DCGAN

发表于 |

UNSUPERVISED REPRESENTATION LEARNING WITH DEEP CONVOLUTIONAL GENERATIVE ADVERSARIAL NETWORKS

  1. 动机

    • unsupervised learning
    • learns a hierarchy of representations from object parts to scenes
    • used for novel tasks

  2. 论点

    • GAN
      • Learning reusable feature representations from large unlabeled datasets
      • generator and discriminator networks can be later used as feature extractors for supervised tasks
      • unstable to train
    • we
      • propose a set of constraints on the architectural topology making it stable to train
      • use the trained discriminators for image classification tasks
      • visualize the filters
      • show that the generators have interesting vector arithmetic properties
    • unsupervised representation learning
      • clustering, hierarchical clustering
      • auto-encoders learn good feature representations
    • generative image models
      • samples often suffer from being blurry, being noisy and incomprehensible
      • further use for supervised tasks

  3. 方法

    • architecture

      • all convolutional net：没有池化，用stride conv
      • eliminating fully connected layers：
        • generator：输入是一个向量，reshape以后接的全是卷积层
        • discriminator：最后一层卷积出来直接flatten
      • Batch Normalization
        • generator输出层 & discriminator输入层不加
        • resulted in sample oscillation and model instability

      • ReLU

        • generator输出层用Tanh
        • discriminator用leakyReLU

    • train

      • image preprocess：rescale to [-1,1]
      • LeakyReLU(0.2)
      • lr：2e-4
      • momentum term $\beta 1$：0.5, default 0.9

  4. 实验

    • evaluate
      • apply them as a feature extractor on supervised datasets
      • evaluate the performance of linear models on top of these features
    • model
      • use the discriminator’s convolutional features from all layers
      • maxpooling to 4x4 grids
      • flattened and concatenated to form a 28672 dimensional vector
      • regularized linear L2-SVM
      • 相比之下：the discriminator has many less feature maps, but larger total feature vector size
    • visualizing
      • walking in the latent space
        • 在vector Z上差值，生成图像可以观察到smooth transitions
      • visualize the discriminator feature
        • 特征图可视化，能观察到床结构
      • manipulate the generator representation
        • generator learns specific object representations for major scene components
        • use logistic regression to find feature maps related with window, drop the spatial locations on feature-maps
        • most result forgets to draw windows in the bedrooms, replacing them with other objects
    • vector arithmetic
      • averaging the Z vector for three examplars
      • semantically obeyed the arithmetic