原文：TensorFlow - 深度卷积生成对抗网络

该文主要是介绍如何使用深度卷积生成对抗网络(DCGAN)生成手写数字图片.

代码是使用 Keras Sequential API 与 tf.GradientTape 训练循环编写的.

1. 什么是生成对抗网络

生成对抗网络(GANs)是当今计算机科学领域最有趣的想法之一. 两个模型通过对抗过程同时训练. 一个生成器(“艺术家”)学习创造看起来真实的图像，而判别器(“艺术评论家”)学习区分真假图像.

训练过程中，生成器在生成逼真图像方面逐渐变强，而判别器在辨别这些图像的能力上逐渐变强. 当判别器不再能够区分真实图片和伪造图片时，训练过程达到平衡.

本文在 MNIST 数据集上演示了该过程. 下方动画展示了当训练了 50 个epoch (全部数据集迭代50次) 时生成器所生成的一系列图片. 图片从随机噪声开始，随着时间的推移越来越像手写数字.

关于 GANs 更多信息：参阅 MIT的 深度学习入门 课程.

2. 配置

首先安装用于生成 GIF 图片的库 - imageio:

pip install -q imageio

import glob
import imageio
import matplotlib.pyplot as plt
import numpy as np
import os
import PIL
import time

import tensorflow as tf
print(tf.__version__)
#2.3.0
from tensorflow.keras import layers

3. 加载和准备数据集

# MNIST 数据集
(train_images, train_labels), (_, _) = tf.keras.datasets.mnist.load_data()

train_images = train_images.reshape(train_images.shape[0], 28, 28, 1).astype('float32')
train_images = (train_images - 127.5) / 127.5 # 将图片标准化到 [-1, 1] 区间内

BUFFER_SIZE = 60000
BATCH_SIZE = 256

# 批量化和打乱数据
train_dataset = tf.data.Dataset.from_tensor_slices(train_images).shuffle(BUFFER_SIZE).batch(BATCH_SIZE)

4. 创建模型

生成器和判别器均使用 Keras Sequential API 定义.

4.1. 生成器

生成器使用 tf.keras.layers.Conv2DTranspose (上采样)层来从种子(随机噪声)中产生图片. 以一个使用该种子作为输入的 Dense 层开始，然后多次上采样直到达到所期望的 28x28x1 的图片尺寸.

注：除了输出层使用 tanh 之外，其他每层均使用 tf.keras.layers.LeakyReLU 作为激活函数.

def make_generator_model():
    model = tf.keras.Sequential()
    model.add(layers.Dense(7*7*256, use_bias=False, input_shape=(100,)))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Reshape((7, 7, 256)))
    assert model.output_shape == (None, 7, 7, 256) # 注意：batch size 没有限制

    model.add(layers.Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False))
    assert model.output_shape == (None, 7, 7, 128)
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False))
    assert model.output_shape == (None, 14, 14, 64)
    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'))
    assert model.output_shape == (None, 28, 28, 1)

    return model

使用(尚未训练的)生成器创建一张图片.

generator = make_generator_model()

noise = tf.random.normal([1, 100])
generated_image = generator(noise, training=False)

plt.imshow(generated_image[0, :, :, 0], cmap='gray')
plt.show()

如图：

4.2. 判别器

判别器是一个基于 CNN 的图片分类器.

def make_discriminator_model():
    model = tf.keras.Sequential()
    model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same',
                                     input_shape=[28, 28, 1]))
    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))

    return model

使用(尚未训练的)判别器来对图片的真伪进行判断. 模型将被训练为为真实图片输出正值，为伪造图片输出负值.

discriminator = make_discriminator_model()
decision = discriminator(generated_image)
print (decision)
#tf.Tensor([[-0.00427552]], shape=(1, 1), dtype=float32)

5. 定义损失函数和优化器

为两个模型定义损失函数和优化器.

# 计算交叉熵损失的辅助函数
cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)

5.1. 判别器损失

判别器损失量化判别器从判断真伪图片的能力. 它将判别器对真实图片的预测值与值全为 1 的数组进行对比，将判别器对伪造(生成的)图片的预测值与值全为 0 的数组进行对比.

