目錄

一、Generative Adversarial Nets

1. GAN簡介

對抗生成網路（GAN）同時訓練兩個模型：能夠得到資料分佈的生成模型（generative model G）和能判夠區別資料是生成的還是真實的判別模型 （discriminative model D）。訓練過程使得G生成的資料儘可能真實，同時又使得D儘可能能夠區分生成的資料和真實的資料，最終G生成資料足以以假亂真，而D輸出資料的概率均為0.5 。 參考論文：Bengio大神的 Generative Adversarial Networks

2. GAN訓練過程

注意：這裡的loss在更新D梯度是上升方向，在caffe具體實現時，為了使得D模型梯度更新為梯度的下降方向，loss等價改為: l

二、Deep Convolutional GANs (DCGAN)

1. DCGAN 網路結構

• DCGAN網路結構圖
  
• DCGAN結構內容
  

2. DCGAN caffe prototxt

1.train.prototxt

#  Create on: 2016/10/22 ShanghaiTech
#  Author:    Yingying Zhang
name: "gan_newface"
layer {
  name: "images"
  type: "ImageData"
  top: "face_images"
 

  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    mirror: true
    mean_value: [127.5]
    scale: 0.00784314
  }
  image_data_param {
    source: "data/face_data.txt"
    root_folder: "data/"
    batch_size: 128
    new_height: 64
    new_width: 64
    is_color: true
    shuffle: true
 

