基于迁移学习的双层生成式对抗网络

摘要

在生成式对抗网络的对抗训练中,目标样本训练集不足会导致模型无法准确学习到对应的特征,但对于需要人工制作、标记的目标样本训练集又很难获取.针对这一问题,提出了基于迁移学习的双层生成式对抗网络模型,在第一层网络中通过伪目标样本让模型学习到目标样本在结构空间的大致分布后,利用迁移学习的思想进行模型迁移,并在第二层网络中根据少量目标样本进行调整.实验中,验证了该模型在中文字体生成与图片框架图转换中的提高,有效地在少量目标样本训练集中训练出更好的模型.%In the confrontation training of the Generative Adversarial Networks(GAN), insufficient training set of target samples will result in the model not being able to accurately learn the corresponding features, but it is difficult to obtain a target sample training set that needs to be manually produced and marked. Aiming at this problem, a two-layer GAN model based on migration learning is proposed. In the first layer network, the pseudo-target samples are used to let the model learn the approximate distribution of the target samples in the structure space, and the model migration is carried out by using the idea of migration learning, and is adjusted according to a small number of target samples in the second layer network. In experiment, the improvement of the model in Chinese font generation and picture frame graph conversion is verified, and a better model is effectively trained in a small number of target sample training sets.