Attentive evolutionary generative adversarial network

摘要

Generative adversarial network (GAN) is an effective method to learn generative models from real data. But there are some drawbacks such as instability, mode collapse and low computational efficiency in the existing GANs. In this paper, attentive evolutionary generative adversarial network (AEGAN) model is proposed in order to improve these disadvantages of GANs. The modified evolutionary algorithm is designed for the AEGAN. In the AEGAN the generator evolves continuously to resist the discriminator by three independent mutations at every batch and only the well-performing offspring (i.e.,the generators) can be preserved at next batch. Furthermore, a normalized self-attention (SA) mechanism is embedded in the discriminator and generator of AEGAN to adaptively assign weights according to the importance of features. We propose careful regulation of the generators evolution and an effective weight assignment to improve diversity and long-range dependence. We also propose a superior training algorithm for AEGAN. With the algorithm, the AEGAN overcomes the shortcomings of traditional GANs brought by single loss function and deep convolution and it greatly improves the training stability and statistical efficiency. Extensive image synthesis experiments on CIFAR-10, CelebA and LSUN datasets are presented to validate the performance of AEGAN. Experimental results and comparisons with other GANs show that the proposed model is superior to the existing models.