  }
}
layer {
  name: "rand_vec"
  type: "RandVec"
  top: "rand_vec"
  rand_vec_param {
    batch_size: 128
    dim: 100
    lower: -1.0
    upper: 1.0
    repeat: 3
  }
}
layer {
  name: "silence"
  type: "Silence"
  bottom: "label"
}
#----------- generative -----------
layer {
  name: "ip1"
  type: "InnerProduct"
  bottom: "rand_vec"
  top: "ip1"
  param {
    name: "ip_w_g"
    lr_mult: 1
  }
  param {
    name: "ip_b_g"
    lr_mult: 2
  }
  inner_product_param{
    num_output: 16384
    gen_mode: true
    weight_filler {
      type: "gaussian"
      std: 0.02
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "ip1_reshape"
  type: "Reshape"
  bottom: "ip1"
  top: "ip1_reshape"
  reshape_param {
    shape {
      dim: 0
      dim: 1024
      dim: 4
      dim: 4
    }
  }
}
layer {
  name: "batch_norm_g1"
  type: "BatchNorm"
  bottom: "ip1_reshape"
  top: "ip1_reshape"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "scale_batch_g1"
  type: "Scale"
  bottom: "ip1_reshape"
  top: "ip1_reshape"
  param {
    name: "gen_s1"
    lr_mult: 1
    decay_mult: 1
  }
  param {
    name: "gen_b1"
    lr_mult: 1
    decay_mult: 1
  }
  scale_param {
    bias_term: true
    gen_mode: true
  }
}
layer {
  name: "relu_g1"
  type: "ReLU"
  bottom: "ip1_reshape"
  top: "ip1_reshape"
}
layer {
  name: "gen_conv2"
  type: "Deconvolution"
  bottom: "ip1_reshape"
  top: "gen_conv2"
  param {
    name: "gen_w_2"
    lr_mult: 1
  }
  param {
    name: "gen_b_2"
    lr_mult: 2
  }
  convolution_param {
    num_output: 512
    pad: 2
    kernel_size: 5
    stride: 2
    gen_mode: true
    shape_offset: [1, 1]
    weight_filler {
      type: "gaussian"
      std: 0.02
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "batch_norm_g2"
  type: "BatchNorm"
  bottom: "gen_conv2"
  top: "gen_conv2"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "scale_batch_g2"
  type: "Scale"
  bottom: "gen_conv2"
  top: "gen_conv2"
  param {
    name: "gen_s2"
    lr_mult: 1
    decay_mult: 1
  }
  param {
    name: "gen_b2"
    lr_mult: 1
    decay_mult: 1
  }
  scale_param {
    gen_mode: true
    bias_term: true
  }
}
layer {
  name: "relu_g2"
  type: "ReLU"
  bottom: "gen_conv2"
  top: "gen_conv2"
}
layer {
  name: "gen_conv3"
  type: "Deconvolution"
  bottom: "gen_conv2"
  top: "gen_conv3"
  param {
    name: "gen_w_3"
    lr_mult: 1
  }
  param {
    name: "gen_b_3"
    lr_mult: 2
  }
  convolution_param {
    num_output: 256
    pad: 2
    kernel_size: 5
    stride: 2
    gen_mode: true
    shape_offset: [1, 1]
    weight_filler {
      type: "gaussian"
      std: 0.02
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "batch_norm_g3"
  type: "BatchNorm"
  bottom: "gen_conv3"
  top: "gen_conv3"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "scale_batch_g3"
  type: "Scale"
  bottom: "gen_conv3"
  top: "gen_conv3"
  param {
    name: "gen_s3"
    lr_mult: 1
    decay_mult: 1
  }
  param {
    name: "gen_b3"
    lr_mult: 1
    decay_mult: 1
  }
  scale_param {
    gen_mode: true
    bias_term: true
  }
}
layer {
  name: "relu_g3"
  type: "ReLU"
  bottom: "gen_conv3"
  top: "gen_conv3"
}
layer {
  name: "gen_conv4"
  type: "Deconvolution"
  bottom: "gen_conv3"
  top: "gen_conv4"
  param {
    name: "gen_w_4"
    lr_mult: 1
  }
  param {
    name: "gen_b_4"
    lr_mult: 2
  }
  convolution_param {
    num_output: 128
    pad: 2
    kernel_size: 5
    stride: 2
    gen_mode: true
    shape_offset: [1, 1]
    weight_filler {
      type: "gaussian"
      std: 0.02
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "batch_norm_g4"
  type: "BatchNorm"
  bottom: "gen_conv4"
  top: "gen_conv4"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "scale_batch_g4"
  type: "Scale"
  bottom: "gen_conv4"
  top: "gen_conv4"
  param {
    name: "gen_s4"
    lr_mult: 1
    decay_mult: 1
  }
  param {
    name: "gen_b4"
    lr_mult: 1
    decay_mult: 1
  }
  scale_param {
    gen_mode: true
    bias_term: true
  }
}
layer {
  name: "relu_g4"
  type: "ReLU"
  bottom: "gen_conv4"
  top: "gen_conv4"
}
layer {
  name: "gen_conv5"
  type: "Deconvolution"
  bottom: "gen_conv4"
  top: "gen_conv5"
  param {
    name: "gen_w_5"
    lr_mult: 1
  }
  param {
    name: "gen_b_5"
    lr_mult: 2
  }
  convolution_param {
    num_output: 3
    pad: 2
    kernel_size: 5
    stride: 2
    gen_mode: true
    shape_offset: [1, 1]
    weight_filler {
      type: "gaussian"
      std: 0.02
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "tanh_g5"
  type: "TanH"
  bottom: "gen_conv5"
  top: "gen_conv5"
}
#-----------   gan gate  ------------
layer {
  name: "gan_gate"
  type: "GANGate"
  bottom: "face_images"
  bottom: "gen_conv5"
  top: "dis_input"
}
