一、SGAN介绍
SGAN来源于这篇论文:《Semi-Supervised Learning with Generative Adversarial Networks》
传统的机器学习分为监督式学习和无监督式学习。前者的数据是有标签的,后者的数据是无标签的。然而,在很多问题中,有标签的数据是非常少的,要想获得有标签的数据,需要人工标注等一些操作。而无标签的数据则比较容易获得。半监督学习就是要结合监督式和无监督式,利用少量标签数据与大量无标签数据进行训练,然后,实现对未标签数据进行分类。
在生成对抗网络中,真实数据可以被看做有标签数据集,生成器随机产生的数据则可以被看做是无标签数据集。在DCGAN中,使用生成模型特征提取后的判别器已经可以实现分类的效果。首先,由判别器D学习到的特征可以提升分类器C的效果。那么一个好的分类器也可以优化判别器的最终效果。而C和D是无法同时训练的。然后C和D是可以相互促进提升的。而提升了判别器的能力。生成器G的效果也会随之变得更好,三者会在一个交替过程中趋向一个理想的平衡点。
因次,提出了一个半监督式GAN,称之为SGAN。希望能够同时训练生成器与半监督式分类器,最终实现一个更优的半监督式分类器,以及一个成像更高的生成模型。在传统的二分类模式基础上,SGAN变成了多分类,类型数量为N+1。分别指代N个标签和一个“假”数据。在实际过程中,判别器和分类器是一体的,记作D/C。共同与生成器G形成一个博弈关系。目标函数为负向最大似然估计(NLL)。下面为伪代码:
输入:I:总迭代次数
for n = 1, ... , I do
从生成器前置随机分布取出m个随机样本
从真实数据分布取出m个真实样本
最小化NLL,更新D/C的参数
从生成器前置随机分布取出m个随机样本
最大化NLL,更新G的参数
end for
相比较于cGAN,SGAN的生成是随机的。而且判别器的输出是一个分类器和判别器的结合。结构图如下:
二、代码实现
1. 导包
rom __future__ import print_function, division
from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout
from keras.layers import BatchNormalization, Activation, ZeroPadding2D
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import UpSampling2D, Conv2D
from keras.models import Sequential, Model
from keras.optimizers import Adam
from keras.utils import to_categorical
import matplotlib.pyplot as plt
import numpy as np
2. 初始化
class SGAN():
def __init__(self):
self.img_rows = 28
self.img_cols = 28
self.channels = 1
self.img_shape = (self.img_cols, self.img_rows, self.channels)
self.num_classes = 10
self.latent_dim = 100
optimizer = Adam(0.0002, 0.5)
# 构建判别器并编译
self.discriminator = self.build_discriminator()
self.discriminator.compile(loss=['binary_crossentropy', 'categorical_crossentropy'],
loss_weights=[0.5, 0.5],
optimizer=optimizer,
metrics=['accuracy'])
# 构建生成器
self.generator = self.build_generator()
noise = Input(shape=(100,))
img = self.generator(noise)
# 固定判别器
self.discriminator.trainable = False
# 生成器生成图像的判别结果
valid, _ = self.discriminator(img)
# 编译模型, 生成器和判别器的堆叠
self.combined = Model(noise, valid)
self.combined.compile(loss=['binary_crossentropy'],
optimizer=optimizer)
3. 构建生成器
def build_generator(self):
model = Sequential()
model.add(Dense(128 * 7 * 7, activation='relu', input_dim=self.latent_dim))
model.add(Reshape((7, 7, 128)))
model.add(BatchNormalization(momentum=0.8))
model.add(UpSampling2D())
model.add(Conv2D(128, kernel_size=3, padding='same'))
model.add(Activation('relu'))
model.add(BatchNormalization(momentum=0.8))
model.add(UpSampling2D())
model.add(Conv2D(64, kernel_size=3, padding='same'))
model.add(Activation('relu'))
model.add(BatchNormalization(momentum=0.8))
model.add(Conv2D(1, kernel_size=3, padding='same'))
model.add(Activation('tanh'))
model.summary()
noise = Input(shape=(self.latent_dim,))
img = model(noise)
return Model(noise, img)
4. 构建判别器
def build_discriminator(self):
model = Sequential()
model.add(Conv2D(32, kernel_size=3, strides=2, input_shape=self.img_shape, padding='same'))
