GAN、DCGANs、WGANS

GAN

GAN模型没有损失函数，优化过程是一个“二元极小极大博弈（minimax two-player game）”问题:

这是关于判别网络D和生成网络G的价值函数（Value Function），训练网络D使得最大概率地分对训练样本的标签（最大化log D(x)），训练网络G最小化log(1 – D(G(z)))，即最大化D的损失。训练过程中固定一方，更新另一个网络的参数，交替迭代，使得对方的错误最大化，最终，G 能估测出样本数据的分布。生成模型G隐式地定义了一个概率分布Pg，我们希望Pg 收敛到数据真实分布Pdata。论文证明了这个极小化极大博弈当且仅当Pg = Pdata时存在最优解，即达到纳什均衡，此时生成模型G恢复了训练数据的分布，判别模型D的准确率等于50%。

DCGANs

改进后相比原始GAN的算法实现流程却只改了四点：
1.将所有的池化层变为带步长的卷积层(分类器)，带分数步长的卷积层(生成器)
2.在分类器和生成器中都使用batchnorm
3.去除全连接层
4.在生成器中除了输出层使用tanh以外，其他激活层均使用ReLU
5.在分类器中激活层均使用LeakyRelu

Generator生成器:

相应的tensorflow代码：

    def generator(self, z, y=None):
      with tf.variable_scope("generator") as scope:
      if not self.y_dim:
        s_h, s_w = self.output_height, self.output_width
        s_h2, s_w2 = conv_out_size_same(s_h, 2), conv_out_size_same(s_w, 2)
        s_h4, s_w4 = conv_out_size_same(s_h2, 2), conv_out_size_same(s_w2, 2)
        s_h8, s_w8 = conv_out_size_same(s_h4, 2), conv_out_size_same(s_w4, 2)
        s_h16, s_w16 = conv_out_size_same(s_h8, 2), conv_out_size_same(s_w8, 2)
        
        # project `z` and reshape
        self.z_, self.h0_w, self.h0_b = linear(
            z, self.gf_dim*8*s_h16*s_w16, 'g_h0_lin', with_w=True)

        self.h0 = tf.reshape(
            self.z_, [-1, s_h16, s_w16, self.gf_dim * 8])
        h0 = tf.nn.relu(self.g_bn0(self.h0))

        self.h1, self.h1_w, self.h1_b = deconv2d(
            h0, [self.batch_size, s_h8, s_w8, self.gf_dim*4], name='g_h1', with_w=True)
        h1 = tf.nn.relu(self.g_bn1(self.h1))

        h2, self.h2_w, self.h2_b = deconv2d(
            h1, [self.batch_size, s_h4, s_w4, self.gf_dim*2], name='g_h2', with_w=True)
        h2 = tf.nn.relu(self.g_bn2(h2))

        h3, self.h3_w, self.h3_b = deconv2d(
            h2, [self.batch_size, s_h2, s_w2, self.gf_dim*1], name='g_h3', with_w=True)
        h3 = tf.nn.relu(self.g_bn3(h3))

        h4, self.h4_w, self.h4_b = deconv2d(
            h3, [self.batch_size, s_h, s_w, self.c_dim], name='g_h4', with_w=True)

        return tf.nn.tanh(h4)
      else:
        s_h, s_w = self.output_height, self.output_width
        s_h2, s_h4 = int(s_h/2), int(s_h/4)
        s_w2, s_w4 = int(s_w/2), int(s_w/4)

        # yb = tf.expand_dims(tf.expand_dims(y, 1),2)
        yb = tf.reshape(y, [self.batch_size, 1, 1, self.y_dim])
        z = concat([z, y], 1)

        h0 = tf.nn.relu(
            self.g_bn0(linear(z, self.gfc_dim, 'g_h0_lin')))
        h0 = concat([h0, y], 1)

        h1 = tf.nn.relu(self.g_bn1(
            linear(h0, self.gf_dim*2*s_h4*s_w4, 'g_h1_lin')))
        h1 = tf.reshape(h1, [self.batch_size, s_h4, s_w4, self.gf_dim * 2])

        h1 = conv_cond_concat(h1, yb)

        h2 = tf.nn.relu(self.g_bn2(deconv2d(h1,
            [self.batch_size, s_h2, s_w2, self.gf_dim * 2], name='g_h2')))
        h2 = conv_cond_concat(h2, yb)

        return tf.nn.sigmoid(
            deconv2d(h2, [self.batch_size, s_h, s_w, self.c_dim], name='g_h3'))

