位置: IT常识 - 正文

Diffusion-GAN: Training GANs with Diffusion 解读

编辑:rootadmin
Diffusion-GAN: Training GANs with Diffusion 解读

推荐整理分享Diffusion-GAN: Training GANs with Diffusion 解读,希望有所帮助,仅作参考,欢迎阅读内容。

文章相关热门搜索词:,内容如对您有帮助,希望把文章链接给更多的朋友!

 Diffusion-GAN: 将GAN与diffusion一起训练 

paper:https://arxiv.org/abs/2206.02262

code:GitHub - Zhendong-Wang/Diffusion-GAN: Official PyTorch implementation for paper: Diffusion-GAN: Training GANs with Diffusion

  第一行从左向右看是diffusion forward的过程,不断由 real image 进行 diffusion,第三行从右向左看是由noise逐步恢复成fake image的过程,第二行是鉴别器D,D对每一个timestep都进行鉴别。 

 Figure 1: Flowchart for Diffusion-GAN. The top-row images represent the forward diffusion process of a real image, while the bottom-row images represent the forward diffusion process of a generated fake image. The discriminator learns to distinguish a diffused real image from a diffused fake image at all diffusion steps.

in Figure 1. In Diffusion-GAN, the input to the diffusion process is either a real or a generated image, and the diffusion process consists of a series of steps that gradually add noise to the  image. The number of diffusion steps is not fixed, but depends on the data and the generator. We also design the diffusion process to be differentiable, which means that we can compute the derivative of the output with respect to the input. This allows us to propagate the gradient from the discriminator to the generator through the diffusion process, and update the generator accordingly. Unlike vanilla GANs, which compare the real and generated images directly, Diffusion-GAN compares the noisy versions of them, which are obtained by sampling from the Gaussian mixture distribution over the diffusion steps, with the help of our timestep-dependent discriminator. This distribution has the property that its components have different noise-to-data ratios, which means that some components add more noise than others. By sampling from this distribution, we can achieve two benefits: first, we can stabilize the training by easing the problem of vanishing gradient, which occurs when the data and generator distributions are too different; second, we can augment the data by creating different noisy versions of the same image, which can improve the data efficiency and the diversity of the generator. We provide a theoretical analysis to support our method, and show that the min-max objective function of Diffusion-GAN, which measures the difference between the data and generator distributions, is continuous and differentiable everywhere. This means that the generator in theory can always receive a useful gradient from the discriminator, and improve its performance.【G可以从D收到有用的梯度,从而提升G的性能】

主要贡献:

1) We show both theoretically and empirically how the diffusion process can be utilized to provide a model- and domain-agnostic differentiable augmentation, enabling data-efficient and leaking-free stable GAN training.【稳定了GAN的训练】 2) Extensive experiments show that Diffusion-GAN boosts the stability and generation performance of strong baselines, including StyleGAN2 , Projected GAN , and InsGen , achieving state-of-the-art results in synthesizing photo-realistic images, as measured by both the Fréchet Inception Distance (FID)  and Recall score.【diffusion提升了原始只有GAN组成的框架的性能,例如styleGAN2,Projected GAN】

Diffusion-GAN: Training GANs with Diffusion 解读

Figure 2: The toy example inherited from Arjovsky et al. [2017]. The first row plots the distributions of data with diffusion noise injected for t. The second row shows the JS divergence and the optimal discriminator value with and without our noise injection. 

Figure 4: Plot of adaptively adjusted maximum diffusion steps T and discriminator outputs of Diffusion-GANs. 

To investigate how the adaptive diffusion process works during training, we illustrate in Figure 4 the convergence of the maximum timestep T in our adaptive diffusion and discriminator outputs. We see that T is adaptively adjusted: The T for Diffusion StyleGAN2 increases as the training goes while the T for Diffusion ProjectedGAN first goes up and then goes down. Note that the T is adjusted according to the overfitting status of the discriminator. The second panel shows that trained with the diffusion-based mixture distribution, the discriminator is always well-behaved and provides useful learning signals for the generator, which validates our analysis in Section 3.4 and Theorem 1.

如图4左所示,随着训练过程的变化,扩散的timestep T也会自适应的改变(T通过鉴别器D过拟合的状态而改变); 如图4右所示,用基于扩散的混合分布训练的鉴别器总是表现良好,并为生成器G提供有用的学习信号。

Effectiveness of Diffusion-GAN for domain-agnostic augmentation(未知域增强的有效性)

25-Gaussians Example.

We conduct experiments on the popular 25-Gaussians generation task. The 25-Gaussians dataset is a 2-D toy data, generated by a mixture of 25 two-dimensional Gaussian distributions. Each data point is a 2-dimensional feature vector. We train a small GAN model, whose generator and discriminator are both parameterized by multilayer perceptrons (MLPs), with two 128-unit hidden layers and LeakyReLu nonlinearities.

Figure 5: The 25-Gaussians example. We show the true data samples, the generated samples from vanilla GANs, the discriminator outputs of the vanilla GANs, the generated samples from our Diffusion-GAN, and the discriminator outputs of Diffusion-GAN. 

