def train(self):
# iteration이 지날때 마다 linearly alpha를 증가시킴
step_pl = tf.placeholder(tf.float32, shape=None)
alpha_tra_assign = self.alpha_tra.assign(step_pl / self.max_iters)
# optimizer를 이용하여 loss를 줄이는 방향으로 variable을 업데이트해나간다.
opti_D = tf.train.AdamOptimizer(learning_rate=self.learning_rate,
beta1=0.0,
beta2=0.99).minimize(
self.D_loss, var_list=self.d_vars)
opti_G = tf.train.AdamOptimizer(learning_rate=self.learning_rate,
beta1=0.0,
beta2=0.99).minimize(
self.G_loss, var_list=self.g_vars)
# 변수들을 초기화하는 그래프
init = tf.global_variables_initializer()
# GPU 메모리를 필요할때 마다 조금씩 늘리도록 한다.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
# 변수들을 초기화하는 그래프를 돌림
sess.run(init)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(self.log_dir, sess.graph)
# progressive 값이 1, 7을 제외한 값이라면
if self.pg != 1 and self.pg != 7:
# transition이 일어나는 때라면
if self.trans:
# d_vars_n_read와 g_vars_n_read를 담당하는 r_saver가 이전 progressive에서
# training 하여 저장했던 weight들을 불러온다
self.r_saver.restore(sess, self.read_model_path)
# 이전에 학습했던 d_vars_n_2_rgb와 g_vars_n_2_rgb를 불러온다
self.rgb_saver.restore(sess, self.read_model_path)
# trainsition이 일어나는 단계가 아니라면
else:
# 전체 이전 단계의 전체 weight를 불러온다.
self.saver.restore(sess, self.read_model_path)
step = 0
batch_num = 0
# 16 x 32000 = 512,000개의 실제 이미지를 discriminator에게 보여줄 때까지 돌림
while step <= self.max_iters:
# optimization D
n_critic = 1
if self.pg == 5 and self.trans:
n_critic = 1
for i in range(n_critic):
# 512 짜리 latent vector를 만듬
sample_z = np.random.normal(
size=[self.batch_size, self.sample_size])
# 실제 이미지의 path를 batch단위씩 얻음 (16개씩)
train_list = self.data_In.getNextBatch(
batch_num, self.batch_size)
# 이 path를 현재 단계의 아웃풋 사이즈만큼 리사이즈 시킴
# 예를 들어 1단계에서 output_size가 4이므로 4x4로 resize한다.
realbatch_array = CelebA.getShapeForData(
train_list, resize_w=self.output_size)
# 만약 transition(fade)가 일어나는 차례라면
if self.trans and self.pg != 0:
alpha = np.float(step) / self.max_iters
# 이미지 해상도 변경
# https://stackoverflow.com/questions/37119071/scipy-rotate-and-zoom-an-image-without-changing-its-dimensions
low_realbatch_array = scipy.ndimage.zoom(
realbatch_array, zoom=[1, 0.5, 0.5, 1])
low_realbatch_array = scipy.ndimage.zoom(
low_realbatch_array, zoom=[1, 2, 2, 1])
# resolution transition중에 실제이미지의 resolution 사이의 보간을 한다.
realbatch_array = alpha * realbatch_array + (
1 - alpha) * low_realbatch_array
sess.run(opti_D,
feed_dict={
self.images: realbatch_array,
self.z: sample_z
})
# 다음 번 배치
batch_num += 1
# optimization G
sess.run(opti_G, feed_dict={self.z: sample_z})
summary_str = sess.run(summary_op,
feed_dict={
self.images: realbatch_array,
self.z: sample_z
})
summary_writer.add_summary(summary_str, step)
# the alpha of fake_in process
# 1씩 증가하는 step을 step_placeholder에 넣고 alpha_transition값을 계산
sess.run(alpha_tra_assign, feed_dict={step_pl: step})
# 400번째 step마다 loss 출력하고 이미지를 저장
if step % 400 == 0:
D_loss, G_loss, D_origin_loss, alpha_tra = sess.run(
[
self.D_loss, self.G_loss, self.D_origin_loss,
self.alpha_tra
],
feed_dict={
self.images: realbatch_array,
self.z: sample_z
})
print(
"PG %d, step %d: D loss=%.7f G loss=%.7f, D_or loss=%.7f, opt_alpha_tra=%.7f"
% (self.pg, step, D_loss, G_loss, D_origin_loss,
alpha_tra))
realbatch_array = np.clip(realbatch_array, -1, 1)
save_images(
realbatch_array[0:self.batch_size],
[2, self.batch_size / 2],
'{}/{:02d}_real.png'.format(self.sample_path, step))
# 만약 transition 단계인 경우에
if self.trans and self.pg != 0:
low_realbatch_array = np.clip(low_realbatch_array, -1,
1)
save_images(
low_realbatch_array[0:self.batch_size],
[2, self.batch_size / 2],
'{}/{:02d}_real_lower.png'.format(
self.sample_path, step))
