1.什么是CGAN
在CGAN训练期间,生成器学习为训练数据集中的每个标签生成逼真的样本,而鉴别器则学习区分真的样本-标签对与假的样本-标签对。只学习接受真实且样本-标签匹配正确的对,拒绝不匹配的对和样本为假的对。
2.生成器
条件标签称为y,生成器使用噪声向量Z和标签y合成一个伪样本。
3.鉴别器
接受带标签的真实样本(x,y),以及生成器生成的伪样本
4.代码实现
4.1导入数据库
%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
from keras.datasets import mnist
from keras.layers import (Activation, BatchNormalization, Concatenate, Dense,
Embedding, Flatten, Input, Multiply, Reshape)
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import Conv2D, Conv2DTranspose
from keras.models import Model, Sequential
#此处如果报错,就将下面改为from keras.optimizer_v2 import adam as Adam
from keras.optimizers import Adam
4.2模型的输入维度
img_rows = 28
img_cols = 28
channels = 1
# 输入图像的维度
img_shape = (img_rows, img_cols, channels)
# 噪声向量的大小,用作生成器的输入
z_dim = 100
# 数据集中的类别数
num_classes = 10
4.3CGAN生成器
def build_generator(z_dim):
model = Sequential()
# 通过全连接层将输入变为7*7*256的张量
model.add(Dense(256 * 7 * 7, input_dim=z_dim))
model.add(Reshape((7, 7, 256)))
# 转置卷积层,张量从7*7*256变成14*14*128
model.add(Conv2DTranspose(128, kernel_size=3, strides=2, padding='same'))
# 批归一化
model.add(BatchNormalization())
# Leaky ReLU 激活函数
model.add(LeakyReLU(alpha=0.01))
# 转置卷积层,张量成14*14*128变成14*14*64
model.add(Conv2DTranspose(64, kernel_size=3, strides=1, padding='same'))
# 批归一化
model.add(BatchNormalization())
# Leaky ReLU 激活函数
model.add(LeakyReLU(alpha=0.01))
# 转置卷积层,14*14*64变成28*28*1
model.add(Conv2DTranspose(1, kernel_size=3, strides=2, padding='same'))
# 带有tanh的输出层
model.add(Activation('tanh'))
return model
def build_cgan_generator(z_dim):
# 随机噪声Z
z = Input(shape=(z_dim, ))
# 条件标签:G应该生成指定数字,整数0-9
label = Input(shape=(1, ), dtype='int32')
# 标签嵌入:
# ----------------
# 将标签转化为大小为z_dim的稠密向量
# 生成形状为 (batch_size, 1, z_dim)的三维张量
label_embedding = Embedding(num_classes, z_dim, input_length=1)(label)
# 将嵌入的三维张量展平为形状为 (batch_size, z_dim)的二维张量
label_embedding = Flatten()(label_embedding)
# 向量z和嵌入标签的元素级乘积
joined_representation = Multiply()([z, label_embedding])
generator = build_generator(z_dim)
# 为给定的标签生成图像
conditioned_img = generator(joined_representation)
return Model([z, label], conditioned_img)
1.使用Embedding层将标签y转换为大小z_dim的稠密向量
2.Multiply层将标签与噪声向量z嵌入联合表示中,就是将2个等长向量的对应项相乘,输出作为结果乘积的单个向量
4.4鉴别器
def build_discriminator(img_shape):
model = Sequential()
# 卷积层,从28*28*2变成14*14*64的张量
model.add(
Conv2D(64,
kernel_size=3,
strides=2,
input_shape=(img_shape[0], img_shape[1], img_shape[2] + 1),
padding='same'))
# Leaky ReLU 激活函数
model.add(LeakyReLU(alpha=0.01))
# 卷积层,从14*14*14变成7*7*64的张量
model.add(
Conv2D(64,
kernel_size=3,
strides=2,
input_shape=img_shape,
padding='same'))
# 批归一化
model.add(BatchNormalization())
# Leaky ReLU 激活函数
model.add(LeakyReLU(alpha=0.01))
# 卷积层,从7*7*64变成3*3*128的张量
model.add(
Conv2D(128,
kernel_size=3,
strides=2,
input_shape=img_shape,
padding='same'))
# 批归一化
model.add(BatchNormalization())
# Leaky ReLU的激活函数
model.add(LeakyReLU(alpha=0.01))
# 带有sigmoid激活函数的输出层
model.add(Flatten())
model.add(Dense(1, activation='sigmoid'))
return model
def build_cgan_discriminator(img_shape):
# 输入图像
img = Input(shape=img_shape)
# 为输入图像加标签
label = Input(shape=(1, ), dtype='int32')
# 标签嵌入:
# ----------------
# 将标签转化为大小为z_dim的稠密向量
# 生成形状为 (batch_size, 1, 28*28*1)的3维向量
label_embedding = Embedding(num_classes,
np.prod(img_shape),
input_length=1)(label)