Discriminator

相应的tensorflow代码

  def discriminator(self, image, y=None, reuse=False):
    with tf.variable_scope("discriminator") as scope:
      if reuse:
        scope.reuse_variables()

      if not self.y_dim:
        h0 = lrelu(conv2d(image, self.df_dim, name='d_h0_conv'))
        h1 = lrelu(self.d_bn1(conv2d(h0, self.df_dim*2, name='d_h1_conv')))
        h2 = lrelu(self.d_bn2(conv2d(h1, self.df_dim*4, name='d_h2_conv')))
        h3 = lrelu(self.d_bn3(conv2d(h2, self.df_dim*8, name='d_h3_conv')))
        h4 = linear(tf.reshape(h3, [self.batch_size, -1]), 1, 'd_h4_lin')

        return tf.nn.sigmoid(h4), h4
      else:
        yb = tf.reshape(y, [self.batch_size, 1, 1, self.y_dim])
        x = conv_cond_concat(image, yb)

        h0 = lrelu(conv2d(x, self.c_dim + self.y_dim, name='d_h0_conv'))
        h0 = conv_cond_concat(h0, yb)

        h1 = lrelu(self.d_bn1(conv2d(h0, self.df_dim + self.y_dim, name='d_h1_conv')))
        h1 = tf.reshape(h1, [self.batch_size, -1])      
        h1 = concat([h1, y], 1)
        
        h2 = lrelu(self.d_bn2(linear(h1, self.dfc_dim, 'd_h2_lin')))
        h2 = concat([h2, y], 1)

        h3 = linear(h2, 1, 'd_h3_lin')
        
        return tf.nn.sigmoid(h3), h3

WGANs

Wasserstein GAN（下面简称WGAN）成功地做到了以下爆炸性的几点：

1.彻底解决GAN训练不稳定的问题，不再需要小心平衡生成器和判别器的训练程度
2.基本解决了collapse mode的问题，确保了生成样本的多样性
3.训练过程中终于有一个像交叉熵、准确率这样的数值来指示训练的进程，这个数值越小代表GAN训练得越好，代表生成器产生的图像质量越高（如题图所示）
4.以上一切好处不需要精心设计的网络架构，最简单的多层全连接网络就可以做到
改进后相比原始GAN的算法实现流程却只改了四点：

1.判别器最后一层去掉sigmoid
2.生成器和判别器的loss不取log
3.每次更新判别器的参数之后把它们的绝对值截断到不超过一个固定常数c
4.不要用基于动量的优化算法（包括momentum和Adam），推荐RMSProp，SGD也行

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import torch
import torch.nn as nn

class MLP_G(nn.Module):
    def __init__(self, isize, nz, nc, ngf, ngpu):
        super(MLP_G, self).__init__()
        self.ngpu = ngpu

        main = nn.Sequential(
            # Z goes into a linear of size: ngf
            nn.Linear(nz, ngf),
            nn.ReLU(True),
            nn.Linear(ngf, ngf),
            nn.ReLU(True),
            nn.Linear(ngf, ngf),
            nn.ReLU(True),
            nn.Linear(ngf, nc * isize * isize),
        )
        self.main = main
        self.nc = nc
        self.isize = isize
        self.nz = nz

    def forward(self, input):
        input = input.view(input.size(0), input.size(1))
        if isinstance(input.data, torch.cuda.FloatTensor) and self.ngpu > 1:
            output = nn.parallel.data_parallel(self.main, input, range(self.ngpu))
        else:
            output = self.main(input)
        return output.view(output.size(0), self.nc, self.isize, self.isize)


class MLP_D(nn.Module):
    def __init__(self, isize, nz, nc, ndf, ngpu):
        super(MLP_D, self).__init__()
        self.ngpu = ngpu

        main = nn.Sequential(
            # Z goes into a linear of size: ndf
            nn.Linear(nc * isize * isize, ndf),
            nn.ReLU(True),
            nn.Linear(ndf, ndf),
            nn.ReLU(True),
            nn.Linear(ndf, ndf),
            nn.ReLU(True),
            nn.Linear(ndf, 1),
        )
        self.main = main
        self.nc = nc
        self.isize = isize
        self.nz = nz

    def forward(self, input):
        input = input.view(input.size(0),
                           input.size(1) * input.size(2) * input.size(3))
        if isinstance(input.data, torch.cuda.FloatTensor) and self.ngpu > 1:
            output = nn.parallel.data_parallel(self.main, input, range(self.ngpu))
        else:
            output = self.main(input)
        output = output.mean(0)
        return output.view(1)

相关函数说明

tensorflow

pytorch

torch.nn.Linear

// An highlighted block
import torch
x = torch.randn(128,20)    # x [128,20]
m = torch.nn.Linear(20,30) # m.weight [30,20]  m.bias[30]
output = m(x)              # output [128,30]
ans = torch.mm(x,m.weight.t()) + m.bias

y = x A^T + b

torch.nn.ConvTranspose2d

class torch.nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride=1, padding=0, output_padding=0, groups=1, bias=True, dilation=1)

参数的含义如下：

对于每一条边输入输出的尺寸的公式如下：
output =(input-1) X stride + outputpadding - 2 X padding + kernelsize

Reference

【1】Goodfellow Ian, Pouget-Abadie J, Mirza M, et al. Generative adversarial nets[C]//Advances in Neural Information Processing Systems. 2014: 2672-2680.