def build(x):
real_label = 1.0
fake_label = 0.0
tf.summary.histogram('input', x)
tf.summary.image(
'input_img',
tf.cast(tf.clip_by_value(input_data.invert_norm(x), 0, 255), tf.uint8))
# (1) Update D network: maximize log(D(x)) + log(1 - D(G(z)))
# train with real
optimizerD = tf.train.AdamOptimizer(learning_rate=opts.learning_rate,
beta1=opts.adam_beta1,
beta2=opts.adam_beta2,
epsilon=opts.opt_epsilon)
output_x = discriminator_net(x, True, opts)
errD_real = criterion(real_label, output_x)
tf.summary.scalar('errD_real', errD_real)
# train with fake
noise = tf.random.uniform([opts.batch_size, 1, 1, opts.nz], -1.0, 1.0)
fake = generator_net(noise, True, opts)
tf.summary.histogram('fake', fake)
tf.summary.image(
'fake_img',
tf.cast(tf.clip_by_value(input_data.invert_norm(fake), 0, 255),
tf.uint8))
output_fake = discriminator_net(fake, True, opts)
errD_fake = criterion(fake_label, output_fake)
errD = errD_real + errD_fake
tf.summary.scalar('errD_fake', errD_fake)
tf.summary.scalar('errD', errD)
# (2) Update G network: maximize log(D(G(z)))
optimizerG = tf.train.AdamOptimizer(learning_rate=opts.learning_rate,
beta1=opts.adam_beta1,
beta2=opts.adam_beta2,
epsilon=opts.opt_epsilon)
errG = criterion(real_label, output_fake)
tf.summary.scalar('errG', errG)
t_vars = tf.trainable_variables()
d_vars = [var for var in t_vars if 'Discriminator' in var.name]
g_vars = [var for var in t_vars if 'Generator' in var.name]
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
train_op_D = optimizerD.minimize(loss=errD, var_list=d_vars)
# first update D newwork, then update G network twice
with tf.control_dependencies([train_op_D]):
train_op_G_1 = optimizerG.minimize(loss=errG, var_list=g_vars)
with tf.control_dependencies([train_op_G_1]):
train_op_G_2 = optimizerG.minimize(loss=errG, var_list=g_vars)
with tf.control_dependencies(update_ops):
train_op = train_op_G_2
return train_op, errD_real, errD_fake, errG
def sample(filename='output.jpg'):
with tf.Graph().as_default():
batch_size = 12
inoise = tf.random.uniform([batch_size, opts.nz], -1.0, 1.0)
categorical_code = tf.fill((batch_size, 1), 9)
categorical_code = tf.one_hot(categorical_code, opts.num_categorical)
categorical_code = tf.reshape(categorical_code, (batch_size, -1))
continuous_code = tf.random.normal((batch_size, opts.num_continuous))
noise = tf.concat([inoise, categorical_code, continuous_code], 1)
noise = tf.reshape(noise, (batch_size, 1, 1, -1))
print(noise)
# set training True for good quality image
fake = generator_net(noise, False, opts)
checkpoint_path = opts.sample_checkpoint_path
if tf.gfile.IsDirectory(checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(checkpoint_path)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, checkpoint_path)
output = sess.run(fake)
output = input_data.invert_norm(output)
for idx, img in enumerate(output):
if opts.nc == 1:
img = np.squeeze(img, -1)
utils.imsave(str(idx + 1) + '_' + filename, img)
def sample(filename='output.jpg'):
with tf.Graph().as_default():
# set training True for good quality image
fake = generator_net(tf.constant(fixed_noise), True, opts)
checkpoint_path = opts.sample_checkpoint_path
if tf.gfile.IsDirectory(checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(checkpoint_path)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, checkpoint_path)
output = sess.run(fake)
output = input_data.invert_norm(output)
output = np.squeeze(output, 0)
if opts.nc == 1:
output = np.squeeze(output, -1)
utils.imsave(filename, output)
def add_summary_img(name, img):
tf.summary.image(
name,
tf.cast(tf.clip_by_value(input_data.invert_norm(img), 0, 255),
tf.uint8))