def train():
"""
"""
print('training')
model = Model()
reader = ore.RandomReader(ore.DATASET_KAGGLE_MNIST_TRAINING)
while True:
images, labels, is_new_batch = next_batch(reader, 128)
loss, step = model.train(images, labels)
print('loss: {}'.format(loss))
if step % 1000 == 0:
model.save_checkpoint()
def train():
"""
"""
reader = ore.RandomReader(ore.DATASET_LSUN_BEDROOM_TRAINING)
# tensorflow
checkpoint_source_path = tf.train.latest_checkpoint(
FLAGS.checkpoints_dir_path)
checkpoint_target_path = os.path.join(FLAGS.checkpoints_dir_path,
'model.ckpt')
gan_graph = build_xwgan()
summaries = build_summaries(gan_graph)
reporter = tf.summary.FileWriter(FLAGS.logs_dir_path)
with tf.Session() as session:
if checkpoint_source_path is None:
session.run(tf.global_variables_initializer())
else:
tf.train.Saver().restore(session, checkpoint_source_path)
# give up overlapped old data
global_step = session.run(gan_graph['global_step'])
reporter.add_session_log(tf.SessionLog(status=tf.SessionLog.START),
global_step=global_step)
while True:
for _ in range(5):
real_sources = next_real_batch(reader)
fake_sources = next_fake_batch()
fetches = [
gan_graph['discriminator_loss'],
gan_graph['discriminator_trainer'],
summaries['discriminator_loss_summary']
]
feeds = {
gan_graph['seed']: fake_sources,
gan_graph['real']: real_sources,
}
returns = session.run(fetches, feed_dict=feeds)
d_loss_summary = returns[2]
reporter.add_summary(d_loss_summary, global_step)
#
fake_sources = next_fake_batch()
log_fakes = (global_step % 500 == 0)
fetches = [
gan_graph['global_step'],
gan_graph['generator_loss'],
gan_graph['generator_trainer'],
summaries['generator_loss_summary'],
]
feeds = {gan_graph['seed']: fake_sources}
if log_fakes:
fetches.append(summaries['generator_fake_summary'])
returns = session.run(fetches, feed_dict=feeds)
global_step = returns[0]
g_loss_summary = returns[3]
reporter.add_summary(g_loss_summary, global_step)
if log_fakes:
reporter.add_summary(returns[4], global_step)
if global_step % 100 == 0:
print('[{}]: {}'.format(global_step, returns[1]))
if global_step % 500 == 0:
tf.train.Saver().save(session,
checkpoint_target_path,
global_step=gan_graph['global_step'])
def train():
"""
"""
if FLAGS.use_lsun:
reader = ore.RandomReader(ore.DATASET_LSUN_BEDROOM_TRAINING)
else:
reader = ore.RandomReader(ore.DATASET_MNIST_TRAINING)
# tensorflow
checkpoint_source_path = tf.train.latest_checkpoint(
FLAGS.checkpoints_dir_path)
checkpoint_target_path = os.path.join(FLAGS.checkpoints_dir_path,
'model.ckpt')
gan_graph = build_dcgan()
summaries = build_summaries(gan_graph)
reporter = tf.summary.FileWriter(FLAGS.logs_dir_path)
with tf.Session() as session:
if checkpoint_source_path is None:
session.run(tf.global_variables_initializer())
else:
tf.train.Saver().restore(session, checkpoint_source_path)
# give up overlapped old data
global_step = session.run(gan_graph['global_step'])
reporter.add_session_log(tf.SessionLog(status=tf.SessionLog.START),
global_step=global_step)
while True:
real_sources = next_real_batch(reader)
fake_sources = next_fake_batch()
fetches = {
'discriminator_loss': gan_graph['discriminator_loss'],
'discriminator_trainer': gan_graph['discriminator_trainer'],
'summary': summaries['summary_discriminator']
}
feeds = {
gan_graph['seed']: fake_sources,
gan_graph['real']: real_sources,
}
fetched = session.run(fetches, feed_dict=feeds)
reporter.add_summary(fetched['summary'], global_step)
#
fetches = {
'global_step': gan_graph['global_step'],
'generator_loss': gan_graph['generator_loss'],
'generator_trainer': gan_graph['generator_trainer'],
}
feeds = {gan_graph['seed']: fake_sources}
if global_step % 500 == 0:
fetches['summary'] = summaries['summary_generator_plus']
else:
fetches['summary'] = summaries['summary_generator']
fetched = session.run(fetches, feed_dict=feeds)
global_step = fetched['global_step']
reporter.add_summary(fetched['summary'], global_step)
if global_step % 100 == 0:
print('[{}]'.format(global_step))
if global_step % 500 == 0:
tf.train.Saver().save(session,
checkpoint_target_path,
global_step=gan_graph['global_step'])