发现一个挺有意思的结构，可以通过同一个输入，生成不同的内容。
COGAN是一种耦合生成式对抗网络，其内部具有一定的耦合，可以对同一个输入有不同的输出。
其具体实现方式就是：
1、建立两个生成模型，两个判别模型。
2、两个生成模型的特征提取部分有一定的重合，在最后生成图片的部分分开，以生成不同类型的图片。
3、两个判别模型的特征提取部分有一定的重合，在最后判别真伪的部分分开，以判别不同类型的图片。
生成模型的输入是一个N维度的符合正太分布的随机数，输出是一个28,28,1的Mnist手写体。
一共存在两个生成模型，两个生成模型的特征提取部分有一定的重合，在最后生成图片的部分分开，以生成不同类型的图片。
即：权值部分有一定的共享。
def build_generators(self):
# 共享权值部分
noise = Input(shape=(self.latent_dim,))
x = Dense(32 * 7 * 7, activation="relu", input_dim=self.latent_dim)(noise)
x = Reshape((7, 7, 32))(x)
x = Conv2D(64, kernel_size=3, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
x = Activation("relu")(x)
x = UpSampling2D()(x)
x = Conv2D(128, kernel_size=3, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
x = Activation("relu")(x)
x = UpSampling2D()(x)
x = Conv2D(128, kernel_size=3, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
feature_repr = Activation("relu")(x)
model = Model(noise,feature_repr)
noise = Input(shape=(self.latent_dim,))
feature_repr = model(noise)
# 生成模型1
g1 = Conv2D(64, kernel_size=1, padding="same")(feature_repr)
g1 = BatchNormalization(momentum=0.8)(g1)
g1 = Activation("relu")(g1)
g1 = Conv2D(64, kernel_size=3, padding="same")(g1)
g1 = BatchNormalization(momentum=0.8)(g1)
g1 = Activation("relu")(g1)
g1 = Conv2D(64, kernel_size=1, padding="same")(g1)
g1 = BatchNormalization(momentum=0.8)(g1)
g1 = Activation("relu")(g1)
g1 = Conv2D(self.channels, kernel_size=1, padding="same")(g1)
img1 = Activation("tanh")(g1)
# 生成模型2
g2 = Conv2D(64, kernel_size=1, padding="same")(feature_repr)
g2 = BatchNormalization(momentum=0.8)(g2)
g2 = Activation("relu")(g2)
g2 = Conv2D(64, kernel_size=3, padding="same")(g2)
g2 = BatchNormalization(momentum=0.8)(g2)
g2 = Activation("relu")(g2)
g2 = Conv2D(64, kernel_size=1, padding="same")(g2)
g2 = BatchNormalization(momentum=0.8)(g2)
g2 = Activation("relu")(g2)
g2 = Conv2D(self.channels, kernel_size=1, padding="same")(g2)
img2 = Activation("tanh")(g2)
return Model(noise, img1), Model(noise, img2)
判别模型的输入一个28,28,1维的图片，输出一个是0到1之间的数，1代表判断这个图片是真的，0代表判断这个图片是假的。
一共存在两个判别模型，两个生成模型的特征提取部分有一定的重合，在最后判别真伪的部分分开，以判别不同类型的图片。
def build_discriminators(self):
# 共享权值部分
img = Input(shape=self.img_shape)
x = Conv2D(64, kernel_size=3, strides=2, padding="same")(img)
x = BatchNormalization(momentum=0.8)(x)
x = Activation("relu")(x)
x = Conv2D(128, kernel_size=3, strides=2, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
x = Activation("relu")(x)
x = Conv2D(64, kernel_size=3, strides=2, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
x = GlobalAveragePooling2D()(x)
feature_repr = Activation("relu")(x)
model = Model(img,feature_repr)
img1 = Input(shape=self.img_shape)
img2 = Input(shape=self.img_shape)
img1_embedding = model(img1)
img2_embedding = model(img2)
# 生成评价模型1
validity1 = Dense(1, activation='sigmoid')(img1_embedding)
# 生成评价模型2
validity2 = Dense(1, activation='sigmoid')(img2_embedding)
return Model(img1, validity1), Model(img2, validity2)
COGAN的训练思路分为如下几个步骤：
1、创建两个风格不同的数据集。
2、随机生成batch_size个N维向量，利用两个不同的生成模型生成图片。
3、利用两个判别模型分别对两个不同的生成模型的生成图片进行判别、对两个风格不同的数据集进行随机选取并进行判别。
4、根据两个判别模型的结果与1对比，对两个生成模型进行训练。
from __future__ import print_function, division
import scipy
from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout
from keras.layers import BatchNormalization, Activation, ZeroPadding2D, GlobalAveragePooling2D
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
import os
import matplotlib.pyplot as plt
import numpy as np
class COGAN():
def __init__(self):
# 输入shape
self.img_rows = 28
self.img_cols = 28
self.channels = 1
self.img_shape = (self.img_rows, self.img_cols, self.channels)
# 分十类
self.num_classes = 10
self.latent_dim = 100
# adam优化器
optimizer = Adam(0.0002, 0.5)
# 生成两个判别器
self.d1, self.d2 = self.build_discriminators()
self.d1.compile(loss='binary_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
self.d2.compile(loss='binary_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
# 建立生成器
self.g1, self.g2 = self.build_generators()
z = Input(shape=(self.latent_dim,))
img1 = self.g1(z)
img2 = self.g2(z)
self.d1.trainable = False
self.d2.trainable = False
valid1 = self.d1(img1)
valid2 = self.d2(img2)
self.combined = Model(z, [valid1, valid2])
self.combined.compile(loss=['binary_crossentropy', 'binary_crossentropy'],
optimizer=optimizer)
def build_generators(self):
