CGAN是最早使目标数据生成成为可能的GAN创新之一,可以说是最具影响力的一种。接下来,介绍CGAN的工作方式以及如何用MNIST数据集实现它的小规模版本。
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。半监督GAN的鉴别器除了区分真实样本与伪样本,还为每个真实样本分配正确的标签;而CGAN中的鉴别器不会学习识别哪个样本是哪个类。它只学习接受真实的且样本-标签匹配正确的对,拒绝不匹配的对和样本为假的对。
例如:无论样本1是真是假,CGAN的判别器都拒绝该(样本1与标签2)对,为了欺骗鉴别器,CGAN生成器仅生成逼真的数据是不够的,生成的样本还需要与标签相匹配。在对生成器进行充分训练之后,就可以通过传递所需的标签来指定希望CGAN合成的样本。
接受带标签的真实样本(x,y),以及带有标签的伪样本(x*|y,y),在真实样本-标签对上,鉴别器学习如何识别真实数据以及如何识别匹配对。在生成器生成的样本中,鉴别器学习识别伪样本-标签对,以将它们与真实样本-标签对区分开来。
判别器输出表明输入是真实的匹配对的概率,它的目标是学会接受所有的真实样本-标签对,并拒绝所有伪样本和所有与标签不匹配的样本。
对于每个伪样本,相同的标签y同时被传递给生成器和鉴别器。另外,通过在带有不匹配标签的真实样本上训练鉴别器来拒绝不匹配的对;它识别不匹配对的能力是被训练成只接收真实匹配对时的副产品。
# 导入包
%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from keras import backend as K
from tensorflow.keras.datasets import mnist
from keras.layers import Embedding, Multiply, Dropout, Lambda, Concatenate, Input, Dense, Flatten, Reshape, Activation, BatchNormalization
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import Conv2D, Conv2DTranspose
from keras.models import Sequential, Model
from keras.optimizers import Adam
from keras.utils import to_categorical
Using TensorFlow backend.
# 模型输入维度
img_rows = 28
img_cols = 28
channels = 1
# 图像大小
img_shape = (img_rows, img_cols, channels)
# 噪声向量大小
z_dim = 100
num_classes = 10
(1)使用Keras的Embedding层将标签y(0到9的整数)转换为大小为z_dim(随机噪声向量的长度)的稠密向量。
(2)使用Keras的Multiply层将标签与噪声向量z嵌入联合表示中。顾名思义,该层将两个等长向量的对应项相乘,并输出作为结果乘积的单个向量。
(3)将得到的向量作为输入,保留CGAN生成器网络的其余部分以合成图像。
def build_generator(z_dim):
model = Sequential()
model.add(Dense(256 * 7 * 7, input_dim=z_dim))
model.add(Reshape((7, 7, 256)))
model.add(Conv2DTranspose(128, kernel_size=3, strides=2, padding='same'))
model.add(BatchNormalization())
model.add(LeakyReLU(alpha=0.01))
model.add(Conv2DTranspose(64, kernel_size=3, strides=1, padding='same'))
model.add(BatchNormalization())
model.add(LeakyReLU(alpha=0.01))
model.add(Conv2DTranspose(1, kernel_size=3, strides=2, padding='same'))
model.add(Activation('tanh'))
return model
def build_cgan_genertator(z_dim):
z = Input(shape=(z_dim, ))
label = Input(shape=(1,), dtype='int32')
label_embedding = Embedding(num_classes, z_dim, input_length=1)(label)
label_embedding = Flatten()(label_embedding)
joined_representation = Multiply()([z, label_embedding])
generator = build_generator(z_dim)
conditioned_img = generator(joined_representation)
return Model([z, label], conditioned_img)
步骤:
(1)取一个标签(0到9的整数),使用Keras的Embedding层将标签变成大小为28 × 28 × 1 = 784(扁平化图像的长度)的稠密向量。
(2)将嵌入标签调整为图像尺寸(28 × 28 × 1)。
(3)将重塑后的嵌入标签连接到对应图像上,生成形状(28 × 28× 2)的联合表示。可以将其视为在顶部“贴有”嵌入标签的图像。
(4)将图像-标签的联合表示输入CGAN的鉴别器网络中。注意,为了训练正常进行,必须将模型输入尺寸调整为(28 × 28 × 2)来对应新的输入形状。
def build_discriminator(img_shape):
model = Sequential()
model.add(Conv2D(32, kernel_size=3, strides=2, input_shape=(img_shape[0], img_shape[1], img_shape[2]+1),padding='same'))
model.add(LeakyReLU(alpha=0.01))
model.add(Conv2D(64, kernel_size=3, strides=2, input_shape=img_shape,padding='same'))
model.add(BatchNormalization())
model.add(LeakyReLU(alpha=0.01))
model.add(Conv2D(128, kernel_size=3, strides=2, input_shape=img_shape,padding='same'))
model.add(BatchNormalization())
model.add(LeakyReLU(alpha=0.01))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(1, activation='sigmoid'))
return model
def build_cgan_discriminator(img_shape):
img = Input(shape=img_shape)
