GAN生成对抗网络-WGAN原理与基本实现-wasserstein GAN-07

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import matplotlib.pyplot as plt
%matplotlib inline
import numpy as np
import glob
import os

os.listdir('../input/anime-faces/data/data')[:10]

image_path = glob.glob('../input/anime-faces/data/data/*.png')

def load_preprosess_image(path):
    image = tf.io.read_file(path)
    image = tf.image.decode_png(image, channels=3)
#    image = tf.image.resize_with_crop_or_pad(image, 256, 256)
    image = tf.cast(image, tf.float32)
    image = (image / 127.5) - 1
    return image

image_ds = tf.data.Dataset.from_tensor_slices(image_path)

AUTOTUNE = tf.data.experimental.AUTOTUNE

image_ds = image_ds.map(load_preprosess_image, num_parallel_calls=AUTOTUNE)

BATCH_SIZE = 256
image_count = len(image_path)

image_ds = image_ds.shuffle(image_count).batch(BATCH_SIZE)
image_ds = image_ds.prefetch(AUTOTUNE)

def generator_model():
    model = tf.keras.Sequential()
    model.add(layers.Dense(8*8*256, use_bias=False, input_shape=(100,)))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Reshape((8, 8, 256)))   #8*8*256

    model.add(layers.Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())             #8*8*128

    model.add(layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())             #16*16*128
    
    model.add(layers.Conv2DTranspose(32, (5, 5), strides=(2, 2), padding='same', use_bias=False))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())             #32*32*32
    

    model.add(layers.Conv2DTranspose(3, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh'))
                                              #64*64*3
    
    return model

generator = generator_model()

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

#plt.imshow((generated_image[0, :, :, :3] + 1)/2)

def discriminator_model():
    model = tf.keras.Sequential()
    model.add(layers.Conv2D(32, (5, 5), strides=(2, 2), padding='same',
                                     input_shape=[64, 64, 3]))
    model.add(layers.LeakyReLU())
    model.add(layers.Dropout(0.3))      # 32*32*32

    model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same'))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())
    model.add(layers.Dropout(0.3))      # 16*16*64
    
    model.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same'))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())
#    model.add(layers.Dropout(0.3))      # 8*8*128
    
    model.add(layers.Conv2D(256, (5, 5), strides=(2, 2), padding='same'))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())       # 4*4*256

    model.add(layers.GlobalAveragePooling2D())
    
    model.add(layers.Dense(1024))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())
    model.add(layers.Dense(1))

    return model

discriminator = discriminator_model()
decision = discriminator(generated_image)
print(decision)

def discriminator_loss(real_output, fake_output):
    return tf.reduce_mean(fake_output) - tf.reduce_mean(real_output)

cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)

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

generator_optimizer = tf.keras.optimizers.RMSprop(1e-4)
discriminator_optimizer = tf.keras.optimizers.RMSprop(5e-5)

EPOCHS = 800
noise_dim = 100
num_examples_to_generate = 4

seed = tf.random.normal([num_examples_to_generate, noise_dim])

@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)
    
    for idx, grad in enumerate(gradients_of_discriminator):
        gradients_of_discriminator[idx] = tf.clip_by_value(grad, -0.01, 0.01)

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

def generate_and_save_images(model, epoch, test_input):
  # Notice `training` is set to False.
  # This is so all layers run in inference mode (batchnorm).
    predictions = model(test_input, training=False)

    fig = plt.figure(figsize=(6, 6))

    for i in range(predictions.shape[0]):
        plt.subplot(2, 2, i+1)
        plt.imshow((predictions[i, :, :, :] + 1)/2)
        plt.axis('off')

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

epoch_loss_avg_gen = tf.keras.metrics.Mean('g_loss')
epoch_loss_avg_disc = tf.keras.metrics.Mean('d_loss')

g_loss_results = []
d_loss_results = []

def train(dataset, epochs):
    for epoch in range(epochs):
        for image_batch in dataset:
            g_loss, d_loss = train_step(image_batch)
            epoch_loss_avg_gen(g_loss)
            epoch_loss_avg_disc(d_loss)
            
        g_loss_results.append(epoch_loss_avg_gen.result())
        d_loss_results.append(epoch_loss_avg_disc.result())
        
        epoch_loss_avg_gen.reset_states()
        epoch_loss_avg_disc.reset_states()
#            print('.', end='')
#        print()
        if epoch%10 == 0:
#            print('epoch:', epoch, 'g_loss:', g_loss, 'd_loss:', d_loss)
            generate_and_save_images(generator,
                                     epoch + 1,
                                     seed)


    generate_and_save_images(generator,
                           epochs,
                           seed)

train(image_ds, EPOCHS)

plt.plot(range(1, len(g_loss_results)+1), g_loss_results, label='g_loss')
plt.plot(range(1, len(d_loss_results)+1), d_loss_results, label='d_loss')
plt.legend()