# 共享权值部分
noise = Input(shape=(self.latent_dim,))
x = Dense(32 * 7 * 7, activation="relu", input_dim=self.latent_dim)(noise)
x = Reshape((7, 7, 32))(x)
x = Conv2D(64, kernel_size=3, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
x = Activation("relu")(x)
x = UpSampling2D()(x)
x = Conv2D(128, kernel_size=3, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
x = Activation("relu")(x)
x = UpSampling2D()(x)
x = Conv2D(128, kernel_size=3, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
feature_repr = Activation("relu")(x)
model = Model(noise,feature_repr)
noise = Input(shape=(self.latent_dim,))
feature_repr = model(noise)
# 生成模型1
g1 = Conv2D(64, kernel_size=1, padding="same")(feature_repr)
g1 = BatchNormalization(momentum=0.8)(g1)
g1 = Activation("relu")(g1)
g1 = Conv2D(64, kernel_size=3, padding="same")(g1)
g1 = BatchNormalization(momentum=0.8)(g1)
g1 = Activation("relu")(g1)
g1 = Conv2D(64, kernel_size=1, padding="same")(g1)
g1 = BatchNormalization(momentum=0.8)(g1)
g1 = Activation("relu")(g1)
g1 = Conv2D(self.channels, kernel_size=1, padding="same")(g1)
img1 = Activation("tanh")(g1)
# 生成模型2
g2 = Conv2D(64, kernel_size=1, padding="same")(feature_repr)
g2 = BatchNormalization(momentum=0.8)(g2)
g2 = Activation("relu")(g2)
g2 = Conv2D(64, kernel_size=3, padding="same")(g2)
g2 = BatchNormalization(momentum=0.8)(g2)
g2 = Activation("relu")(g2)
g2 = Conv2D(64, kernel_size=1, padding="same")(g2)
g2 = BatchNormalization(momentum=0.8)(g2)
g2 = Activation("relu")(g2)
g2 = Conv2D(self.channels, kernel_size=1, padding="same")(g2)
img2 = Activation("tanh")(g2)
return Model(noise, img1), Model(noise, img2)
def build_discriminators(self):
# 共享权值部分
img = Input(shape=self.img_shape)
x = Conv2D(64, kernel_size=3, strides=2, padding="same")(img)
x = BatchNormalization(momentum=0.8)(x)
x = Activation("relu")(x)
x = Conv2D(128, kernel_size=3, strides=2, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
x = Activation("relu")(x)
x = Conv2D(64, kernel_size=3, strides=2, padding="same")(x)
x = BatchNormalization(momentum=0.8)(x)
x = GlobalAveragePooling2D()(x)
feature_repr = Activation("relu")(x)
model = Model(img,feature_repr)
img1 = Input(shape=self.img_shape)
img2 = Input(shape=self.img_shape)
img1_embedding = model(img1)
img2_embedding = model(img2)
# 生成评价模型1
validity1 = Dense(1, activation='sigmoid')(img1_embedding)
# 生成评价模型2
validity2 = Dense(1, activation='sigmoid')(img2_embedding)
return Model(img1, validity1), Model(img2, validity2)
def train(self, epochs, batch_size=128, sample_interval=50):
(X_train, _), (_, _) = mnist.load_data()
X_train = (X_train.astype(np.float32) - 127.5) / 127.5
X_train = np.expand_dims(X_train, axis=3)
X1 = X_train[:int(X_train.shape[0]/2)]
X2 = X_train[int(X_train.shape[0]/2):]
X2 = scipy.ndimage.interpolation.rotate(X2, 90, axes=(1, 2))
valid = np.ones((batch_size, 1))
fake = np.zeros((batch_size, 1))
for epoch in range(epochs):
# ---------------------- #
# 训练评价者
# ---------------------- #
idx = np.random.randint(0, X1.shape[0], batch_size)
imgs1 = X1[idx]
imgs2 = X2[idx]
noise = np.random.normal(0, 1, (batch_size, 100))
gen_imgs1 = self.g1.predict(noise)
gen_imgs2 = self.g2.predict(noise)
d1_loss_real = self.d1.train_on_batch(imgs1, valid)
d2_loss_real = self.d2.train_on_batch(imgs2, valid)
d1_loss_fake = self.d1.train_on_batch(gen_imgs1, fake)
d2_loss_fake = self.d2.train_on_batch(gen_imgs2, fake)
d1_loss = 0.5 * np.add(d1_loss_real, d1_loss_fake)
d2_loss = 0.5 * np.add(d2_loss_real, d2_loss_fake)
# ------------------ #
# 训练生成模型
# ------------------ #
g_loss = self.combined.train_on_batch(noise, [valid, valid])
print ("%d [D1 loss: %f, acc.: %.2f%%] [D2 loss: %f, acc.: %.2f%%] [G loss: %f]" \
% (epoch, d1_loss[0], 100*d1_loss[1], d2_loss[0], 100*d2_loss[1], g_loss[0]))
if epoch % sample_interval == 0:
self.sample_images(epoch)
def sample_images(self, epoch):
r, c = 4, 4
noise = np.random.normal(0, 1, (r * int(c/2), 100))
gen_imgs1 = self.g1.predict(noise)
gen_imgs2 = self.g2.predict(noise)
gen_imgs = np.concatenate([gen_imgs1, gen_imgs2])
gen_imgs = 0.5 * gen_imgs + 0.5
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()
if __name__ == '__main__':
if not os.path.exists("./images"):
os.makedirs("./images")
gan = COGAN()
gan.train(epochs=30000, batch_size=256, sample_interval=200)
实现效果为：