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

5.2. 生成器损失

生成器损失量化其欺骗判别器的能力. 直观来讲，如果生成器表现良好，判别器将会把伪造图片判断为真实图片(或 1). 这里将把判别器在生成图片上的判断结果与一个值全为 1 的数组进行对比.

def generator_loss(fake_output):
    return cross_entropy(tf.ones_like(fake_output), fake_output)

由于需要分别训练两个网络，判别器和生成器的优化器是不同的.

generator_optimizer = tf.keras.optimizers.Adam(1e-4)
discriminator_optimizer = tf.keras.optimizers.Adam(1e-4)

5.3. 保存checkpoint

保存和恢复模型，在长时间训练任务被中断的情况下比较有帮助.

checkpoint_dir = './training_checkpoints'
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
checkpoint = tf.train.Checkpoint(generator_optimizer=generator_optimizer,
                                 discriminator_optimizer=discriminator_optimizer,
                                 generator=generator,
                                 discriminator=discriminator)

6. 定义训练循环

EPOCHS = 50
noise_dim = 100
num_examples_to_generate = 16

# 将重复使用该种子(因此在动画 GIF 中更容易可视化进度)
seed = tf.random.normal([num_examples_to_generate, noise_dim])

训练循环在生成器接收到一个随机种子作为输入时开始. 该种子用于生产一张图片. 判别器随后被用于区分真实图片(选自训练集)和伪造图片(由生成器生成). 针对这里的每一个模型都计算损失函数，并且计算梯度用于更新生成器与判别器.

# 注意 `tf.function` 的使用
# 该注解使函数被“编译”
@tf.function
def train_step(images):
    noise = tf.random.normal([BATCH_SIZE, noise_dim])

    with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:
      generated_images = generator(noise, training=True)

      real_output = discriminator(images, training=True)
      fake_output = discriminator(generated_images, training=True)

      gen_loss = generator_loss(fake_output)
      disc_loss = discriminator_loss(real_output, fake_output)

    gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables)
    gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables)

    generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables))
    discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables))

#
def generate_and_save_images(model, epoch, test_input):
  # 注意 training 设定为 False
  # 因此，所有层都在推理模式下运行(batchnorm). 
  predictions = model(test_input, training=False)

  fig = plt.figure(figsize=(4,4))

  for i in range(predictions.shape[0]):
      plt.subplot(4, 4, i+1)
      plt.imshow(predictions[i, :, :, 0] * 127.5 + 127.5, cmap='gray')
      plt.axis('off')

  plt.savefig('image_at_epoch_{:04d}.png'.format(epoch))
  plt.show()
    
#
def train(dataset, epochs):
  for epoch in range(epochs):
    start = time.time()

    for image_batch in dataset:
      train_step(image_batch)

    # 继续进行时为 GIF 生成图像
    generate_and_save_images(generator, epoch + 1, seed)

    # 每 15 个 epoch 保存一次模型
    if (epoch + 1) % 15 == 0:
      checkpoint.save(file_prefix = checkpoint_prefix)

    print ('Time for epoch {} is {} sec'.format(epoch + 1, time.time()-start))

  # 最后一个 epoch 结束后生成图片
  generate_and_save_images(generator, epochs, seed)

7. 训练模型

调用上面定义的 train() 方法来同时训练生成器和判别器.

注：训练 GANs 可能是棘手的. 重要的是，生成器和判别器不能够互相压制对方(例如，他们以相似的学习率训练).

在训练之初，生成的图片看起来像是随机噪声. 随着训练过程的进行，生成的数字将越来越真实. 在大概 50 个 epoch 之后，这些图片看起来像是 MNIST 数字. 使用 Colab 中的默认设置可能需要大约 1 分钟每 epoch.

train(train_dataset, EPOCHS)

恢复最新的checkpoint.

checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir))

9. 完整实现

#!/usr/bin/python3
#!--*-- coding: utf-8 --*--
import glob
import imageio
import matplotlib.pyplot as plt
import numpy as np
import os
import PIL
import time

import tensorflow as tf
print(tf.__version__)
#2.3.0
from tensorflow.keras import layers


# 生成网络
def make_generator_model():
    model = tf.keras.Sequential()
    model.add(layers.Dense(7 * 7 * 256, use_bias=False, input_shape=(100,)))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Reshape((7, 7, 256)))
    assert model.output_shape == (None, 7, 7, 256)  # 注意：batch size 没有限制

    model.add(layers.Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False))
    assert model.output_shape == (None, 7, 7, 128)
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False))
    assert model.output_shape == (None, 14, 14, 64)
    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'))
    assert model.output_shape == (None, 28, 28, 1)

    return model


# 判别网络
def make_discriminator_model():
    model = tf.keras.Sequential()
    model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same',
                            input_shape=[28, 28, 1]))
    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))

