![[GAN实战] DCGAN实现](http://www.mshxw.com/aiimages/31/444538.png "[GAN实战] DCGAN实现")
'''
>>> import tensorflow as tf;tf.__version__
2021-11-07 23:53:04.980446: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1
'2.3.0'
>>> import tensorflow_datasets as tfds;tfds.__version__
'4.3.0'
>>>
'''
import tensorflow as tf
from tensorflow.keras import layers, Model
from tensorflow.keras.activations import relu
from tensorflow.keras.models import Sequential, load_model
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
from tensorflow.keras.losses import BinaryCrossentropy
from tensorflow.keras.optimizers import RMSprop, Adam
from tensorflow.keras.metrics import binary_accuracy
import tensorflow_datasets as tfds
import numpy as np
import matplotlib.pyplot as plt
# load datasets
ds_train, ds_info = tfds.load('fashion_mnist', split='train', shuffle_files=True, with_info=True)
fig = tfds.show_examples(ds_info, ds_train)
batch_size = 200
image_shape = (28, 28, 1)
def preprocess(features):
image = tf.image.resize(features['image'], image_shape[:2])
image = tf.cast(image, tf.float32)
image = (image-127.5)/127.5
return image
ds_train = ds_train.map(preprocess)
ds_train = ds_train.cache() # put dataset into memory
ds_train = ds_train.shuffle(ds_info.splits['train'].num_examples)
ds_train = ds_train.batch(batch_size).repeat()
train_num = ds_info.splits['train'].num_examples
train_steps_per_epoch = round(train_num/batch_size)
print(train_steps_per_epoch)
class GAN():
def __init__(self, generator, discriminator):
# discriminator
self.D = discriminator
self.G = generator
self.bce = tf.keras.losses.BinaryCrossentropy()
self.d_loss = {}
self.g_loss = {}
self.accuracy = {}
self.g_gradients = []
def discriminator_loss(self, pred_fake, pred_real):
real_loss = self.bce(tf.ones_like(pred_real), pred_real)
fake_loss = self.bce(tf.zeros_like(pred_fake), pred_fake)
d_loss = 0.5*(real_loss + fake_loss)
return d_loss
def generator_loss(self, pred_fake):
g_loss = self.bce(tf.ones_like(pred_fake), pred_fake)
return g_loss
def train_step(self, g_input, real_input):
with tf.GradientTape() as g_tape,
tf.GradientTape() as d_tape:
# Feed forward
fake_input = self.G(g_input)
pred_fake = self.D(fake_input)
pred_real = self.D(real_input)
# Calculate losses
d_loss = self.discriminator_loss(pred_fake, pred_real)
g_loss = self.generator_loss(pred_fake)
# Accuracy
fake_accuracy = tf.math.reduce_mean(binary_accuracy(tf.zeros_like(pred_fake), pred_fake))
real_accuracy = tf.math.reduce_mean(binary_accuracy(tf.ones_like(pred_real), pred_real))
# backprop gradients
gradient_g = g_tape.gradient(g_loss, self.G.trainable_variables)
gradient_d = d_tape.gradient(d_loss, self.D.trainable_variables)
gradient_g_l1_norm = [tf.norm(gradient).numpy() for gradient in gradient_g]
self.g_gradients.append(gradient_g_l1_norm)
# update weights
self.G_optimizer.apply_gradients(zip(gradient_g, self.G.trainable_variables))
self.D_optimizer.apply_gradients(zip(gradient_d, self.D.trainable_variables))
return g_loss, d_loss, fake_accuracy, real_accuracy
def train(self, data_generator,
z_generator,
g_optimizer, d_optimizer,
steps, interval=100):
self.D_optimizer = d_optimizer
self.G_optimizer = g_optimizer
val_g_input = next(z_generator)
for i in range(steps):
g_input = next(z_generator)
real_input = next(data_generator)
g_loss, d_loss, fake_accuracy, real_accuracy = self.train_step(g_input, real_input)
self.d_loss[i] = d_loss.numpy()
self.g_loss[i] = g_loss.numpy()
self.accuracy[i] = 0.5*(fake_accuracy.numpy() + real_accuracy.numpy())
if i%interval == 0:
msg = "Step {}: d_loss {:.4f} g_loss {:.4f} Accuracy. real : {:.3f} fake : {:.3f}"
.format(i, d_loss, g_loss, real_accuracy, fake_accuracy)
print(msg)
fake_images = self.G(val_g_input)
self.plot_images(fake_images)
def plot_images(self, images):
pass
class DCGAN(GAN):
def __init__(self, z_dim, input_shape):
discriminator = self.Discriminator(input_shape)
generator = self.Generator(z_dim)
GAN.__init__(self, generator, discriminator)
def Discriminator(self, input_shape):
model = tf.keras.Sequential(name='Discriminator')
model.add(layers.Input(shape=input_shape))
model.add(layers.Conv2D(32, 3, strides=(2,2), padding='same'))
model.add(layers.BatchNormalization(momentum=0.9))
model.add(layers.LeakyReLU(0.2))
model.add(layers.Dropout(0.2))
model.add(layers.Conv2D(64, 3, strides=(2,2), padding='same'))