model.add(LeakyReLU(alpha=0.2))
model.add(Dropout(0.25))
model.add(Conv2D(64, kernel_size=3, strides=2, padding='same'))
model.add(ZeroPadding2D(padding=((0, 1), (0, 1))))
model.add(LeakyReLU(alpha=0.2))
model.add(Dropout(0.25))
model.add(BatchNormalization(momentum=0.8))
model.add(Conv2D(128, kernel_size=3, strides=2, padding='same'))
model.add(LeakyReLU(alpha=0.2))
model.add(Dropout(0.25))
model.add(BatchNormalization(momentum=0.8))
model.add(Conv2D(256, kernel_size=3, strides=1, padding='same'))
model.add(LeakyReLU(alpha=0.2))
model.add(Dropout(0.25))
model.add(Flatten())
model.summary()
img = Input(shape=self.img_shape)
features = model(img) # 经判别器提取的特征
valid = Dense(1, activation='sigmoid')(features) # 分类器输出结果
label = Dense(self.num_classes + 1, activation='softmax')(features) # 判别器结果
return Model(img, [valid, label])
5. 训练模型
def train(self, epochs, batch_size=128, sample_interval=50):
# 加载数据
(X_train, y_train), (_, _) = mnist.load_data()
half_batch = batch_size / 2
# 归一化到-1~~1
X_train = (X_train.astype(np.float32) - 127.5) / 127.5
X_train = np.expand_dims(X_train, axis=3)
y_train = y_train.reshape(-1, 1)
# 分类权重
cw1 = {0: 1, 1: 1}
cw2 = {i: self.num_classes / half_batch for i in range(self.num_classes)}
cw2[self.num_classes] = 1 / half_batch
# 真实值
valid = np.ones((batch_size, 1))
fake = np.zeros((batch_size, 1))
for epoch in range(epochs):
'''训练判别器'''
# 选择图像批度
idx = np.random.randint(0, X_train.shape[0], batch_size)
imgs = X_train[idx]
# 噪声样本,生成新图像
noise = np.random.normal(0, 1, (batch_size, self.latent_dim))
gen_imgs = self.generator.predict(noise)
# 标签的one-hot变量
labels = to_categorical(y_train[idx], num_classes=self.num_classes + 1)
fake_labels = to_categorical(np.full((batch_size, 1), self.num_classes), num_classes=self.num_classes + 1)
# 训练判别器
d_loss_real = self.discriminator.train_on_batch(imgs, [valid, labels], class_weight=[cw1, cw2])
d_loss_fake = self.discriminator.train_on_batch(gen_imgs, [fake, fake_labels], class_weight=[cw1, cw2])
d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
'''训练生成器'''
g_loss = self.combined.train_on_batch(noise, valid, class_weight=[cw1, cw2])
print("%d [D loss: %f, acc: %.2f%%, op_acc: %.2f%%] [G loss: %f]" % (epoch, d_loss[0], 100 * d_loss[3], 100 * d_loss[4], g_loss))
if epoch % sample_interval == 0:
self.sample_images(epoch)
6. 显示图像
def sample_images(self, epoch):
r, c = 5, 5
noise = np.random.normal(0, 1, (r * c, self.latent_dim))
gen_imgs = self.generator.predict(noise)
gen_imgs = 0.5 * gen_imgs + 1
fig, axs = plt.subplots(r, c)
cnt = 0
for i in range(r):
for j in range(c):
axs[i, j].imshow(gen_imgs[cnt, :, :, 0], cmap='gray')
axs[i, j].axis('off')
cnt += 1
fig.savefig("images/mnist_%d.png" % epoch)
plt.close()
7. 运行代码
if __name__ == '__main__':
sgan = SGAN()
sgan.train(epochs=20000, batch_size=32, sample_interval=50)
结果: