network (GCN) in both generator and discriminator of GCGAN for feature learning. We evaluate our proposal over a large real traffic dataset in the arterial road network of downtown Chicago. The results show that GCGAN significantly outperforms both traditional statistic based methods and recent state-of-the-art deep learning methods.