第二次作业：卷积神经网络 part 2

Part Ⅰ 问题总结

Part Ⅱ 生成式对抗网络

2.1 用途

• 旧图像修复
• 图像超像素
• 人脸生成
• 人脸定制
• 文本生成图片
• 字体变换
• 风格变换
• 帧预测

2.2 GAN

生成式对抗网络的目的是训练一个生成模型，生成我们想要的数据。（低维向量-》生成模型-》高维数据（图片、文本、语音））

Part Ⅲ 代码练习

3.1 生成式对抗网络

定义生成器和判别器

z_dim = 32
hidden_dim = 128

# 定义生成器
net_G = nn.Sequential(
            nn.Linear(z_dim,hidden_dim),
            nn.ReLU(), 
            nn.Linear(hidden_dim, 2))

# 定义判别器
net_D = nn.Sequential(
            nn.Linear(2,hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim,1),
            nn.Sigmoid())

# 网络放到 GPU 上
net_G = net_G.to(device)
net_D = net_D.to(device)

# 定义网络的优化器
optimizer_G = torch.optim.Adam(net_G.parameters(),lr=0.0001)
optimizer_D = torch.optim.Adam(net_D.parameters(),lr=0.0001)

开始训练

batch_size = 50
nb_epochs = 1000

loss_D_epoch = []
loss_G_epoch = []

for e in range(nb_epochs):
    np.random.shuffle(X)
    real_samples = torch.from_numpy(X).type(torch.FloatTensor)
    loss_G = 0
    loss_D = 0
    for t, real_batch in enumerate(real_samples.split(batch_size)):
        # 固定生成器G，改进判别器D
        # 使用normal_()函数生成一组随机噪声，输入G得到一组样本
        z = torch.empty(batch_size,z_dim).normal_().to(device)
        fake_batch = net_G(z)
        # 将真、假样本分别输入判别器，得到结果
        D_scores_on_real = net_D(real_batch.to(device))
        D_scores_on_fake = net_D(fake_batch)
        # 优化过程中，假样本的score会越来越小，真样本的score会越来越大，下面 loss 的定义刚好符合这一规律，
        # 要保证loss越来越小，真样本的score前面要加负号
        # 要保证loss越来越小，假样本的score前面是正号（负负得正）
        loss = -torch.mean(torch.log(1-D_scores_on_fake) + torch.log(D_scores_on_real))
        # 梯度清零
        optimizer_D.zero_grad()
        # 反向传播优化
        loss.backward()
        # 更新全部参数
        optimizer_D.step()
        loss_D += loss
                    
        # 固定判别器，改进生成器
        # 生成一组随机噪声，输入生成器得到一组假样本
        z = torch.empty(batch_size,z_dim).normal_().to(device)
        fake_batch = net_G(z)
        # 假样本输入判别器得到 score
        D_scores_on_fake = net_D(fake_batch)
        # 我们希望假样本能够骗过生成器，得到较高的分数，下面的 loss 定义也符合这一规律
        # 要保证 loss 越来越小，假样本的前面要加负号
        loss = -torch.mean(torch.log(D_scores_on_fake))
        optimizer_G.zero_grad()
        loss.backward()
        optimizer_G.step()
        loss_G += loss
    
    if e % 50 ==0:
        print(f'\n Epoch {e} , D loss: {loss_D}, G loss: {loss_G}') 

    loss_D_epoch.append(loss_D)
    loss_G_epoch.append(loss_G)

展示loss的变化情况：

plt.plot(loss_D_epoch)
plt.plot(loss_G_epoch)

可以发现直到训练结束，模型都没有收敛。
利用生成器生成一组假样本，观察是否符合两个半月形状的数据分布：

z = torch.empty(n_samples,z_dim).normal_().to(device)
fake_samples = net_G(z)
fake_data = fake_samples.cpu().data.numpy()

fig, ax = plt.subplots(1, 1, facecolor='#4B6EA9')
all_data = np.concatenate((X,fake_data),axis=0)
Y2 = np.concatenate((np.ones(n_samples),np.zeros(n_samples)))
plot_data(ax, all_data, Y2)
plt.show()

