Accuracies of most fingerprinting approaches for WiFi-based indoor localization applications are affected by the qualities of fingerprint databases, which are time-consuming and labor-intensive. Recently, many methods have been proposed to reduce the localization accuracy reliance on the qualities of the established fingerprint databases. However, studies on establishing fingerprint databases are relatively rare under the condition of sparse reference points. In this paper, we propose a novel data augmenter based on the adversarial networks to build fingerprint databases with sparse reference points. Additionally, two conditions of these networks are designed to generate data effectively and stably, which are 0-1 sketch and Gaussian sketch. Based on the networks, we design two augmenters with different cyclic training strategies to evaluate the augmenting effects comparatively. Meanwhile, five quantitative evaluation metrics of the augmenters are proposed from two perspectives of the artificial experiences and the data features, and some of them are also used as the gradient penalties for generators. Finally, experiments corresponding to these metrics and localization accuracies demonstrate that the data augmenter with the 0-1 sketch adversarial network is more efficient, effective and stable totally