【Tensorflow+Keras】keras实现条件生成对抗网络DCGAN--以Minis和fashion_mnist数据集为例

1 引言

条件生成对抗网络(Conditional Generative Adversarial Nets，简称CGAN)是GAN的改进。

举例如图所示，如果使用Minist数据集

• 在GAN中，在训练时，随机初始化一个和图片大小一致的矩阵和原始图片的矩阵进行博弈，产生一个新的类似于原始图片的网络。
• 在Conditional GAN中，在训练时，会同时输入label，告诉当前网络生成的图片是数字8，而不是生成其他数字的图片

图1 GAN原理图

图2 Conditional GAN原理图

2 实现

Github源码

Mian.py

指定条件即条件输入是Label

import tensorflow as tf
from tensorflow.keras.datasets import fashion_mnist,mnist
import utils
from models import build_discriminator_model,build_generator_model
import numpy as np

# 图片维度
noise_dim = 100
# 学习率
learning_rate = 1e-4
# 交叉熵用来计算生成器Generator和鉴别器Disctiminator的损失函数
cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)
# 指定使用哪个数据集
dataset = 'fashion_mnist'
if dataset == 'mnist':
    (X_train, y_train), (X_test, y_test) = mnist.load_data()
if dataset == 'fashion_mnist':
    (X_train, y_train), (X_test, y_test) = fashion_mnist.load_data()
else:
    raise RuntimeError('Dataset not found')
# 数据标准化
X_train, X_test = utils.normalize(X_train, X_test)
# 初始化G和D
discriminator = build_discriminator_model()
generator = build_generator_model()
# 数据标准化
def normalize(train, test):
  # convert from integers to floats
  train_norm = train.astype('float32')
  test_norm = test.astype('float32')
  # normalize to range 0-1
  train_norm = train_norm / 255.0
  test_norm = test_norm / 255.0
  # return normalized images
  return train_norm, test_norm

# 生成器和鉴别器的优化器
generator_optimizer = tf.keras.optimizers.Adam(learning_rate = 1e-4)
discriminator_optimizer = tf.keras.optimizers.Adam(learning_rate = 1e-4)

# 鉴别器的损失函数
def discriminator_loss(real_output, fake_output):
    real_loss = cross_entropy(tf.ones_like(real_output), real_output)
    fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)
    total_loss = real_loss + fake_loss
    return total_loss
# 生成器的损失函数
def generator_loss(fake_output):
    return cross_entropy(tf.ones_like(fake_output), fake_output)
# 保存模型
def save_models(epochs, learning_rate):
    generator.save(f'generator-epochs-{epochs}-learning_rate-{learning_rate}.h5')
    discriminator.save(f'discriminator-epochs-{epochs}-learning_rate-{learning_rate}.h5')

# 训练
tf.function
def train_step(batch_size=512):
    # 随机产生一组下标，从训练数据中随机抽取训练集
    idx = np.random.randint(0, X_train.shape[0], batch_size)
    # 随机抽取训练集
    Xtrain, labels = X_train[idx], y_train[idx]
    with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:
        # 随机初始化一个和图片大小的矩阵
        z = np.random.normal(0, 1, size=(batch_size, noise_dim))
        # 经过生成器，产生一个图片。并指定条件是label，把label嵌入到图片中
        generated_images = generator([z, labels], training=True)
        real_output = discriminator([Xtrain, labels], training=True)
        fake_output = discriminator([generated_images, labels], training=True)
        gen_loss = generator_loss(fake_output)
        disc_loss = discriminator_loss(real_output, fake_output)
    # 打印G和D的损失函数
    tf.print(f'Genrator loss: {gen_loss} Discriminator loss: {disc_loss}')
    gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables)
    gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables)
    # 更新梯度
    generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables))
    discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables))
if __name__  =="__main__":
    epochs = 100
    for epoch in range(1, epochs + 1):
        print(f'Epoch {epoch}/{epochs}')
        train_step()
        if epoch % 500 == 0:
            save_models(epoch, learning_rate)

Model.py

模型采用深度卷卷积的GAN网络结构

import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras.models import Sequential, Model
import numpy as np

WIDTH, HEIGHT = 28, 28
num_classes = 10
img_channel = 1
img_shape = (WIDTH, HEIGHT, img_channel)
noise_dim = 100

def build_generator_model():
    model = tf.keras.Sequential()

    model.add(layers.Dense(7*7*256, use_bias=False,input_shape=(noise_dim,)))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Reshape((7, 7, 256)))

    model.add(layers.Conv2DTranspose(128, (1, 1), strides=(1, 1), padding='same', use_bias=False))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Conv2DTranspose(64, (3, 3), strides=(2, 2), padding='same', use_bias=False))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Conv2DTranspose(1, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh'))

    z = layers.Input(shape=(noise_dim,))
    label = layers.Input(shape=(1,))

    label_embedding = layers.Embedding(num_classes, noise_dim, input_length = 1)(label)
    label_embedding = layers.Flatten()(label_embedding)
    joined = layers.multiply([z, label_embedding])

    img = model(joined)
    return Model([z, label], img)

def build_discriminator_model():
    model = tf.keras.Sequential()
    model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same',
                                      input_shape=[28, 28, 2]))
    model.add(layers.LeakyReLU())
    model.add(layers.Dropout(0.3))

    model.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same'))
    model.add(layers.LeakyReLU())
    model.add(layers.Dropout(0.3))

    model.add(layers.Flatten())
    model.add(layers.Dense(1))

    img = layers.Input(shape=(img_shape))
    label = layers.Input(shape=(1,))

    label_embedding = layers.Embedding(input_dim=num_classes, output_dim=np.prod(img_shape), input_length = 1)(label)
    label_embedding = layers.Flatten()(label_embedding)
    label_embedding = layers.Reshape(img_shape)(label_embedding)

    concat = layers.Concatenate(axis=-1)([img, label_embedding])
    prediction = model(concat)
    return Model([img, label], prediction)