其中，白色的是原来的真实样本，黑色的点是生成器生成的样本。效果并不好！
进行改进：把学习率修改为 0.001，batch_size改大到250，在运行一次

# 定义网络的优化器
optimizer_G = torch.optim.Adam(net_G.parameters(),lr=0.001)
optimizer_D = torch.optim.Adam(net_D.parameters(),lr=0.001)

batch_size = 250

loss_D_epoch = []
loss_G_epoch = []

for e in range(nb_epochs):
    np.random.shuffle(X)
    real_samples = torch.from_numpy(X).type(torch.FloatTensor)
    loss_G = 0
    loss_D = 0
    for t, real_batch in enumerate(real_samples.split(batch_size)):
        # 固定生成器G，改进判别器D
        # 使用normal_()函数生成一组随机噪声，输入G得到一组样本
        z = torch.empty(batch_size,z_dim).normal_().to(device)
        fake_batch = net_G(z)
        # 将真、假样本分别输入判别器，得到结果
        D_scores_on_real = net_D(real_batch.to(device))
        D_scores_on_fake = net_D(fake_batch)
        # 优化过程中，假样本的score会越来越小，真样本的score会越来越大，下面 loss 的定义刚好符合这一规律，
        # 要保证loss越来越小，真样本的score前面要加负号
        # 要保证loss越来越小，假样本的score前面是正号（负负得正）
        loss = -torch.mean(torch.log(1-D_scores_on_fake) + torch.log(D_scores_on_real))
        # 梯度清零
        optimizer_D.zero_grad()
        # 反向传播优化
        loss.backward()
        # 更新全部参数
        optimizer_D.step()
        loss_D += loss
                    
        # 固定判别器，改进生成器
        # 生成一组随机噪声，输入生成器得到一组假样本
        z = torch.empty(batch_size,z_dim).normal_().to(device)
        fake_batch = net_G(z)
        # 假样本输入判别器得到 score
        D_scores_on_fake = net_D(fake_batch)
        # 我们希望假样本能够骗过生成器，得到较高的分数，下面的 loss 定义也符合这一规律
        # 要保证 loss 越来越小，假样本的前面要加负号
        loss = -torch.mean(torch.log(D_scores_on_fake))
        optimizer_G.zero_grad()
        loss.backward()
        optimizer_G.step()
        loss_G += loss
    
    if e % 50 ==0:
        print(f'\n Epoch {e} , D loss: {loss_D}, G loss: {loss_G}') 

    loss_D_epoch.append(loss_D)
    loss_G_epoch.append(loss_G)

展示结果

z = torch.empty(n_samples,z_dim).normal_().to(device)
fake_samples = net_G(z)
fake_data = fake_samples.cpu().data.numpy()

fig, ax = plt.subplots(1, 1, facecolor='#4B6EA9')
all_data = np.concatenate((X,fake_data),axis=0)
Y2 = np.concatenate((np.ones(n_samples),np.zeros(n_samples)))
plot_data(ax, all_data, Y2)
plt.show()

随着batch size的增大，loss的降低，效果明显改善
下面生成更多的样本观察一下

z = torch.empty(10*n_samples,z_dim).normal_().to(device)
fake_samples = net_G(z)
fake_data = fake_samples.cpu().data.numpy()
fig, ax = plt.subplots(1, 1, facecolor='#4B6EA9')
all_data = np.concatenate((X,fake_data),axis=0)
Y2 = np.concatenate((np.ones(n_samples),np.zeros(10*n_samples)))
plot_data(ax, all_data, Y2)
plt.show();

3.2 CGAN

首先下载数据集

import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
import numpy as np
import matplotlib.pyplot as plt

# 基本参数
z_dim = 100
batch_size = 128
learning_rate = 0.0002
total_epochs = 30

# 使用GPU训练，可以在菜单 "代码执行工具" -> "更改运行时类型" 里进行设置
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# 加载MNIST数据集
dataloader = torch.utils.data.DataLoader(
    datasets.MNIST('./data', train=True, download=True,
        transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,), (0.3081,))])
        ), batch_size, shuffle=False, drop_last=True)

实现CGAN，分别是生成器和判别器的网络结构。

class Discriminator(nn.Module):
	'''全连接判别器，用于1x28x28的MNIST数据,输出是数据和类别'''
	def __init__(self):
		super(Discriminator, self).__init__()
		self.model = nn.Sequential(
			  nn.Linear(28*28+10, 512),
			  nn.LeakyReLU(0.2, inplace=True),
			  nn.Linear(512, 256),
			  nn.LeakyReLU(0.2, inplace=True),
			  nn.Linear(256, 1),
			  nn.Sigmoid()
		)
  
	def forward(self, x, c):
		x = x.view(x.size(0), -1)
		validity = self.model(torch.cat([x, c], -1))
		return validity

class Generator(nn.Module):
	'''全连接生成器，用于1x28x28的MNIST数据，输入是噪声和类别'''
	def __init__(self, z_dim):
		super(Generator, self).__init__()
		self.model = nn.Sequential(
			  nn.Linear(z_dim+10, 128),
			  nn.LeakyReLU(0.2, inplace=True),
			  nn.Linear(128, 256),
			  nn.BatchNorm1d(256, 0.8),
			  nn.LeakyReLU(0.2, inplace=True),
			  nn.Linear(256, 512),
			  nn.BatchNorm1d(512, 0.8),
			  nn.LeakyReLU(0.2, inplace=True),
			  nn.Linear(in_features=512, out_features=28*28),
			  nn.Tanh()
	 	)

	def forward(self, z, c):
		x = self.model(torch.cat([z, c], dim=1))
		x = x.view(-1, 1, 28, 28)
		return x

定义相关的模型

# 初始化构建判别器和生成器
discriminator = Discriminator().to(device)
generator = Generator(z_dim=z_dim).to(device)

# 初始化二值交叉熵损失
bce = torch.nn.BCELoss().to(device)
ones = torch.ones(batch_size).to(device)
zeros = torch.zeros(batch_size).to(device)

# 初始化优化器，使用Adam优化器
g_optimizer = optim.Adam(generator.parameters(), lr=learning_rate)
d_optimizer = optim.Adam(discriminator.parameters(), lr=learning_rate)

开始训练

# 开始训练，一共训练total_epochs
for epoch in range(total_epochs):

	# torch.nn.Module.train() 指的是模型启用 BatchNormalization 和 Dropout
	# torch.nn.Module.eval() 指的是模型不启用 BatchNormalization 和 Dropout
	# 因此，train()一般在训练时用到， eval() 一般在测试时用到
	generator = generator.train()

	# 训练一个epoch
	for i, data in enumerate(dataloader):

		# 加载真实数据
		real_images, real_labels = data
		real_images = real_images.to(device)
		# 把对应的标签转化成 one-hot 类型
		tmp = torch.FloatTensor(real_labels.size(0), 10).zero_()
		real_labels = tmp.scatter_(dim=1, index=torch.LongTensor(real_labels.view(-1, 1)), value=1)
		real_labels = real_labels.to(device)

		# 生成数据
		# 用正态分布中采样batch_size个随机噪声
		z = torch.randn([batch_size, z_dim]).to(device)
		# 生成 batch_size 个 ont-hot 标签
		c = torch.FloatTensor(batch_size, 10).zero_()
		c = c.scatter_(dim=1, index=torch.LongTensor(np.random.choice(10, batch_size).reshape([batch_size, 1])), value=1)
		c = c.to(device)
		# 生成数据
		fake_images = generator(z,c)

		# 计算判别器损失，并优化判别器
		real_loss = bce(discriminator(real_images, real_labels), ones)
		fake_loss = bce(discriminator(fake_images.detach(), c), zeros)
		d_loss = real_loss + fake_loss

		d_optimizer.zero_grad()
		d_loss.backward()
		d_optimizer.step()

		# 计算生成器损失，并优化生成器
		g_loss = bce(discriminator(fake_images, c), ones)

		g_optimizer.zero_grad()
		g_loss.backward()
		g_optimizer.step()

	# 输出损失
	print("[Epoch %d/%d] [D loss: %f] [G loss: %f]" % (epoch, total_epochs, d_loss.item(), g_loss.item()))

下面我们用随机噪声生成一组图像，看看CGAN的效果

#用于生成效果图
# 生成100个随机噪声向量
fixed_z = torch.randn([100, z_dim]).to(device)
# 生成100个one_hot向量，每类10个
fixed_c = torch.FloatTensor(100, 10).zero_()
fixed_c = fixed_c.scatter_(dim=1, index=torch.LongTensor(np.array(np.arange(0, 10).tolist()*10).reshape([100, 1])), value=1)
fixed_c = fixed_c.to(device)

generator = generator.eval()
fixed_fake_images = generator(fixed_z, fixed_c)

plt.figure(figsize=(8, 8))
for j in range(10):
    for i in range(10):
        img = fixed_fake_images[j*10+i, 0, :, :].detach().cpu().numpy()
        img = img.reshape([28, 28])
        plt.subplot(10, 10, j*10+i+1)
        plt.imshow(img, 'gray')

3.3 DCGAN

判别器 和 生成器 的网络结构

class D_dcgan(nn.Module):
	'''滑动卷积判别器'''
	def __init__(self):
		super(D_dcgan, self).__init__()
		self.conv = nn.Sequential(
            # 第一个滑动卷积层，不使用BN，LRelu激活函数
            nn.Conv2d(in_channels=1, out_channels=16, kernel_size=3, stride=2, padding=1),
            nn.LeakyReLU(0.2, inplace=True),
            # 第二个滑动卷积层，包含BN，LRelu激活函数
            nn.Conv2d(in_channels=16, out_channels=32, kernel_size=3, stride=2, padding=1),
            nn.BatchNorm2d(32),
            nn.LeakyReLU(0.2, inplace=True),
            # 第三个滑动卷积层，包含BN，LRelu激活函数
            nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.BatchNorm2d(64),
            nn.LeakyReLU(0.2, inplace=True),
            # 第四个滑动卷积层，包含BN，LRelu激活函数
            nn.Conv2d(in_channels=64, out_channels=128, kernel_size=4, stride=1),
            nn.BatchNorm2d(128),
            nn.LeakyReLU(0.2, inplace=True)
        )

