Content loss and conditional space relationship in conditional generative adversarial networks

Authors: ENES EKEN

Abstract: In the machine learning community, generative models, especially generative adversarial networks (GANs) continue to be an attractive yet challenging research topic. Right after the invention of GAN, many GAN models have been proposed by the researchers with the same goal: creating better images. The first and foremost feature that a GAN model should have is that creating realistic images that cannot be distinguished from genuine ones. A large portion of the GAN models proposed to this end have a common approach which can be defined as factoring the image generation process into multiple states for decomposing the difficult task into several more manageable sub tasks. This can be realized by using sequential conditional/unconditional generators. Although images generated by sequential generators experimentally prove the effectiveness of this approach, visually inspecting the generated images are far away of being objective and it is not yet quantitatively showed in an objective manner. In this paper, we quantitatively show the effectiveness of shrinking the conditional space by using the sequential generators instead of utilizing single but large generator. At the light of the content loss we demonstrate that in sequential designs, each generator helps to shrink the conditional space, and therefore reduces the loss and the uncertainties at the generated images. In order to quantitatively validate this approach, we tried different combinations of connecting generators sequentially and/or increasing the capacity of generators and using single or multiple discriminators under four different scenarios applied to image-to-image translation tasks. Scenario-1 uses the conventional pix2pix GAN model which serves as the based line model for the rest of the scenarios. In Scenario-2, we utilized two generators connected sequentially. Each generator is identical to the one used in Scenario-1. Another possibility is just doubling the size of a single generator which is evaluated in the Scenario-3. In the last scenario, we used two different discriminators in order to train two sequentially connected generators. Our quantitative results support that simply increasing the capacity of one generator, instead of using sequential generators, does not help a lot to reduce the content loss which is used in addition to adversarial loss and hence does not create better images.

Keywords: Generative adversarial networks, conditional space, content loss, sequential generators, image-to-image translation

Full Text: PDF