gan图注意力网络_首先深入研究高级GAN：探索自我注意力和频谱规范

gan图注意力网络

by Thalles Silva

由Thalles Silva

首先深入研究高级GAN：探索自我注意力和频谱规范 (Dive head first into advanced GANs: exploring self-attention and spectral norm)

Lately, Generative Models are drawing a lot of attention. Much of that comes from Generative Adversarial Networks (GANs). Invented by Goodfellow et al, GANs are a framework in which two players compete with one another. The two actors, the generator G and discriminator D, are both represented by function approximators. Moreover, they play different roles in the game.

最近，生成模型引起了很多关注。 其中大部分来自生殖对抗网络(GAN)。 由Goodfellow等人发明的GAN是一个框架，其中两个参与者相互竞争。 生成器G和鉴别器D这两个角色都由函数逼近器表示。 而且，他们在游戏中扮演不同的角色。

Given a training data Dt, G creates samples as an attempt to mimic the ones from the same probability distribution as Dt.

给定训练数据Dt ， G创建样本以尝试模仿与Dt相同的概率分布中的样本。

D, on the other hand, is a common binary classifier. It has two main jobs. First, it categorizes whether its received input comes from the true data distribution (Dt) or from the generator distribution. In addition, D also guides G to create more realistic samples by passing to G its gradients. In fact, taking the gradients from D is the only way G can optimize its parameters.

另一方面， D是常见的二进制分类器。 它有两个主要工作。 首先，它对接收到的输入是来自真实数据分布( Dt )还是来自生成器分布进行分类。 此外， D还通过将G的渐变传递给G来指导G创建更逼真的样本。 实际上，从D取梯度是G可以优化其参数的唯一方法。

In this game, G takes random noise as input and generates a sample image Gsample. This sample is designed to maximize the probability of making D mistake it as coming from real training set Dt.

在此游戏中， G将随机噪声作为输入并生成样本图像Gsample 。 此样本旨在最大程度地使D误认为来自真实训练集Dt的可能性。

During training, half of the time D receives images from the training set Dt. The other half of the time, D receives images from the generator network — Gsample. D is trained to maximize the probability of assigning the correct class label to both: real images (from the training set) and fake samples (from G). In the end, the hope is that the game finds an equilibrium — the Nash equilibrium.

在训练期间，时间D的一半从训练集Dt接收图像。 D的另一半时间从生成器网络Gsample接收图像。 对D进行训练，以最大程度地为以下两者分配正确的类别标签：真实图像(来自训练集)和假样本(来自G )。 最后，希望是游戏找到一个平衡点-纳什平衡点。

In this situation, G would capture the data probability distribution. And D, in turn, would not be able to distinguish between real or fake samples.

在这种情况下， G将捕获数据概率分布。 而D则无法区分真实样本或假样本。

GANs have been used in a lot of different applications in the past few years. Some of them include: generating synthetic data, Image in-paining, semi-supervised learning, super-resolution, and text to image generation.

在过去的几年中，GAN已用于许多不同的应用程序中。 其中包括：生成合成数据，图像痛苦，半监督学习，超分辨率和文本到图像生成。

However, much of the recent work on GANs is focused on developing techniques to stabilize training. Indeed, GANs are known to be unstable during training, and very sensitive to the choice of hyper-parameters.

但是，最近有关GAN的许多工作都集中在开发稳定培训的技术上。 确实，众所周知，GAN在训练过程中不稳定，并且对超参数的选择非常敏感。

In this context, this piece presents an overview of two relevant techniques for improving GANs. Specifically, we aim to describe recent methods for improving the quality of G samples. To do that, we address two techniques explored in the recent paper: Self-Attention Generative Adversarial Networks.

在这种情况下，本文概述了两种用于改善GAN的相关技术。 具体而言，我们旨在描述提高G样品质量的最新方法。 为此，我们解决了最近一篇论文中探讨的两种技术： 自我注意生成对抗网络 。

All the code developed with the Tensorflow Eager execution