model.add(layers.BatchNormalization(momentum=0.9))
model.add(layers.LeakyReLU(0.2))
model.add(layers.Dropout(0.2))
model.add(layers.Flatten())
model.add(layers.Dense(1, activation='sigmoid'))
return model
def Generator(self, z_dim):
model = tf.keras.Sequential(name='Generator')
model.add(layers.Input(shape=[z_dim]))
model.add(layers.Dense(7*7*64))
model.add(layers.BatchNormalization(momentum=0.9))
model.add(layers.ReLU())
model.add(layers.Reshape((7,7,64)))
model.add(layers.Conv2D(64, 3, padding='same'))
model.add(layers.BatchNormalization(momentum=0.9))
model.add(layers.ReLU())
model.add(layers.UpSampling2D((2,2), interpolation="bilinear"))
model.add(layers.Conv2D(32, 3, padding='same'))
model.add(layers.ReLU())
model.add(layers.UpSampling2D((2,2), interpolation="bilinear"))
model.add(layers.Conv2D(image_shape[-1], 3, padding='same', activation='tanh'))
return model
def plot_images(self, images):
grid_row = 1
grid_col = 8
f, axarr = plt.subplots(grid_row, grid_col, figsize=(grid_col*1.5, grid_row*1.5))
for col in range(grid_col):
axarr[col].imshow((images[col,:,:,0]+1)/2, cmap='gray')
axarr[col].axis('off')
plt.show()
z_dim = 100
def z_generator(batch_size, z_dim):
while True:
yield tf.random.normal((batch_size, z_dim))
z_gen = z_generator(batch_size, z_dim)
#gan.D.summary()
#gan.G.summary()
gan = DCGAN(z_dim, image_shape)
gan.D.summary()
'''
Model: "Discriminator"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 14, 14, 32) 320
_________________________________________________________________
batch_normalization (BatchNo (None, 14, 14, 32) 128
_________________________________________________________________
leaky_re_lu (LeakyReLU) (None, 14, 14, 32) 0
_________________________________________________________________
dropout (Dropout) (None, 14, 14, 32) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 7, 7, 64) 18496
_________________________________________________________________
batch_normalization_1 (Batch (None, 7, 7, 64) 256
_________________________________________________________________
leaky_re_lu_1 (LeakyReLU) (None, 7, 7, 64) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 7, 7, 64) 0
_________________________________________________________________
flatten (Flatten) (None, 3136) 0
_________________________________________________________________
dense (Dense) (None, 1) 3137
=================================================================
Total params: 22,337
Trainable params: 22,145
Non-trainable params: 192
'''
gan.G.summary()
'''
Model: "Generator"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_1 (Dense) (None, 3136) 316736
_________________________________________________________________
batch_normalization_2 (Batch (None, 3136) 12544
_________________________________________________________________
re_lu (ReLU) (None, 3136) 0
_________________________________________________________________
reshape (Reshape) (None, 7, 7, 64) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 7, 7, 64) 36928
_________________________________________________________________
batch_normalization_3 (Batch (None, 7, 7, 64) 256
_________________________________________________________________
re_lu_1 (ReLU) (None, 7, 7, 64) 0
_________________________________________________________________
up_sampling2d (UpSampling2D) (None, 14, 14, 64) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 14, 14, 32) 18464
_________________________________________________________________
re_lu_2 (ReLU) (None, 14, 14, 32) 0
_________________________________________________________________
up_sampling2d_1 (UpSampling2 (None, 28, 28, 32) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None, 28, 28, 1) 289
=================================================================
Total params: 385,217
Trainable params: 378,817
Non-trainable params: 6,400
'''
gan.train(iter(ds_train), z_gen,
RMSprop(3e-4), RMSprop(3e-4),
25*train_steps_per_epoch,
2*train_steps_per_epoch)
#plt.figure(figsize=(10,6))
fig, (ax1, ax2) = plt.subplots(2, sharex=True)
fig.set_figwidth(10)
fig.set_figheight(8)
ax1.plot(list(gan.d_loss.values())[:5000], label='D loss', alpha=0.7)
ax1.set_title("D loss")
ax2.plot(list(gan.g_loss.values())[:5000], label='G loss', alpha=0.7)
ax2.set_title("G loss")
#ax3.plot([grad[0] for grad in gan.g_gradients], label='G loss', alpha=0.7)
#ax3.set_title("Gradient")
plt.xlabel('Steps')
训练过程：
损失函数:
