def main():
global k_p
if not os.path.exists(FLAGS.working_directory):
os.makedirs(FLAGS.working_directory)
if not os.path.exists(FLAGS.data_directory):
os.makedirs(FLAGS.data_directory)
if os.path.exists(os.path.join(FLAGS.working_directory,"save","ver")):
ver_desc = open(os.path.join(FLAGS.working_directory,"save","ver"))
FLAGS.version = int(ver_desc.readline())
ver_desc.close()
else:
if not os.path.exists(os.path.join(FLAGS.working_directory,"save")):
os.mkdir(os.path.join(FLAGS.working_directory,"save"))
ver_desc = open(os.path.join(FLAGS.working_directory,"save","ver"),'w')
ver_desc.write(str(FLAGS.version)+"\n")
ver_desc.close()
load_desc()
gan = GAN(FLAGS.version, FLAGS.clip_abs, FLAGS.hidden_size, FLAGS.batch_size, FLAGS.learning_rate, FLAGS.data_directory, os.path.join(FLAGS.working_directory, "log"))
saver = tf.train.Saver()
load_model(gan.sess, saver)
# start the queue runners
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=gan.sess, coord=coord)
for epoch in range(start_epoch, FLAGS.max_epoch):
pbar = ProgressBar()
for update in pbar(range(FLAGS.updates_per_epoch)):
k_p = gan.update_params(FLAGS.global_step, FLAGS.learning_rate, k_p)
FLAGS.global_step = FLAGS.global_step + 1
if FLAGS.global_step % FLAGS.lr_update == FLAGS.lr_update - 1:
FLAGS.learning_rate = FLAGS.learning_rate*0.5
print "update learning rate:"+str(FLAGS.learning_rate)
cm = gan.get_loss()
print "loss: " + str(cm)
gan.generate_and_save_images(64, FLAGS.working_directory) #(int(math.sqrt(FLAGS.batch_size))+1)**2
save_model(gan.sess, saver)
# ask threads to stop
coord.request_stop()
# wait for threads to finish
coord.join(threads)
gan.sess.close()
training_loss = 0.0
batch_size = 128
for i in range(40):
images, _ = mnist.train.next_batch(128)
flagReal = kneighbors_graph(images,
5,
mode='connectivity',
include_self=True)
disReal = kneighbors_graph(images,
5,
mode='distance',
include_self=True)
flagReal = flagReal.toarray()
disReal = disReal.toarray()
flagReal2 = (flagReal - 1)
flagReal2 = flagReal2 * flagReal2
disReal = 100 * flagReal2 + flagReal * 1
for i in range(batch_size):
Xtemp2 = np.matlib.repmat(images[i, :], batch_size, 1)
disTemp = disReal[i, :].reshape(1, batch_size)
flagTemp = flagReal[i, :].reshape(1, batch_size)
# images=orderedImage[i*128:(i+1)*128,:]
loss_value = model.update_params(images, images, flagTemp,
disTemp)
training_loss += loss_value
training_loss = training_loss / (1000 * 128)
print("Loss %f" % training_loss)
model.generate_and_save_images(128, "")
f_inputs=images[1],
sum_epoch=epoch)
del images
if FLAGS.save_imgs:
images = novaSet.get_data(which_config = ['data'],\
which_type = None,\
indices = image_indices,\
is_train=FLAGS.save_train)
print("Generating and saving images for epoch {}...".
format(epoch))
image_dir = join(FLAGS.working_directory, "") + join(
config, config_type + "/imgs")
model.generate_and_save_images(images[0],
image_indices,
join(image_dir,
"fake"), epoch)
print("Saving final model...")
model.save_model(FLAGS.max_epoch + 1)
print("Generating and saving final image summary")
images = novaSet.next_batch(which_config=[config, 'data'],
which_type=config_type,
batch_size=5)
model.save_image_summary(r_inputs=images[0],
f_inputs=images[1],
sum_epoch=epoch)
del images
if FLAGS.save_imgs:
