Discriminator Feature-Based Inference by Recycling the Discriminator of GANs
- PDF / 11,558,346 Bytes
- 23 Pages / 595.276 x 790.866 pts Page_size
- 36 Downloads / 172 Views
Discriminator Feature-Based Inference by Recycling the Discriminator of GANs Duhyeon Bang1 · Seoungyoon Kang1 · Hyunjung Shim1 Received: 29 April 2019 / Accepted: 19 February 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Generative adversarial networks (GANs) successfully generate high quality data by learning a mapping from a latent vector to the data. Various studies assert that the latent space of a GAN is semantically meaningful and can be utilized for advanced data analysis and manipulation. To analyze the real data in the latent space of a GAN, it is necessary to build an inference mapping from the data to the latent vector. This paper proposes an effective algorithm to accurately infer the latent vector by utilizing GAN discriminator features. Our primary goal is to increase inference mapping accuracy with minimal training overhead. Furthermore, using the proposed algorithm, we suggest a conditional image generation algorithm, namely a spatially conditioned GAN. Extensive evaluations confirmed that the proposed inference algorithm achieved more semantically accurate inference mapping than existing methods and can be successfully applied to advanced conditional image generation tasks. Keywords Generative adversarial networks · Inference mapping · Conditional image generation · Quality metric for inference mapping · Spatial semantic manipulation
1 Introduction Generative adversarial networks (GANs) have demonstrated remarkable progress in successfully reproducing real data distribution, particularly for natural images. Although GANs impose few constraints or assumptions on their model definition, they are capable of producing sharp and realistic images. To this end, training GANs involves adversarial competition between a generator and discriminator: the generator learns the generation process formulated by mapping from the latent distribution Pz to the data distribution Pdata ; and the disCommunicated by Jun-Yan Zhu, Hongsheng Li, Eli Shechtman, MingYu Liu, Jan Kautz, Antonio Torralba. Duhyeon Bang and Seoungyoon Kang have contributed equally to this work.
B
Hyunjung Shim [email protected] Duhyeon Bang [email protected] Seoungyoon Kang [email protected]
1
School of Integrated Technology, Yonsei Institute of Convergence Technology, Yonsei University, Seoul, Republic of Korea
criminator evaluates the generation quality by distinguishing generated images from real images. Goodfellow et al. (2014) formulated the objective of this adversarial training using the minimax game min max G
D
E [log(D(x))] + E [log(1 − D(G(z))],
x∼Pdata
(1)
z∼Pz
where E denotes expectation; G and D are the generator and discriminator, respectively; and z and x are samples drawn from Pz and Pdata , respectively. Once the generator learns the mapping from the latent vector to the data (i.e., z → x), it is possible to generate arbitrary data corresponding to randomly drawn z. Inspired by this pioneering work, various GAN models have been developed to improve training stability,
Data Loading...
