def _create(self):
# placeholders for training data
x_images_holder = tf.placeholder(
tf.float32,
shape=[None, A.img_size, A.img_size, A.input_channels],
name='x_images_holder')
y_images_holder = tf.placeholder(
tf.float32,
shape=[None, A.img_size, A.img_size, A.output_channels],
name='y_images_holder')
# forward pass
G = Generator()
D = Discriminator()
generated_images = G.create(x_images_holder)
Dx = D.score_patches(x_images_holder, y_images_holder)
Dg = D.score_patches(x_images_holder, generated_images, reuse=True)
prob_x = tf.reduce_mean(Dx, name='prob_x')
prob_g = tf.reduce_mean(Dg, name='prob_g')
# compute losses
loss_d, loss_g, loss_L1 = self._loss(Dx, Dg, y_images_holder,
generated_images)
# optimizers
optimizer_g = tf.train.AdamOptimizer(learning_rate=0.0002, beta1=0.5)
optimizer_d = tf.train.AdamOptimizer(learning_rate=0.0002, beta1=0.5)
# backprop
g_vars = tf.trainable_variables(scope='generator')
d_vars = tf.trainable_variables(scope='discriminator')
train_g = optimizer_g.minimize(loss_g, var_list=g_vars, name='train_g')
train_d = optimizer_d.minimize(loss_d, var_list=d_vars, name='train_d')
# prepare summaries
loss_d_summary_op = tf.summary.scalar('Discriminator_Loss', loss_d)
loss_g_summary_op = tf.summary.scalar('Generator_Loss', loss_g)
images_summary_op = tf.summary.image('Generated_Image',
generated_images,
max_outputs=1)
x_summary_op = tf.summary.image('X_Image',
x_images_holder,
max_outputs=1)
y_summary_op = tf.summary.image('Y_Image',
y_images_holder,
max_outputs=1)
dx_summary_op = tf.summary.scalar('prob_x', prob_x)
dg_summary_op = tf.summary.scalar('prob_g', prob_g)
l1_summary_op = tf.summary.scalar('loss_L1', loss_L1)
summary_op = tf.summary.merge_all()
def test_discriminator(self):
D = Discriminator()
# create two fake tensors of shape [batch_size, img_size, img_size, output_channels]
batch_size = 128
output_shape = [batch_size, A.img_size, A.img_size, A.output_channels]
input_images = tf.random_uniform(output_shape, minval=-1, maxval=1)
output_images = tf.random_uniform(output_shape, minval=-1, maxval=1)
# get output from D
output = D.score_patches(input_images, output_images)
# confirm output shape has one channel
self.assertEqual(output.shape[3], 1)