#----------- discrimitive -----------
layer {
  name: "dis_conv_d1"
  type: "Convolution"
  bottom: "dis_input"
  top: "dis_conv_d1"
  param {
    name: "dis_w_1"
    lr_mult: 1
  }
  param {
    name: "dis_b_1"
    lr_mult: 2
  }
  convolution_param {
    num_output: 128
    pad: 2
    kernel_size: 5
    stride: 2
    dis_mode: true
    weight_filler {
      type: "gaussian"
      std: 0.02
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "batch_norm_d1"
  type: "BatchNorm"
  bottom: "dis_conv_d1"
  top: "dis_conv_d1"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "scale_batch_d1"
  type: "Scale"
  bottom: "dis_conv_d1"
  top: "dis_conv_d1"
  param {
    name: "dis_s1"
    lr_mult: 1
    decay_mult: 1
  }
  param {
    name: "dis_b1"
    lr_mult: 1
    decay_mult: 1
  }
  scale_param {
    dis_mode: true
    bias_term: true
  }
}
layer {
  name: "relu_d1"
  type: "ReLU"
  bottom: "dis_conv_d1"
  top: "dis_conv_d1"
  relu_param {
    negative_slope: 0.2
  }
}
layer {
  name: "dis_conv_d2"
  type: "Convolution"
  bottom: "dis_conv_d1"
  top: "dis_conv_d2"
  param {
    name: "dis_w_2"
    lr_mult: 1
  }
  param {
    name: "dis_b_2"
    lr_mult: 2
  }
  convolution_param {
    num_output: 256
    pad: 2
    kernel_size: 5
    stride: 2
    dis_mode: true
    weight_filler {
      type: "gaussian"
      std: 0.02
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "batch_norm_d2"
  type: "BatchNorm"
  bottom: "dis_conv_d2"
  top: "dis_conv_d2"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "scale_batch_d2"
  type: "Scale"
  bottom: "dis_conv_d2"
  top: "dis_conv_d2"
  param {
    name: "dis_s2"
    lr_mult: 1
    decay_mult: 1
  }
  param {
    name: "dis_b2"
    lr_mult: 1
    decay_mult: 1
  }
  scale_param {
    dis_mode: true
    bias_term: true
  }
}
layer {
  name: "relu_d2"
  type: "ReLU"
  bottom: "dis_conv_d2"
  top: "dis_conv_d2"
  relu_param {
    negative_slope: 0.2
  }
}
layer {
  name: "dis_conv_d3"
  type: "Convolution"
  bottom: "dis_conv_d2"
  top: "dis_conv_d3"
  param {
    name: "dis_w_3"
    lr_mult: 1
  }
  param {
    name: "dis_b_3"
    lr_mult: 2
  }
  convolution_param {
    num_output: 512
    pad: 2
    kernel_size: 5
    stride: 2
    dis_mode: true
    weight_filler {
      type: "gaussian"
      std: 0.02
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "batch_norm_d3"
  type: "BatchNorm"
  bottom: "dis_conv_d3"
  top: "dis_conv_d3"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "scale_batch_d3"
  type: "Scale"
  bottom: "dis_conv_d3"
  top: "dis_conv_d3"
  param {
    name: "dis_s3"
    lr_mult: 1
    decay_mult: 1
  }
  param {
    name: "dis_b3"
    lr_mult: 1
    decay_mult: 1
  }
  scale_param {
    dis_mode: true
    bias_term: true
  }
}
layer {
  name: "relu_d3"
  type: "ReLU"
  bottom: "dis_conv_d3"
  top: "dis_conv_d3"
  relu_param {
    negative_slope: 0.2
  }
}
layer {
  name: "dis_conv_d4"
  type: "Convolution"
  bottom: "dis_conv_d3"
  top: "dis_conv_d4"
  param {
    name: "dis_w_4"
    lr_mult: 1
  }
  param {
    name: "dis_b_4"
    lr_mult: 2
  }
  convolution_param {
    num_output: 1024
    pad: 2
    kernel_size: 5
    stride: 2
    dis_mode: true
    weight_filler {
      type: "gaussian"
      std: 0.02
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "batch_norm_d4"
  type: "BatchNorm"
  bottom: "dis_conv_d4"
  top: "dis_conv_d4"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "scale_batch_d4"
  type: "Scale"
  bottom: "dis_conv_d4"
  top: "dis_conv_d4"
  param {
    name: "dis_s4"
    lr_mult: 1
    decay_mult: 1
  }
  param {
    name: "dis_b4"
    lr_mult: 1
    decay_mult: 1
  }
  scale_param {
    dis_mode: true
    bias_term: true
  }
}
layer {
  name: "relu_d4"
  type: "ReLU"
  bottom: "dis_conv_d4"
  top: "dis_conv_d4"
  relu_param {
    negative_slope: 0.2
  }
}
layer {
  name: "score"
  type: "InnerProduct"
  bottom: "dis_conv_d4"
  top: "score"
  param {
    name: "score_w"
    lr_mult: 1
  }
  param {
    name: "score_b"
    lr_mult: 2
  }
  inner_product_param{
    num_output: 1
    dis_mode: true
    weight_filler {
      type: "gaussian"
      std: 0.0002
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "sigmoid"
  type: "Sigmoid"
  bottom: "score"
  top: "score"
}
layer {
  name: "gan_loss"
  type: "GANDGLoss"
  bottom: "score"
  top: "gan_loss"
  gan_loss_param {
    dis_iter: 1
    gen_iter: 2
  }
}

2.solver.prototxt

#  Create on: 2016/10/22 ShanghaiTech
#  Author:    Yingying Zhang

net: "gan_configs/train.prototxt"
debug_info: false
display: 10
solver_type: ADAM
average_loss: 100
base_lr: 2e-4
lr_policy: "fixed"
max_iter: 15000
momentum: 0.5
snapshot: 100
gan_solver: true
snapshot_prefix: "models/gan_"

3.生成結果