Pytorch学习记录（八）：生成对抗网络GAN

1. GAN 原理

2. 纳什均衡

3. JS散度的缺陷

• 当两个分布完全不重叠的时候，对于任意的x输入，其JS散度值等于log2；
• 在固定D从而训练G时（此时网络损失和JS散度呈线性关系），会导致JS散度值为一个常量值，从而网络的损失也固定为一个常量值，从而网络参数梯度为0，参数得不到更新，训练不稳定；
  

4. EM距离

• 能够解决由于训练数据和虚假数据分布不重叠而导致的网络训练不稳定的问题，使用Wssserstein距离替代原来JS散度的损失；
  
  
  

5. GAN 实现

import torch
from torch import nn, optim, autograd
import numpy as np
import visdom
import random

h_dim = 400
batchsz = 512
viz = visdom.Visdom()

class Generator(nn.Module):
    def __init__(self, h_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(2, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, 2)
        )
    def forward(self, x):
        return self.net(x)

class Discriminator(nn.Module):
    def __init__(self, h_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(2, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, 1),  #输出分类概率
            nn.Sigmoid()
        )

    def forward(self, x):
        return self.net(x).view(-1)

def data_gernerator():
    scale = 2
    centers = [(1,0),(-1,0),(0,1),(0,-1),
               (1. / np.sqrt(2), 1. / np.sqrt(2)),
               (1. / np.sqrt(2), -1. / np.sqrt(2)),
               (-1. / np.sqrt(2), 1. / np.sqrt(2)),
               (-1. / np.sqrt(2), -1. / np.sqrt(2))]
    centers = [(scale * x, scale * y) for x,y in centers]
    while True:
        dataset = []
        for i in range(batchsz):
            point = np.random.randn(2) * 0.02  #随机采样两个数据
            center = random.choice(centers) #从八个高斯分布中随机选一个
            # N(0,1) + center_x1/x2 加上一个均值，方差还是为1
            point[0] += center[0]
            point[1] += center[1]
            dataset.append(point)
        dataset = np.array(dataset).astype(np.float32)
        dataset /= 1.414
        yield dataset #保存状态，返回dataset，下次运行时又继续从while处开始执行

def main():
    # 固定随机性
    torch.manual_seed(23)
    np.random.seed(23)

    data_iter = data_gernerator()
    x = next(data_iter)
    # print(x.shape) [b, 2]

    G = Generator(h_dim)
    D = Discriminator(h_dim)
    
    optim_G = optim.Adam(G.parameters(), lr = 5e-4, betas=(0.5, 0.9))
    optim_D = optim.Adam(D.parameters(), lr = 5e-4, betas=(0.5, 0.9))

    for epoch in range(50000):
        # Generator 和 Discriminator交替训练
        # 一般先训 Discrimator
        # 1. train Discriminator firstly, max(D(x)), min(D(G(z)))
        for _ in range(5):
            # 1.1 train on real data
            x = next(data_iter)
            x = torch.from_numpy(x)
            pred_real = D(x)  #真实数据的判决结果
            loss_real = -pred_real.mean() # max pred_real
            # 1.2 train on fake data
            z = torch.randn(batchsz, 2) #随机生成噪声
            x_fake = G(z).detach() #根据噪声生成假的数据 tf.stop_gradient() 
            pred_fake = D(x_fake) #假的数据的判决结果
            loss_fake = pred_fake.mean()

            # 1.3 aggreate loss
            loss_D = loss_real + loss_fake  #最大化真实数据的判决概率， 最小化生成数据的判决概率

            # 1.4 optimize
            optim_D.zero_grad()
            loss_D.backward()
            optim_D.step()

        # 2. train Generator， max D(G(z))
        z = torch.randn(batchsz, 2)  #随机生成噪声
        x_fake = G(z)  #不能detach， 因为D在G后面
        pred_fake = D(x_fake)  # D和G都有梯度反向传播，但只有G更新
        loss_G = - pred_fake.mean() # 最大化生成数据的判决概率

        # optimize
        optim_G.zero_grad() 
        loss_G.backward() #反向传播计算G和D的梯度
        optim_G.step()    #只更新G的参数

    
        if epoch % 100 == 0:
            print(loss_D.item(), loss_G.item()) 
            #会出现 0.0， -1.0两个值
            # 原因是Discriminator训练得很好，可以很好地将真假数据区分开来, 所以loss为0
            # 但由于Generator训练过程中， 使用JS散度不能衡量两个没有重叠的分布，使得网络没有梯度信息，
            # Generator长期得不到更新，所以loss长期处于-1.0. 


if __name__ == '__main__':
    main()

  
''' -0.14626020193099976 -0.47136804461479187 0.0 -1.0 0.0 -1.0 0.0 -1.0 0.0 -1.0 0.0 -1.0 0.0 -1.0 0.0 -1.0 '''
...