fake_image = sess.run(self.fake_images,
feed_dict={
self.images: realbatch_array,
self.z: sample_z
})
fake_image = np.clip(fake_image, -1, 1)
save_images(
fake_image[0:self.batch_size],
[2, self.batch_size / 2],
'{}/{:02d}_train.png'.format(self.sample_path, step))
# 4000번째 마다 g_vars와 d_vars (network의 weight)를 중간저장함
if np.mod(step, 4000) == 0 and step != 0:
self.saver.save(sess, self.gan_model_path)
step += 1
# max_iter 끝나고 최종 모델을 저장함
save_path = self.saver.save(sess, self.gan_model_path)
print("Model saved in file: %s" % save_path)
tf.reset_default_graph()
def train(self):
step_pl = tf.placeholder(tf.float32, shape=None)
alpha_tra_assign = self.alpha_tra.assign(step_pl / self.max_iters)
opti_D = tf.train.AdamOptimizer(learning_rate=self.learning_rate,
beta1=0.0,
beta2=0.99).minimize(
self.D_loss, var_list=self.d_vars)
opti_G = tf.train.AdamOptimizer(learning_rate=self.learning_rate,
beta1=0.0,
beta2=0.99).minimize(
self.G_loss, var_list=self.g_vars)
init = tf.global_variables_initializer()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(self.log_dir, sess.graph)
if self.pg != 1 and self.pg != 7:
if self.trans:
self.r_saver.restore(sess, self.read_model_path)
self.rgb_saver.restore(sess, self.read_model_path)
else:
self.saver.restore(sess, self.read_model_path)
step = 0
batch_num = 0
while step <= self.max_iters:
# optimization D
n_critic = 1
if self.pg == 5 and self.trans:
n_critic = 1
for i in range(n_critic):
sample_z = np.random.normal(
size=[self.batch_size, self.sample_size])
train_list = self.data_In.getNextBatch(
batch_num, self.batch_size)
realbatch_array = CelebA.getShapeForData(
train_list, resize_w=self.output_size)
if self.trans and self.pg != 0:
alpha = np.float(step) / self.max_iters
low_realbatch_array = scipy.ndimage.zoom(
realbatch_array, zoom=[1, 0.5, 0.5, 1])
low_realbatch_array = scipy.ndimage.zoom(
low_realbatch_array, zoom=[1, 2, 2, 1])
realbatch_array = alpha * realbatch_array + (
1 - alpha) * low_realbatch_array
sess.run(opti_D,
feed_dict={
self.images: realbatch_array,
self.z: sample_z
})
batch_num += 1
# optimization G
sess.run(opti_G, feed_dict={self.z: sample_z})
summary_str = sess.run(summary_op,
feed_dict={
self.images: realbatch_array,
self.z: sample_z
})
summary_writer.add_summary(summary_str, step)
# the alpha of fake_in process
sess.run(alpha_tra_assign, feed_dict={step_pl: step})
if step % 1000 == 0:
D_loss, G_loss, D_origin_loss, alpha_tra = sess.run(
[
self.D_loss, self.G_loss, self.D_origin_loss,
self.alpha_tra
],
feed_dict={
self.images: realbatch_array,
self.z: sample_z
})
print(
"PG %d, step %d: D loss=%.7f G loss=%.7f, D_or loss=%.7f, opt_alpha_tra=%.7f"
% (self.pg, step, D_loss, G_loss, D_origin_loss,
alpha_tra))
realbatch_array = np.clip(realbatch_array, -1, 1)
save_images(
realbatch_array[0:self.batch_size],
[2, self.batch_size / 2],
'{}/{:02d}_real.png'.format(self.sample_path, step))
if self.trans and self.pg != 0:
low_realbatch_array = np.clip(low_realbatch_array, -1,
1)
save_images(
low_realbatch_array[0:self.batch_size],
[2, self.batch_size / 2],
'{}/{:02d}_real_lower.png'.format(
self.sample_path, step))
fake_image = sess.run(self.fake_images,
feed_dict={
self.images: realbatch_array,
self.z: sample_z
})
fake_image = np.clip(fake_image, -1, 1)
save_images(
fake_image[0:self.batch_size],
[2, self.batch_size / 2],
'{}/{:02d}_train.png'.format(self.sample_path, step))
if np.mod(step, 4000) == 0 and step != 0:
self.saver.save(sess, self.gan_model_path)
step += 1
save_path = self.saver.save(sess, self.gan_model_path)
print("Model saved in file: %s" % save_path)
tf.reset_default_graph()