    return model

#
# 计算交叉熵损失的辅助函数
cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)

# 生成器损失函数
def generator_loss(fake_output):
  return cross_entropy(tf.ones_like(fake_output), fake_output)

# 判别器损失函数
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

#
def generate_and_save_images(model, epoch, test_input):
  # 注意 training 设定为 False
  # 因此，所有层都在推理模式下运行(batchnorm). 
  predictions = model(test_input, training=False)

  plt.figure(figsize=(4, 4))
  for i in range(predictions.shape[0]):
    plt.subplot(4, 4, i + 1)
    plt.imshow(predictions[i, :, :, 0] * 127.5 + 127.5, cmap='gray')
    plt.axis('off')

  plt.savefig('image_at_epoch_{:04d}.png'.format(epoch))
  plt.show()

#
@tf.function
def train_step(images, generator, discriminator, generator_optimizer, discriminator_optimizer, batch_size, noise_dim):
  noise = tf.random.normal([batch_size, noise_dim])

  with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:
    generated_images = generator(noise, training=True)

    real_output = discriminator(images, training=True)
    fake_output = discriminator(generated_images, training=True)

    gen_loss = generator_loss(fake_output)
    disc_loss = discriminator_loss(real_output, fake_output)

  gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables)
  gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables)

  generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables))
  discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables))

#
def main():
  # MNIST 数据集
  (train_images, train_labels), (_, _) = tf.keras.datasets.mnist.load_data()

  train_images = train_images.reshape(train_images.shape[0], 28, 28, 1).astype('float32')
  train_images = (train_images - 127.5) / 127.5 # 将图片标准化到 [-1, 1] 区间内

  BUFFER_SIZE = 60000
  BATCH_SIZE = 256

  # 批量化和打乱数据
  train_dataset = tf.data.Dataset.from_tensor_slices(train_images).shuffle(BUFFER_SIZE).batch(BATCH_SIZE)

  #
  generator = make_generator_model()

  noise = tf.random.normal([1, 100])
  generated_image = generator(noise, training=False)
  #
  # plt.imshow(generated_image[0, :, :, 0], cmap='gray')
  # plt.show()

  #
  discriminator = make_discriminator_model()
  # decision = discriminator(generated_image)
  # print(decision)

  #
  generator_optimizer = tf.keras.optimizers.Adam(1e-4)
  discriminator_optimizer = tf.keras.optimizers.Adam(1e-4)

  #
  checkpoint_dir = './training_checkpoints'
  checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
  checkpoint = tf.train.Checkpoint(
      generator_optimizer=generator_optimizer,
      discriminator_optimizer=discriminator_optimizer,
      generator=generator,
      discriminator=discriminator)

  #
  EPOCHS = 50
  noise_dim = 100
  num_examples_to_generate = 16

  # 将重复使用该种子
  seed = tf.random.normal([num_examples_to_generate, noise_dim])

  # train
  for epoch in range(EPOCHS):
      start = time.time()
      for image_batch in train_dataset:
          train_step(image_batch, generator, discriminator, generator_optimizer, discriminator_optimizer, BATCH_SIZE, noise_dim)

      # 继续进行时为 GIF 生成图像
      generate_and_save_images(generator, epoch + 1, seed)

      # 每 15 个 epoch 保存一次模型
      if (epoch + 1) % 15 == 0:
          checkpoint.save(file_prefix=checkpoint_prefix)

      print('Time for epoch {} is {} sec'.format(epoch + 1, time.time() - start))

  # 最后一个 epoch 结束后生成图片
  generate_and_save_images(generator, epochs, seed)
    
  #checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir))
  print('[INFO]Done.')

#
if __name__ == '__main__':
    main()

9. 创建 GIF

# 使用 epoch 数生成单张图片
def display_image(epoch_no):
  return PIL.Image.open('image_at_epoch_{:04d}.png'.format(epoch_no))
#
display_image(EPOCHS)

使用训练过程中生成的图片通过 imageio 生成动态 gif:

anim_file = 'dcgan.gif'

with imageio.get_writer(anim_file, mode='I') as writer:
  filenames = glob.glob('image*.png')
  filenames = sorted(filenames)
  last = -1
  for i,filename in enumerate(filenames):
    frame = 2*(i**0.5)
    if round(frame) > round(last):
      last = frame
    else:
      continue
    image = imageio.imread(filename)
    writer.append_data(image)
  image = imageio.imread(filename)
  writer.append_data(image)

10. 相关材料

[1] - 大规模名人面部属性(CelebA)数据集 - Kaggle

[2] - NIPS 2016 教程： 生成对抗网络