6. WGAN 实现

from re import L
import torch
from torch import nn, optim, autograd
import numpy as np
import visdom
import random

h_dim = 400
batchsz = 512
viz = visdom.Visdom()

class Generator(nn.Module):
    def __init__(self, h_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(2, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, 2)
        )
    def forward(self, x):
        return self.net(x)

class Discriminator(nn.Module):
    def __init__(self, h_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(2, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, h_dim),
            nn.ReLU(inplace=True),
            nn.Linear(h_dim, 1),  #输出分类概率
            nn.Sigmoid()
        )

    def forward(self, x):
        return self.net(x).view(-1)

def data_gernerator():
    scale = 2
    centers = [(1,0),(-1,0),(0,1),(0,-1),
               (1. / np.sqrt(2), 1. / np.sqrt(2)),
               (1. / np.sqrt(2), -1. / np.sqrt(2)),
               (-1. / np.sqrt(2), 1. / np.sqrt(2)),
               (-1. / np.sqrt(2), -1. / np.sqrt(2))]
    centers = [(scale * x, scale * y) for x,y in centers]
    while True:
        dataset = []
        for i in range(batchsz):
            point = np.random.randn(2) * 0.02  #随机采样两个数据
            center = random.choice(centers) #从八个高斯分布中随机选一个
            # N(0,1) + center_x1/x2 加上一个均值，方差还是为1
            point[0] += center[0]
            point[1] += center[1]
            dataset.append(point)
        dataset = np.array(dataset).astype(np.float32)
        dataset /= 1.414
        yield dataset #保存状态，返回dataset，下次运行时又继续从while处开始执行

def gradient_penalty(D, x_real, x_fake): #dskj1
    t = torch.rand(batchsz, 1) #随机sample
    # [b,1]=>[b,2]
    t = t.expand_as(x_real)
    x = t*x_real + (1-t)*x_fake  #真实数据和fake数据之间做一个线性插值
    x.requires_grad_() #设置x需要导数信息 
    pred = D(x)
    grads = autograd.grad(outputs = pred, inputs=x, 
                        grad_outputs=torch.ones_like(pred),
                        create_graph=True, #用于二阶求导
                        retain_graph=True, #如果还需要backward一次，就把这个梯度信息保留下来，否则会报错
                        only_inputs=True
                        )[0]
    gp = torch.pow((grads.norm(2, dim = 1) -1), 2).mean()
    return gp

def main():
    # 固定随机性
    torch.manual_seed(23)
    np.random.seed(23)

    data_iter = data_gernerator()
    x = next(data_iter)
    # print(x.shape) [b, 2]

    G = Generator(h_dim)
    D = Discriminator(h_dim)
    
    optim_G = optim.Adam(G.parameters(), lr = 5e-4, betas=(0.5, 0.9))
    optim_D = optim.Adam(D.parameters(), lr = 5e-4, betas=(0.5, 0.9))

    for epoch in range(50000):
        # Generator 和 Discriminator交替训练
        # 一般先训 Discrimator
        # 1. train Discriminator firstly, max(D(x)), min(D(G(z)))
        for _ in range(5):
            # 1.1 train on real data
            x = next(data_iter)
            x = torch.from_numpy(x)
            pred_real = D(x)  #真实数据的判决结果
            loss_real = -pred_real.mean() # max pred_real
            # 1.2 train on fake data
            z = torch.randn(batchsz, 2) #随机生成噪声
            x_fake = G(z).detach() #根据噪声生成假的数据 tf.stop_gradient() 
            pred_fake = D(x_fake) #假的数据的判决结果
            loss_fake = pred_fake.mean()
            # 1.3 gradient penalty
            gp = gradient_penalty(D, x, x_fake.detach()) #这里x_fake要detach一下，因为不需要对它进行求导

            # 1.4 aggreate loss
            loss_D = loss_real + loss_fake + 0.2 * gp  #最大化真实数据的判决概率， 最小化生成数据的判决概率

            # 1.4 optimize
            optim_D.zero_grad()
            loss_D.backward()
            optim_D.step()

        # 2. train Generator， max D(G(z))
        z = torch.randn(batchsz, 2)  #随机生成噪声
        x_fake = G(z)  #不能detach， 因为D在G后面
        pred_fake = D(x_fake)  # D和G都有梯度反向传播，但只有G更新
        loss_G = - pred_fake.mean() # 最大化生成数据的判决概率

        # optimize
        optim_G.zero_grad() 
        loss_G.backward() #反向传播计算G和D的梯度
        optim_G.step()    #只更新G的参数

    
        if epoch % 100 == 0:
            print(loss_D.item(), loss_G.item()) 
            #会出现 0.0， -1.0两个值
            # 原因是Discriminator训练得很好，可以很好地将真假数据区分开来, 所以loss为0
            # 但由于Generator训练过程中， 使用JS散度不能衡量两个没有重叠的分布，使得网络没有梯度信息，
            # Generator长期得不到更新，所以loss长期处于-1.0. 


if __name__ == '__main__':
    main()

''' 0.010273948311805725 -0.4702531099319458 -0.5763486623764038 -0.1938347965478897 -0.20412319898605347 -0.4627363085746765 -0.21898901462554932 -0.4972558617591858 -0.13111859560012817 -0.504425048828125 -0.1431763768196106 -0.4709075689315796 -0.09945613145828247 -0.5348900556564331 '''