Partial Transfer Learning for Fast Evolutionary Generative Adversarial Networks

Generative Adversarial Networks (GAN) are well known for their capability of generating photo-realistic images or data collections that appear real. One of state-of-the-art approaches in GAN is Evolutionary GAN (E-GAN) which can outperform other GAN methods by leveraging the advantages of evolutionary computing, including population based search, mutation and elitism operators. However evolutionary search is often demanding in terms of resources, e.g. computational power and time. That limits its applicability when resource is limited. Hence we propose Partial Transfer training based E-GAN (PT-EGAN) to improve the efficiency of E-GAN. PT-EGAN aims to train the generator and the discriminator with smaller data sets and transfer the gained features across different stages of training. Our comparative experiments on the CIFAR-10 data set shows that PT-EGAN can reach better performance than E-GAN when using similar resources. Alternatively PT-EGAN requires less resources to achieve a similar performance as E-GAN. PT-EGAN is effective in speeding up generative adversarial learning.

Automated summary

Generative Adversarial Networks (GAN) are known for generating realistic images, with Evolutionary GAN (E-GAN) being a state-of-the-art approach that requires significant computational resources. To improve efficiency, Partial Transfer training based E-GAN (PT-EGAN) is proposed, which uses smaller datasets and transfers learned features across training stages, achieving better performance on CIFAR-10 dataset with similar resources as E-GAN. PT-EGAN is effective in accelerating generative adversarial learning.