Explore the SN-MSA Patch Discriminator architecture for suitable Discriminator in Transformer Generators

Index

Abstract

In the past Jiang et al.[1] created GAN architecture just using Transformers, and it’s conclusion was that the transformers was not suitable for the Discriminator.

Based on this, Durall et al.[2] explored the suitable CNN-based Discriminators for Transformers Generators in SOTA architecture. and found out that SN GAN Discriminator has good performance

However Durall et al.[2] was proceed on a simple search Therefore, it seems that there is a lack of understanding of why SN-GAN Discriminator works well.

So, we propose to examine the correlation between SN-Patch GAN Discriminators and Transformers Generators.

Furthermore, we propose Hybrid-Discriminators that can solve the fundamental reason why Transformers Discriminator was not suitable.

연구 주제 요약

• Transformer Discriminator도 장점이 존재할 것입니다.
• CNN-based Discriminator가 좋은 이유에 대해 새로운 고찰이 필요합니다.
• Patch GAN Discriminator를 사용하는 것이 가장 성능이 좋을 것입니다.
• Spectral Normalization은 성능향상에 매우 도움이 될 것입니다.