ViT-GAN: Using Vision Transformer as Discriminator with Adaptive Data Augmentation

摘要

These days, attention is thought to be an efficient way to recognize an image. Vision Transformer (ViT) uses a Transformer for images and has very high performance in image recognition. ViT has fewer parameters than Big Transfer (BiT) and Noisy Student. Therefore, we consider that Self-Attention-based networks are slimmer than convolution-based networks. We use a ViT as a Discriminator in a Generative Adversarial Network (GAN) to get the same performance with a smaller model. We name it ViT-GAN. Besides, we find parameter sharing is very useful to make parameter-efficient ViT. However, the performances of ViT heavily depend on the number of data samples. Therefore, we propose a new method of Data Augmentation. Our Data Augmentation, in which the strength of Data Augmentation varies adaptively, helps ViT for faster convergence and better performance. With our Data Augmentation, we show ViT-based discriminator can achieve almost the same FID but the number of the parameters of the discriminator is 35% fewer than the original discriminator.