pdfcoffee

Generative Adversarial Networks
Now using the preceding defined generator and discriminator, we construct the
CycleGAN:
discriminator_A = Discriminator()
discriminator_B = Discriminator()
generator_AB = Generator()
generator_BA = Generator()
We next define the loss and optimizers:
loss_object = tf.keras.losses.BinaryCrossentropy(from_logits=True)
@tf.function
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_loss
return total_disc_loss
optimizer = tf.keras.optimizers.Adam(1e-4, beta_1=0.5)
discriminator_optimizer = tf.keras.optimizers.Adam(1e-4, beta_1=0.5)
We create placeholders for the labels of real and fake images:
valid = np.ones((BATCH_SIZE, 16, 16, 1)).astype('float32')
fake = np.zeros((BATCH_SIZE, 16, 16, 1)).astype('float32')
Now we define the function that trains the generator and discriminator in a batch,
a pair of images at a time. The two discriminators and the two generators are trained
via this function with the help of the tape gradient:
@tf.function
def train_batch(imgs_A, imgs_B):
with tf.GradientTape() as g, tf.GradientTape() as d_tape:
fake_B = generator_AB(imgs_A, training=True)
fake_A = generator_BA(imgs_B, training=True)
logits_real_A = discriminator_A(imgs_A, training=True)
logits_fake_A = discriminator_A(fake_A, training=True)
dA_loss = discriminator_loss(logits_real_A, logits_fake_A)
logits_real_B = discriminator_B(imgs_B, training=True)
logits_fake_B = discriminator_B(fake_B, training=True)
dB_loss = discriminator_loss(logits_real_B, logits_fake_B)
[ 224 ]