CycleGAN的原理可以概述为：
将一类图片转换成另一类图片 。也就是说，现在有两个样
本空间，X和Y，我们希望把X空间中的样本转换成Y空间中
的样本。（获取一个数据集的特征，并转化成另一个数据
集的特征）
这样来看：实际的目标就是学习从X到Y的映射。我们设这
个映射为F。它就对应着GAN中的 生成器 ，F可以将X中的
图片x转换为Y中的图片F(x)。对于生成的图片，我们还需要
GAN中的 判别器 来判别它是否为真实图片，由此构成对抗
生成网络
在足够大的样本容量下，网络可以将相同的输入图像集合
映射到目标域中图像的任何随机排列，其中任何学习的映
射可以归纳出与目标分布匹配的输出分布（即：映射F完全
可以将所有x都映射为Y空间中的同一张图片，使损失无效
化）。
因此，单独的对抗损失Loss不能保证学习函数可以
将单个输入Xi映射到期望的输出Yi。
对此，作者又提出了所谓的“循环一致性损失”
（cycle consistency loss）。
我们希望能够把 domain A 的图片（命名为 a）转
化为 domain B 的图片（命名为图片 b）。
为了实现这个过程，我们需要两个生成器 G_AB 和
G_BA，分别把 domain A 和 domain B 的图片进行
互相转换。
将X的图片转换到Y空间后，应该还可以转换回来。
这样就杜绝模型把所有X的图片都转换为Y空间中的
同一张图片了
最后为了训练这个单向 GAN 需要两个 loss，分别是
生成器的重建 loss 和判别器的判别 loss。
判别 loss：判别器 D_B 是用来判断输入的图片是否
是真实的 domain B 图片
CycleGAN 其实就是一个 A→B 单向 GAN 加上一个
B→A 单向 GAN。两个 GAN 共享两个生成器，然
后各自带一个判别器，所以加起来总共有两个判别器
和两个生成器。
一个单向 GAN 有两个 loss，而 CycleGAN 加起来
总共有四个 loss。
对颜色、纹理等的转换效果比较好，对多样性高的、
多变的转换效果不好（如几何转换）
代码
import tensorflow as tfimport globfrom matplotlib import pyplot as plt%matplotlib inlineAUTOTUNE = tf.data.experimental.AUTOTUNEimport os
os.listdir('../input/apple2orange/apple2orange')
imgs_A = glob.glob('../input/apple2orange/apple2orange/trainA/*.jpg')
imgs_B = glob.glob('../input/apple2orange/apple2orange/trainB/*.jpg')
test_A = glob.glob('../input/apple2orange/apple2orange/testA/*.jpg')test_B = glob.glob('../input/apple2orange/apple2orange/testB/*.jpg')
def read_jpg(path):img = tf.io.read_file(path)img = tf.image.decode_jpeg(img, channels=3)return img
def normalize(input_image):input_image = tf.cast(input_image, tf.float32)/127.5 - 1return input_image
def load_image(image_path):image = read_jpg(image_path)image = tf.image.resize(image, (256, 256))image = normalize(image)return image
train_a = tf.data.Dataset.from_tensor_slices(imgs_A)train_b = tf.data.Dataset.from_tensor_slices(imgs_B)test_a = tf.data.Dataset.from_tensor_slices(test_A)test_b = tf.data.Dataset.from_tensor_slices(test_B)
BUFFER_SIZE = 200
train_a = train_a.map(load_image, num_parallel_calls=AUTOTUNE).cache().shuffle(BUFFER_SIZE).batch(1)train_b = train_b.map(load_image, num_parallel_calls=AUTOTUNE).cache().shuffle(BUFFER_SIZE).batch(1)test_a = test_a.map(load_image, num_parallel_calls=AUTOTUNE).cache().shuffle(BUFFER_SIZE).batch(1)test_b = test_b.map(load_image, num_parallel_calls=AUTOTUNE).cache().shuffle(BUFFER_SIZE).batch(1)
data_train = tf.data.Dataset.zip((train_a, train_b))data_test = tf.data.Dataset.zip((test_a, test_b))
plt.figure(figsize=(6, 3))for img, musk in zip(train_a.take(1), train_b.take(1)):plt.subplot(1,2,1)plt.imshow(tf.keras.preprocessing.image.array_to_img(img[0]))plt.subplot(1,2,2)plt.imshow(tf.keras.preprocessing.image.array_to_img(musk[0]))
实例归一化
!pip install tensorflow_addons
import tensorflow_addons as tfa
OUTPUT_CHANNELS = 3
def downsample(filters, size, apply_batchnorm=True):# initializer = tf.random_normal_initializer(0., 0.02)result = tf.keras.Sequential()result.add(tf.keras.layers.Conv2D(filters, size, strides=2, padding='same',use_bias=False))if apply_batchnorm:result.add(tfa.layers.InstanceNormalization())result.add(tf.keras.layers.LeakyReLU())return result