label = Input(shape=(1, ), dtype='int32')
label_embedding = Embedding(num_classes, np.prod(img_shape), input_length=1)(label)
label_embedding = Flatten()(label_embedding)
label_embedding = Reshape(img_shape)(label_embedding) # 将标签调整和输入图像一样的维度
concatenated = Concatenate(axis= -1)([img, label_embedding])# 将图像与其嵌入标签链接
discriminator = build_discriminator(img_shape)
classification = discriminator(concatenated)
return Model([img, label], classification)
def build_cgan(generator, discriminator):
z = Input(shape=(z_dim, ))
label = Input(shape=(1, ))
img = generator([z, label])
classification = discriminator([img, label])
model = Model([z, label], classification)
return model
discriminator = build_cgan_discriminator(img_shape)
discriminator.compile(loss='binary_crossentropy', optimizer=Adam(), metrics=['accuracy'])
generator = build_cgan_genertator(z_dim)
discriminator.trainable = False
cgan = build_cgan(generator, discriminator)
cgan.compile(loss='binary_crossentropy', optimizer=Adam())
losses = []
accuracies = []
def train(iterations, batch_size, sample_interval):
(X_train, y_train), (_, _) = mnist.load_data('./MNIST')
X_train = X_train / 127.5 - 1.0
X_train = np.expand_dims(X_train, axis=3)
real = np.ones((batch_size, 1))
fake = np.zeros((batch_size, 1))
for iteration in range(iterations):
idx = np.random.randint(0, X_train.shape[0], batch_size)
# print(X_train.shape[0])
imgs, labels = X_train[idx], y_train[idx]
z = np.random.normal(0, 1, (batch_size, z_dim))
gen_imgs = generator.predict([z, labels])
# print(imgs.shape)
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)
z = np.random.normal(0, 1, (batch_size, z_dim))
labels = np.random.randint(0, num_classes, batch_size).reshape(-1, 1)
g_loss = cgan.train_on_batch([z, labels], real)
if (iteration + 1) % sample_interval == 0:
losses.append((d_loss[0], g_loss))
accuracies.append(100.0 * d_loss[1])
print("%d [D loss: %f, acc.: %.2f%%] [G loss:%f]"%(iteration + 1, d_loss[0], 100.0 * d_loss[1], g_loss))
sample_images()
def sample_images (image_grid_rows=2, image_grid_columns=5):
z = np.random.normal(0, 1, (image_grid_rows * image_grid_columns, z_dim))
labels = np.arange(0, 10).reshape(-1, 1)
gen_imgs = generator.predict([z, labels])
gen_imgs = 0.5 * gen_imgs + 0.5
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):
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
iterations = 12000
batch_size = 32
sample_interval = 1000
train(iterations, batch_size, sample_interval)
1000 [D loss: 0.000204, acc.: 100.00%] [G loss:9.885448]
2000 [D loss: 0.000059, acc.: 100.00%] [G loss:9.908726]
3000 [D loss: 0.230777, acc.: 90.62%] [G loss:4.183795]
4000 [D loss: 0.040735, acc.: 98.44%] [G loss:3.380749]
5000 [D loss: 0.192189, acc.: 90.62%] [G loss:3.410103]
6000 [D loss: 0.134279, acc.: 98.44%] [G loss:3.005539]
7000 [D loss: 0.412724, acc.: 82.81%] [G loss:1.312850]
8000 [D loss: 0.211682, acc.: 90.62%] [G loss:3.666016]
9000 [D loss: 0.080928, acc.: 98.44%] [G loss:7.182220]
10000 [D loss: 0.107635, acc.: 98.44%] [G loss:2.332113]
11000 [D loss: 0.194184, acc.: 93.75%] [G loss:3.737709]
12000 [D loss: 0.191671, acc.: 89.06%] [G loss:4.127837]
训练1000次
CGAN实现了不光是生成类似于真的样本而且还要生成一个符合条件的真实样本。通过增加生成器判别器的输入进一步提高了GAN的功能。
Github地址:https://github.com/yunlong-G/tensorflow_learn/blob/master/GAN/CGAN.ipynb