summary4

发表于 更新于 分类于
本文字数： 1.2k 阅读时长 ≈ 1 分钟

Learning summary 4

公开课

Machine Learning 2021 Spring：

  • Introduction of Generative Model: NetWork that can output distribution is called Generater

  • Generative Adversarial Network

    • Generator
    • Discriminator : CNN, Transformer
    • Algorithm
    • Unconditional Generation
      • Anime Face Generation: Style-GAN
        • the complex distribution generated by Generator and the data distribution are as close as possible to each other.
        • Divergence： KL Divergence, JS Divergence
        • \(\underset{D}{max}\space V(G,D)\) is related to JS divergence, replace \(Div(P_G, P_{data})\)： \(G^* =\text{arg } \underset{G}{\text{min }} \underset{D}{\text{max }} V(G,D)\)
        • Wasserstein distance （replace JS Div）-- WGAN
      • Sequence Generation
    • Conditional Generation
      • Text-to-Image
      • Image Translation （Pix2pix）
      • Sound to Image
    • Evaluation of Generator
      • Inception Score
      • Question: Mode Collapse and Mode Dropping
      • Fréchet Inception Distance (FID)

  • GAN in Unsupervised Learning: turn data of one domain into data of another domain

    • Cycle GAN、Disco GAN、Dual GAN、StarGAN
    • Bi-directional Cycle GAN
    • Application：风格迁移（图像，文本）、无监督翻译

PyTorch

《Deep Learing with PyTorch》p1-p65

  • The deep learning revolution

  • recognizes the subject of an image

    • Dataset: ImageNet
    • ResNet

  • 风格迁移：turns horses into zebras

    • Cycle GAN
    • Two Generators：9 ResNetBlock( in_channels=256, output_channals=256 ) + ResNetGenerator
    • Discriminator: Classifier

    9

  • imagecaption: NeuralTalk2