		# 全连接层+Sigmoid激活函数
		self.linear = nn.Sequential(nn.Linear(in_features=128, out_features=1), nn.Sigmoid())

	def forward(self, x):
		x = self.conv(x)
		x = x.view(x.size(0), -1)
		validity = self.linear(x)
		return validity

class G_dcgan(nn.Module):
	'''反滑动卷积生成器'''

	def __init__(self, z_dim):
		super(G_dcgan, self).__init__()
		self.z_dim = z_dim
		# 第一层：把输入线性变换成256x4x4的矩阵，并在这个基础上做反卷机操作
		self.linear = nn.Linear(self.z_dim, 4*4*256)
		self.model = nn.Sequential(
            # 第二层：bn+relu
            nn.ConvTranspose2d(in_channels=256, out_channels=128, kernel_size=3, stride=2, padding=0),
            nn.BatchNorm2d(128),
            nn.ReLU(inplace=True),
            # 第三层：bn+relu
            nn.ConvTranspose2d(in_channels=128, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            # 第四层:不使用BN，使用tanh激活函数
            nn.ConvTranspose2d(in_channels=64, out_channels=1, kernel_size=4, stride=2, padding=2),
            nn.Tanh()
        )

	def forward(self, z):
		# 把随机噪声经过线性变换，resize成256x4x4的大小
		x = self.linear(z)
		x = x.view([x.size(0), 256, 4, 4])
		# 生成图片
		x = self.model(x)
		return x

定义相关的模型

# 构建判别器和生成器
d_dcgan = D_dcgan().to(device)
g_dcgan = G_dcgan(z_dim=z_dim).to(device)

def weights_init_normal(m):
    classname = m.__class__.__name__
    if classname.find('Conv') != -1:
        torch.nn.init.normal_(m.weight.data, 0.0, 0.02)
    elif classname.find('BatchNorm2d') != -1:
        torch.nn.init.normal_(m.weight.data, 1.0, 0.02)
        torch.nn.init.constant_(m.bias.data, 0.0)

# 使用均值为0，方差为0.02的正态分布初始化神经网络
d_dcgan.apply(weights_init_normal)
g_dcgan.apply(weights_init_normal)

# 初始化优化器，使用Adam优化器
g_dcgan_optim = optim.Adam(g_dcgan.parameters(), lr=learning_rate)
d_dcgan_optim = optim.Adam(d_dcgan.parameters(), lr=learning_rate)

# 加载MNIST数据集，和之前不同的是，DCGAN输入的图像被 resize 成 32*32 像素
dcgan_dataloader = torch.utils.data.DataLoader(
    datasets.MNIST('./data', train=True, download=True,
        transform=transforms.Compose([transforms.Resize(32), transforms.ToTensor(),transforms.Normalize((0.5,), (0.5,))])
        ), batch_size, shuffle=True, drop_last=True)

开始训练模型

# 开始训练，一共训练 total_epochs

for e in range(total_epochs):

	# 给generator启用 BatchNormalization
	g_dcgan = g_dcgan.train()
	# 训练一个epoch
	for i, data in enumerate(dcgan_dataloader):

		# 加载真实数据，不加载标签
		real_images, _ = data
		real_images = real_images.to(device)

		# 用正态分布中采样batch_size个噪声，然后生成对应的图片
		z = torch.randn([batch_size, z_dim]).to(device)
		fake_images = g_dcgan(z)

		# 计算判别器损失，并优化判别器
		real_loss = bce(d_dcgan(real_images), ones)
		fake_loss = bce(d_dcgan(fake_images.detach()), zeros)
		d_loss = real_loss + fake_loss

		d_dcgan_optim.zero_grad()
		d_loss.backward()
		d_dcgan_optim.step()

		# 计算生成器损失，并优化生成器
		g_loss = bce(d_dcgan(fake_images), ones)

		g_dcgan_optim.zero_grad()
		g_loss.backward()
		g_dcgan_optim.step()
		
    # 输出损失
	print ("[Epoch %d/%d] [D loss: %f] [G loss: %f]" % (e, total_epochs, d_loss.item(), g_loss.item()))

用一组随机噪声输出图像，看看DCGAN的效果

#用于生成效果图
# 生成100个随机噪声向量
fixed_z = torch.randn([100, z_dim]).to(device)
g_dcgan = g_dcgan.eval()
fixed_fake_images = g_dcgan(fixed_z)

plt.figure(figsize=(8, 8))
for j in range(10):
    for i in range(10):
        img = fixed_fake_images[j*10+i, 0, :, :].detach().cpu().numpy()
        img = img.reshape([32, 32])
        plt.subplot(10, 10, j*10+i+1)
        plt.imshow(img, 'gray')