(1)groundtruth数据集的数据分布,在25个Gaussians example均匀分布; (2)vanilla GANs的输出结果产生了mode collapsing,只在几个model上生成数据; (3)vanilla GANs鉴别器输出很快就会彼此分离。这意味着发生了鉴别器的强烈过拟合,使得鉴别器停止为发生器提供有用的学习信号。 (4)Diffusion-GAN在25个example上均匀分布,意味着它在所有的model上学到了采样分布; (5)Diffusion-GAN的鉴别器输出,D在持续的为G提供有用的学习信号

我们从两个角度来解释这种改进: 首先,non-leaking augmentation(无泄漏增强)有助于提供关于数据空间的更多信息;第二,自适应调整的基于扩散的噪声注入,鉴别器表现良好。

关于 Difffferentiable augmentation. (可微分增强)

As Diffusion-GAN transforms both the data and generated samples before sending them to the discriminator, we can also relate it to differentiable augmentation proposed for data-efficient GAN training. Karras et al introduce a stochastic augmentation pipeline with 18 transformationsand develop an adaptive mechanism for controlling the augmentation probability. Zhao et al. [2020] propose to use Color + Translation + Cutout as differentiable augmentations for both generated and real images.

While providing good empirical results on some datasets, these augmentation methods are developed with domain-specific knowledge and have the risk of leaking augmentation  into generation [Karras et al., 2020a]. As observed in our experiments, they sometime worsen the results when applied to a new dataset, likely because the risk of augmentation leakage overpowers the benefits of enlarging the training set, which could happen especially if the training set size is already sufficiently large.(在数据量足够大的情况下,数据增强带来的负面效果可能大于正面效果)

By contrast, Diffusion-GAN uses a differentiable forward diffusion process to stochastically transform the data and can be considered as both a domain-agnostic and a model-agnostic augmentation method. In other words, Diffusion-GAN can be applied to non-image data or even latent features, for which appropriate data augmentation is difficult to be defined, and easily plugged into an existing GAN to improve its generation performance. Moreover, we prove in theory and show in experiments that augmentation leakage is not a concern for Diffusion-GAN. Tran et al. [2021] provide a theoretical analysis for deterministic non-leaking transformation with differentiable and invertible mapping functions. Bora et al. [2018] show similar theorems to us for specific stochastic transformations, such as Gaussian Projection, Convolve+Noise, and stochastic Block-Pixels, while our Theorem 2 includes more satisfying possibilities as discussed in Appendix B.

本文链接地址:https://www.jiuchutong.com/zhishi/294494.html 转载请保留说明!

上一篇:Vue|非单文件组件(vuecli非根目录打包)

下一篇:【HTML】原生js实现的图书馆管理系统(javascript原生)

  • 提足折旧是指
  • 个税汇算清缴可以修改收入吗
  • 增值税视同销售账务处理怎么做?
  • 公司网站建设费做账是流量应计入
  • 融资租赁和经营租赁的特点
  • 测绘费发票需要备注么
  • 房地产契税新政策2021退税
  • 小规模餐饮业会计核算
  • 小规模开票数量怎么算
  • 材料盘亏可收回成本吗
  • 收到样品费计入什么科目
  • 从支付宝里可以查出结婚个人信息吗
  • 增值税怎么查看
  • 机打发票丢失如何处理报销
  • 过了认证期的发票怎么办
  • 修理费要开税控清单吗
  • 2017年个税税率表及个税计算公式
  • 个人买卖二手房增值税
  • 公司之间借款收据要领导签字吗
  • 个体工商户起征点10万执行时间
  • 社保退回的钱怎么做会计分录
  • 其他应收款计提坏账比例
  • 如何防止程序被关闭
  • 学php的书
  • shell检查变量是否为空
  • 以合同条款无法达成一致要求返还定金
  • 银行进账单和现金缴款单的区别
  • react-use
  • 回迁安置房是否有产权
  • php7.0新特性
  • 利润表利息费用怎么填
  • 文化事业建设费减免政策
  • 主营业务收入借贷方向
  • php多进程处理大数据
  • 远程调试时,gdbserver运行在调试机
  • ps制作折扇效果图
  • layui树形下拉框
  • python中的生成器
  • 应收账款计提坏账准备是什么意思
  • mongodb subtract
  • 已经抵扣的进项税额转出会计分录
  • 开始送加盟费
  • 公司注销其他应付款有余额有影响吗
  • 工资计提啥意思
  • 常用sql脚本
  • 扇贝的储存方式
  • 定额备用金怎么做分录
  • 业务招待费如何扣除
  • 留抵进项税太多怎么办
  • 增量成本属于什么成本
  • 高温补贴发放管理制度
  • 疫情期间水电费补贴收入申报企业所得税吗
  • 客户退货不退赠品怎么办
  • 价税合计怎么求税额
  • 政府会计制度累计盈余为负数
  • 车辆使用费报销制度
  • sql server存储过程教程
  • mysql 数据修改
  • 网络硬盘在哪
  • windows查找命令
  • ubuntu 12.04 intel集成显卡设置分辨率的步骤分享
  • sxs.exe病毒
  • ubuntu20录屏
  • win7总是弹出广告怎么办
  • win7使用技巧
  • win7更新8007000e
  • win 7怎么办
  • win8桌面不见了
  • unity 移动应用开发
  • unity-3d
  • js中的filter方法
  • jquery validator
  • python 基础入门
  • jquery简写符号
  • JavaScript中的this指向
  • jquery实现ajax提交表单信息的简单方法(推荐)
  • python安装后怎么用
  • 稳岗补贴是否需要发放
  • 留抵税额退税政策2022
  • 税务局的纳税服务中心是干什么的
  • 免责声明:网站部分图片文字素材来源于网络,如有侵权,请及时告知,我们会第一时间删除,谢谢! 邮箱:opceo@qq.com

    鄂ICP备2023003026号

    网站地图: 企业信息 工商信息 财税知识 网络常识 编程技术

    友情链接: 武汉网站建设