# 将嵌入的三维张量展平成形状为 (batch_size, 28*28*1)的二维张量
label_embedding = Flatten()(label_embedding)
# 将嵌入标签调整为和输入图像一样的维度
label_embedding = Reshape(img_shape)(label_embedding)
# 将图像与其嵌入标签连接
concatenated = Concatenate(axis=-1)([img, label_embedding])
discriminator = build_discriminator(img_shape)
# 将图像-标签对进行分类
classification = discriminator(concatenated)
return Model([img, label], classification)
4.5搭建整个模型
def build_cgan(generator, discriminator):
# 随机噪声向量z
z = Input(shape=(z_dim, ))
# 图像标签
label = Input(shape=(1, ))
#为指定标签生成图像
img = generator([z, label])
classification = discriminator([img, label])
# Combined Generator -> Discriminator model
# G([z, lablel]) = x*
# D(x*) = classification
model = Model([z, label], classification)
return model
# 构建并编译鉴别器,如果报错,替换optimizer=Adam.Adam())
discriminator = build_cgan_discriminator(img_shape)
discriminator.compile(loss='binary_crossentropy',
optimizer=Adam(),
metrics=['accuracy'])
#构建生成器
generator = build_cgan_generator(z_dim)
# 生成器训练时鉴别器参数保持不变
discriminator.trainable = False
# 构建并编译鉴别器固定的CGAN模型来训练生成器
cgan = build_cgan(generator, discriminator)
cgan.compile(loss='binary_crossentropy', optimizer=Adam())
4.6训练
accuracies = []
losses = []
def train(iterations, batch_size, sample_interval):
# 导入mnist数据集
(X_train, y_train), (_, _) = mnist.load_data()
# 灰度像素值从[0,255]缩放到[-1, 1]
X_train = X_train / 127.5 - 1.
X_train = np.expand_dims(X_train, axis=3)
# 真实图像的标签都为1
real = np.ones((batch_size, 1))
# 假图像的标签都为0
fake = np.zeros((batch_size, 1))
for iteration in range(iterations):
# -------------------------
# 训练鉴定器
# -------------------------
# 生成一批量伪样本及其标签
idx = np.random.randint(0, X_train.shape[0], batch_size)
imgs, labels = X_train[idx], y_train[idx]
# 生成一批为图像
z = np.random.normal(0, 1, (batch_size, z_dim))
gen_imgs = generator.predict([z, labels])
# 训练鉴别器
d_loss_real = discriminator.train_on_batch([imgs, labels], real)
d_loss_fake = discriminator.train_on_batch([gen_imgs, labels], fake)
d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
# ---------------------
# 训练生成器
# ---------------------
#生成一批噪声向量
z = np.random.normal(0, 1, (batch_size, z_dim))
# 得到一批次随机标签
labels = np.random.randint(0, num_classes, batch_size).reshape(-1, 1)
# 训练生成器
g_loss = cgan.train_on_batch([z, labels], real)
if (iteration + 1) % sample_interval == 0:
# 输出训练过程
print("%d [D loss: %f, acc.: %.2f%%] [G loss: %f]" %
(iteration + 1, d_loss[0], 100 * d_loss[1], g_loss))
losses.append((d_loss[0], g_loss))
accuracies.append(100 * d_loss[1])
sample_images()
4.7显示生成图像
def sample_images(image_grid_rows=2, image_grid_columns=5):
# 随机噪声采样
z = np.random.normal(0, 1, (image_grid_rows * image_grid_columns, z_dim))
# 得到图像标签0-9
labels = np.arange(0, 10).reshape(-1, 1)
# 从随机噪声生成图像
gen_imgs = generator.predict([z, labels])
# 图像像素缩放到[0,1]
gen_imgs = 0.5 * gen_imgs + 0.5
# 设置图像网格
fig, axs = plt.subplots(image_grid_rows,
image_grid_columns,
figsize=(10, 4),
sharey=True,
sharex=True)
cnt = 0
for i in range(image_grid_rows):
for j in range(image_grid_columns):
输出图像网格
axs[i, j].imshow(gen_imgs[cnt, :, :, 0], cmap='gray')
axs[i, j].axis('off')
axs[i, j].set_title("Digit: %d" % labels[cnt])
cnt += 1
4.8训练模型
# 设置超参数
iterations = 12000
batch_size = 32
sample_interval = 1000
# 训练模型直到指定的迭代次数
train(iterations, batch_size, sample_interval)