代码二

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

%matplotlib inline

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

real_img = tf.placeholder(tf.float32, [None, 784], name='real_img')
noise_img = tf.placeholder(tf.float32, [None, 100], name='noise_img')

def generator(noise_img, hidden_units, out_dim, reuse=False):
    with tf.variable_scope("generator", reuse=reuse):
        hidden1 = tf.layers.dense(noise_img, hidden_units)
        hidden1 = tf.nn.relu(hidden1)

        # logits & outputs
        logits = tf.layers.dense(hidden1, out_dim)
        outputs = tf.nn.sigmoid(logits)
        
        return logits, outputs

def discriminator(img, hidden_units, reuse=False):
    with tf.variable_scope("discriminator", reuse=reuse):
        # hidden layer
        hidden1 = tf.layers.dense(img, hidden_units)
        hidden1 = tf.nn.relu(hidden1)
        
        # logits & outputs
        outputs = tf.layers.dense(hidden1, 1)
        
        return outputs

def plot_images(samples):
    fig, axes = plt.subplots(nrows=1, ncols=25, sharex=True, sharey=True, figsize=(50,2))
    for img, ax in zip(samples, axes):
        ax.imshow(img.reshape((28, 28)), cmap='Greys_r')
        ax.get_xaxis().set_visible(False)
        ax.get_yaxis().set_visible(False)
    fig.tight_layout(pad=0)

img_size = 784
noise_size = 100
hidden_units = 128
learning_rate = 0.0001

g_logits, g_outputs = generator(noise_img, hidden_units, img_size)

# discriminator
d_real = discriminator(real_img, hidden_units)
d_fake = discriminator(g_outputs, hidden_units, reuse=True)

d_loss = tf.reduce_mean(d_real) - tf.reduce_mean(d_fake)
g_loss = -tf.reduce_mean(d_fake)

train_vars = tf.trainable_variables()

g_vars = [var for var in train_vars if var.name.startswith("generator")]

d_vars = [var for var in train_vars if var.name.startswith("discriminator")]


d_train_opt = tf.train.RMSPropOptimizer(learning_rate).minimize(-d_loss, var_list=d_vars)
g_train_opt = tf.train.RMSPropOptimizer(learning_rate).minimize(g_loss, var_list=g_vars)


clip_d = [p.assign(tf.clip_by_value(p, -0.01, 0.01)) for p in d_vars]         # 对discriminator变量使用clip

batch_size = 32

n_sample = 25

samples = []

losses = []

saver = tf.train.Saver(var_list = g_vars)

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    
    for it in range(1000000):
        for _ in range(5):
            batch_images, _ = mnist.train.next_batch(batch_size)
            
            batch_noise = np.random.uniform(-1, 1, size=(batch_size, noise_size))
            _ = sess.run([d_train_opt, clip_d], feed_dict={real_img: batch_images, noise_img: batch_noise})
            
        _ = sess.run(g_train_opt, feed_dict={noise_img: batch_noise})   

        if it%10000 == 0:
            sample_noise = np.random.uniform(-1, 1, size=(n_sample, noise_size))
            _, samples = sess.run(generator(noise_img, hidden_units, img_size, reuse=True),
                                   feed_dict={noise_img: sample_noise})
            plot_images(samples)
            
        
        # 每一轮结束计算loss
            train_loss_d = sess.run(d_loss, 
                                feed_dict = {real_img: batch_images, 
                                             noise_img: batch_noise})
        # generator loss
            train_loss_g = sess.run(g_loss, 
                                feed_dict = {noise_img: batch_noise})
        
            
            print("Epoch {}/{}...".format(it, 1000000),
              "Discriminator Loss: {:.4f}...".format(train_loss_d),
              "Generator Loss: {:.4f}".format(train_loss_g))    
        # 记录各类loss值
            losses.append((train_loss_d, train_loss_g))
        # 存储checkpoints
        saver.save(sess, './checkpoints/generator.ckpt')