def upsample(filters, size, apply_dropout=False):# initializer = tf.random_normal_initializer(0., 0.02)result = tf.keras.Sequential()result.add(tf.keras.layers.Conv2DTranspose(filters, size, strides=2,padding='same',use_bias=False))result.add(tfa.layers.InstanceNormalization())if apply_dropout:result.add(tf.keras.layers.Dropout(0.5))result.add(tf.keras.layers.ReLU())return result
def Generator():inputs = tf.keras.layers.Input(shape=[256,256,3])down_stack = [downsample(64, 4, apply_batchnorm=False), # (bs, 128, 128, 64)downsample(128, 4), # (bs, 64, 64, 128)downsample(256, 4), # (bs, 32, 32, 256)downsample(512, 4), # (bs, 16, 16, 512)downsample(512, 4), # (bs, 8, 8, 512)downsample(512, 4), # (bs, 4, 4, 512)downsample(512, 4), # (bs, 2, 2, 512)downsample(512, 4), # (bs, 1, 1, 512)]up_stack = [upsample(512, 4, apply_dropout=True), # (bs, 2, 2, 1024)upsample(512, 4, apply_dropout=True), # (bs, 4, 4, 1024)upsample(512, 4, apply_dropout=True), # (bs, 8, 8, 1024)upsample(512, 4), # (bs, 16, 16, 1024)upsample(256, 4), # (bs, 32, 32, 512)upsample(128, 4), # (bs, 64, 64, 256)upsample(64, 4), # (bs, 128, 128, 128)]# initializer = tf.random_normal_initializer(0., 0.02)last = tf.keras.layers.Conv2DTranspose(OUTPUT_CHANNELS, 4,strides=2,padding='same',activation='tanh') # (bs, 256, 256, 3)x = inputs# Downsampling through the modelskips = []for down in down_stack:x = down(x)skips.append(x)skips = reversed(skips[:-1])# Upsampling and establishing the skip connectionsfor up, skip in zip(up_stack, skips):x = up(x)x = tf.keras.layers.Concatenate()([x, skip])x = last(x)return tf.keras.Model(inputs=inputs, outputs=x)
generator_x = Generator() # a——>ogenerator_y = Generator() # o——>a#tf.keras.utils.plot_model(generator, show_shapes=True, dpi=64)
def Discriminator():# initializer = tf.random_normal_initializer(0., 0.02)inp = tf.keras.layers.Input(shape=[256, 256, 3], name='input_image')down1 = downsample(64, 4, False)(inp) # (bs, 128, 128, 64)down2 = downsample(128, 4)(down1) # (bs, 64, 64, 128)down3 = downsample(256, 4)(down2) # (bs, 32, 32, 256)zero_pad1 = tf.keras.layers.ZeroPadding2D()(down3) # (bs, 34, 34, 256)conv = tf.keras.layers.Conv2D(512, 4, strides=1,use_bias=False)(zero_pad1) # (bs, 31, 31, 512)norm1 = tfa.layers.InstanceNormalization()(conv)leaky_relu = tf.keras.layers.LeakyReLU()(norm1)zero_pad2 = tf.keras.layers.ZeroPadding2D()(leaky_relu) # (bs, 33, 33, 512)last = tf.keras.layers.Conv2D(1, 4, strides=1)(zero_pad2) # (bs, 30, 30, 1)return tf.keras.Model(inputs=inp, outputs=last)
discriminator_x = Discriminator() # discriminator adiscriminator_y = Discriminator() # discriminator o#tf.keras.utils.plot_model(discriminator, show_shapes=True, dpi=64)
loss_object = tf.keras.losses.BinaryCrossentropy(from_logits=True)
def discriminator_loss(disc_real_output, disc_generated_output):real_loss = loss_object(tf.ones_like(disc_real_output), disc_real_output)generated_loss = loss_object(tf.zeros_like(disc_generated_output), disc_generated_output)total_disc_loss = real_loss + generated_lossreturn total_disc_loss
