Building your own Self-attention GANs

Building your own Self-attention GANs

A PyTorch implementation of SAGAN with MNIST and CelebA dataset

GANs, as known as Generative Adversarial Networks , is one of the most popular topics in the Machine Learning fields recently. It consists of two different Neural Network models, one called Generator , and one called Discriminator . It sounds hard to understand, but let me try to put it in this way: Let’s say we want to forge famous paintings starting without any knowledge of painting, what should we do? Most would say, just look at the paintings and learn how to do it. But it’s not a one-man job, to some point, I am sure you will be better and better in painting. You will need to have your friend come in front of one real painting and one that you forged and let him guess which one is the real one. It will be pretty easy for him to guess in the beginning, but keep it going and you will eventually confuse your friend.

In GANs, the generator is like you who forge paintings, and the discriminator is the friend who specializes in telling which painting is fake. Think about the goal here, you want to make it hard for your friend to tell real or fake. If your friend were to give out a probability of being real from 0 to 1 for each painting, you would want him to give 0.5 to any paintings you show him, either real or forged. This will also be the objective of GANs, as reflected in loss functions.

We also see DCGAN a lot, which stands for Deep Convolutional GAN . It is a GAN design that specialized in image generation, using convolution layers for both generator and discriminator . It works just like a CNN . A Self-attention GAN is a DCGAN that utilizes self-attention layers. The idea of self-attention has been out there for years, also known as non-local in some researches. Think about how does convolution works: they convolve nearby pixels and extract features out of local blocks. They work “locally” in each layer. In contrast, self-attention layers learn from distant blocks. In 2017, Google published a paper “ Attention Is All You Need ”, bringing more hypes about the topic. For a single image input, it works like this:

Related Articles

Search on MC.AI