该专栏为热销专栏榜 第29名
超级会员免费看
该博客详细介绍了生成对抗网络（GAN）的改进，包括网络结构的调整，如Deep Convolutional GAN (DCGAN)、ResNet、Self-Attention GAN (SAGAN)等，以及引入编码器的模型，如VAE-GAN、Bidirectional GAN (BiGAN)等。此外，还讨论了使用能量模型的方法，如Energy-based GAN (EBGAN)和Loss-Sensitive GAN (LSGAN)。最后，提到了优化过程的改进，如Two Time-Scale Update Rule (TTUR)和Wasserstein GANs。这些改进旨在提升GAN的生成质量和稳定性。
文章目录
- 【人工智能】生成对抗网络(Generative Adversarial Network)[2]
-
- （2）改进网络结构
-
- a. 调整神经网络
- ⚪ [Deep Convolutional GAN (DCGAN)](https://0809zheng.github.io/2022/02/05/dcgan.html)
- ⚪ ResNet
- ⚪ [Self-Attention GAN (SAGAN)](https://0809zheng.github.io/2022/05/25/sagan.html)
- ⚪ [BigGAN](https://0809zheng.github.io/2022/05/27/biggan.html)
- ⚪ [Self-Modulation](https://0809zheng.github.io/2022/05/28/selfmod.html)
- ⚪ [StyleGAN](https://0809zheng.github.io/2022/05/30/stylegan.html)
- ⚪ [StyleGAN2](https://0809zheng.github.io/2022/05/31/styleganv2.html)
- ⚪ [StyleGAN3](https://0809zheng.github.io/2022/06/15/stylegan3.html)
- ⚪ [Transformer GAN (TransGAN)](https://0809zheng.github.io/2021/03/02/transgan.html)
- b. 引入编码器
- ⚪ [VAE-GAN](https://0809zheng.github.io/2022/02/17/vaegan.html)
- ⚪ [Bidirectional GAN (BiGAN)](https://0809zheng.github.io/2022/02/18/bigan.html)
- ⚪ [VQGAN](https://0809zheng.github.io/2022/05/26/vqgan.html)
- c. 使用能量模型
- ⚪ [Energy-based GAN (EBGAN)](https://0809zheng.github.io/2022/02/16/ebgan.html)
- ⚪ [Loss-Sensitive GAN (LSGAN)](https://0809zheng.github.io/2022/02/25/lsgan.html)
- ⚪ [Boundary Equilibrium GAN (BEGAN)](https://0809zheng.github.io/2022/02/27/began.html)
- ⚪ [Margin Adaptation GAN (MAGAN)](https://0809zheng.github.io/2022/02/24/magan.html)
- ⚪ [Maximum Entropy Generator (MEG)](https://0809zheng.github.io/2022/02/23/meg.html)
- d. 由粗到细的生成 (Coarse-to-Fine)
- ⚪ [Laplacian Pyramid GAN (LAPGAN)](https://0809zheng.github.io/2022/03/22/lapgan.html)
- ⚪ [Stacked GAN (StackGAN)](https://0809zheng.github.io/2022/05/23/stackgan.html)
- ⚪ [Progressive Growing GAN (PGGAN)](https://0809zheng.github.io/2022/05/21/pggan.html)
- ⚪ [Single Natural Image GAN (SinGAN)](https://0809zheng.github.io/2022/05/22/singan.html)
- （3）改进优化过程
-
- ⚪ [Improved Techniques for Training GANs](https://0809zheng.github.io/2022/02/02/improve.html)
- ⚪ [Towards Principled Methods for Training Generative Adversarial Networks](https://0809zheng.github.io/2022/02/03/principle.html)
- ⚪ [Wasserstein GANs Work Because They Fail (to Approximate the Wasserstein Distance)](https://0809zheng.github.io/2022/02/13/fail.html)
- ⚪ [Two Time-Scale Update Rule (TTUR)](https://0809zheng.github.io/2022/03/24/ttur.html)
- ⚪ [Dirac GAN](https://0809zheng.github.io/2022/03/23/diracgan.html)
- ⚪ [Orthogonal GAN (O-GAN)](https://0809zheng.github.io/2022/06/16/ogan.html)
- ⚪ [Cascading Rejection](https://0809zheng.github.io/2022/04/25/cascading.html)
- ⚪ [Adaptive Discriminator Augmentation (ADA)](https://0809zheng.github.io/2022/06/11/ganaug.html)
- ⚪ [Hubness Prior](https://0809zheng.github.io/2022/03/31/hubness.html)
- （4）其他应用
-
- a. 条件生成 (Conditional Generation)
- ⚪ [Conditional GAN (CGAN)](https://0809zheng.github.io/2022/03/02/cgan.html)
- ⚪ [Information Maximizing GAN (InfoGAN)](https://0809zheng.github.io/2022/03/06/infogan.html)
- ⚪ [Auxiliary Classifier GAN (ACGAN)](https://0809zheng.github.io/2022/03/03/acgan.html)
- ⚪ [Projection Discriminator](https://0809zheng.github.io/2022/05/24/cgan.html)
- b. [图像到图像翻译 (Image-to-Image Translation)](https://0809zheng.github.io/2020/05/23/image_translation.html)
- c. 超分辨率 (Super Resolution)
- ⚪ [Super Resolution GAN (SRGAN)](https://0809zheng.github.io/2020/08/10/srresnet.html)
- ⚪ [Enhanced Super Resolution GAN (ESRGAN)](https://0809zheng.github.io/2020/08/12/esrgan.html)
- d. 图像修补 (Image Completion / Inpainting)
- ⚪ [Context Encoder](https://0809zheng.github.io/2022/05/19/context.html)
- ⚪ [Context-Conditional GAN (CCGAN)](https://0809zheng.github.io/2022/05/20/ccgan.html)
- ⚪ [Spatially-Adaptive Denormalization (SPADE)](https://0809zheng.github.io/2022/05/18/gaugan.html)
- e. 机器学习应用
- ⚪ [Semi-Supervised GAN](https://0809zheng.github.io/2022/06/14/ssgan.html)
- ⚪ [Anomaly Detection GAN (AnoGAN)](https://0809zheng.github.io/2020/10/23/anogan.html)
- ⚪ [Clustering GAN (ClusterGAN)](https://0809zheng.github.io/2022/06/13/clustergan.html)
- ⚪ 参考文献
【人工智能】生成对抗网络(Generative Adversarial Network)[2]
（2）改进网络结构
a. 调整神经网络
⚪ Deep Convolutional GAN (DCGAN)
DCGAN使用卷积神经网络构造生成对抗网络，为了稳定卷积网络的训练过程，作者提出了以下几点设计思路：
- 去掉网络中的pooling层，在判别器中使用Strided convolution (步幅卷积)进行下采样，在生成器中使用transposed convolution（转置卷积）进行上采样；
- 在判别器和生成器中使用batch norm；
- 移除网络中的所有全连接层；
- 生成器的输出层使用Tanh激活函数，其他
订阅专栏 解锁全文
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