def generator_loss(disc_generated_output):gan_loss = loss_object(tf.ones_like(disc_generated_output), disc_generated_output)return gan_loss
LAMBDA = 7
def calc_cycle_loss(real_image, cycled_image):loss1 = tf.reduce_mean(tf.abs(real_image - cycled_image))return LAMBDA * loss1
generator_x_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)generator_y_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)discriminator_x_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)discriminator_y_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)
def generate_images(model, test_input):prediction = model(test_input, training=True)plt.figure(figsize=(15,15))display_list = [test_input[0], prediction[0]]title = ['Input Image', 'Predicted Image']for i in range(2):plt.subplot(1, 2, i+1)plt.title(title[i])# getting the pixel values between [0, 1] to plot it.plt.imshow(display_list[i] * 0.5 + 0.5)plt.axis('off')plt.show()
@tf.functiondef train_step(image_a, image_b):with tf.GradientTape(persistent=True) as tape:fake_b = generator_x(image_a, training=True)cycled_a = generator_y(fake_b, training=True)fake_a = generator_y(image_b, training=True)cycled_b = generator_x(fake_a, training=True)disc_real_a = discriminator_x(image_a, training=True)disc_real_b = discriminator_y(image_b, training=True)disc_fake_a = discriminator_x(fake_a, training=True)disc_fake_b = discriminator_y(fake_b, training=True)gen_x_loss = generator_loss(disc_fake_b)gen_y_loss = generator_loss(disc_fake_a)total_cycle_loss = (calc_cycle_loss(image_a, cycled_a) + calc_cycle_loss(image_b, cycled_b))# 总生成器损失 = 对抗性损失 + 循环损失。total_gen_x_loss = gen_x_loss + total_cycle_losstotal_gen_y_loss = gen_y_loss + total_cycle_lossdisc_x_loss = discriminator_loss(disc_real_a, disc_fake_a)disc_y_loss = discriminator_loss(disc_real_b, disc_fake_b)# 计算生成器和判别器损失。generator_x_gradients = tape.gradient(total_gen_x_loss, generator_x.trainable_variables)generator_y_gradients = tape.gradient(total_gen_y_loss, generator_y.trainable_variables)discriminator_x_gradients = tape.gradient(disc_x_loss, discriminator_x.trainable_variables)discriminator_y_gradients = tape.gradient(disc_y_loss, discriminator_y.trainable_variables)# 将梯度应用于优化器。generator_x_optimizer.apply_gradients(zip(generator_x_gradients, generator_x.trainable_variables))generator_y_optimizer.apply_gradients(zip(generator_y_gradients, generator_y.trainable_variables))discriminator_x_optimizer.apply_gradients(zip(discriminator_x_gradients,discriminator_x.trainable_variables))discriminator_y_optimizer.apply_gradients(zip(discriminator_y_gradients,discriminator_y.trainable_variables))
def fit(train_ds, test_ds, epochs):for epoch in range(epochs+1):for img_a, img_b in train_ds:train_step(img_a, img_b)print ('.', end='')if epoch % 5 == 0:print()for test_a, test_b in test_ds.take(1):print("Epoch: ", epoch)generate_images(generator_x, test_a)generate_images(generator_x, test_a)
EPOCHS = 100
fit(data_train, data_test, EPOCHS)
