def load_data(args):
img_ext = '/*.png'
print("Using input directories:")
if args.triplet:
dir_train_phase1 = './datasets/' + args.dataset + '/train' + args.stage1 + img_ext
dir_test_phase1 = './datasets/' + args.dataset + '/test' + args.stage1 + img_ext
print(dir_train_phase1)
print(dir_test_phase1)
a_img_paths = glob(dir_train_phase1)
a_data_pool = data.ImageData(sess, a_img_paths, args.batch_size, load_size=args.load_size, crop_size=args.crop_size)
a_test_img_paths = glob(dir_test_phase1)
a_test_pool = data.ImageData(sess, a_test_img_paths, args.batch_size, load_size=args.load_size, crop_size=args.crop_size)
else:
a_data_pool = a_test_pool = None
dir_train_phase2 = './datasets/' + args.dataset + '/train' + args.stage2 + img_ext
dir_train_phase3 = './datasets/' + args.dataset + '/train' + args.stage3 + img_ext
dir_test_phase2 = './datasets/' + args.dataset + '/test' + args.stage2 + img_ext
dir_test_phase3 = './datasets/' + args.dataset + '/test' + args.stage3 + img_ext
print(dir_train_phase2)
print(dir_train_phase3)
print(dir_test_phase2)
print(dir_test_phase3)
b_img_paths = glob(dir_train_phase2)
c_img_paths = glob(dir_train_phase3)
b_data_pool = data.ImageData(sess, b_img_paths, args.batch_size, load_size=args.load_size, crop_size=args.crop_size)
c_data_pool = data.ImageData(sess, c_img_paths, args.batch_size, load_size=args.load_size, crop_size=args.crop_size)
b_test_img_paths = glob(dir_test_phase2)
c_test_img_paths = glob(dir_test_phase3)
b_test_pool = data.ImageData(sess, b_test_img_paths, args.batch_size, load_size=args.load_size, crop_size=args.crop_size)
c_test_pool = data.ImageData(sess, c_test_img_paths, args.batch_size, load_size=args.load_size, crop_size=args.crop_size)
return a_data_pool, b_data_pool, c_data_pool, a_test_pool, b_test_pool, c_test_pool
""" train """
''' init '''
# session
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# counter
it_cnt, update_cnt = utils.counter()
''' data '''
a_img_paths = glob('./datasets/' + dataset + '/trainA/*.png')
b_img_paths = glob('./datasets/' + dataset + '/trainB/*.png')
a_data_pool = data.ImageData(sess,
a_img_paths,
batch_size,
load_size=load_size,
crop_size=crop_size,
num_threads=num_threads,
buffer_size=buffer_size)
b_data_pool = data.ImageData(sess,
b_img_paths,
batch_size,
load_size=load_size,
crop_size=crop_size,
num_threads=num_threads,
buffer_size=buffer_size)
a_test_img_paths = glob('./datasets/' + dataset + '/testA/*.png')
b_test_img_paths = glob('./datasets/' + dataset + '/testB/*.png')
a_test_pool = data.ImageData(sess,
a_test_img_paths,
var_list=g_var)
""" train """
''' init '''
# session
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# counter
it_cnt, update_cnt = ops.counter()
'''data'''
a_img_paths = glob('./datasets/' + dataset + '/trainA/*.jpg')
b_img_paths = glob('./datasets/' + dataset + '/trainB/*.jpg')
a_data_pool = data.ImageData(sess,
a_img_paths,
batch_size,
load_size=load_size,
crop_size=crop_size)
b_data_pool = data.ImageData(sess,
b_img_paths,
batch_size,
load_size=load_size,
crop_size=crop_size)
a_test_img_paths = glob('./datasets/' + dataset + '/testA/*.jpg')
b_test_img_paths = glob('./datasets/' + dataset + '/testB/*.jpg')
a_test_pool = data.ImageData(sess,
a_test_img_paths,
batch_size,
load_size=load_size,
crop_size=crop_size)
# counter
it_cnt, update_cnt = utils.counter()
''' data '''
a_img_paths = glob('./datasets/' + dataset + '/trainA/*')
b_img_paths = glob('./datasets/' + dataset + '/trainB/*')
ab_pair_data_pool = data.ImageDataPair(sess,
a_img_paths,
batch_size,
load_size=load_size,
crop_size=crop_size)
a_data_pool = data.ImageData(sess,
a_img_paths,
batch_size,
load_size=load_size,
crop_size=crop_size,
channels=3)
b_data_pool = data.ImageData(sess,
b_img_paths,
batch_size,
load_size=load_size,
crop_size=crop_size,
channels=1)
a_test_img_paths = glob('./datasets/' + dataset + '/testA/*')
ab_pair_test_pool = data.ImageDataPair(sess,
a_test_img_paths,
batch_size=len(a_test_img_paths),
load_size=load_size,
# Session management
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
it_cnt, update_cnt = ops.counter()
if do_train:
summary_writer = tf.summary.FileWriter('./summaries/' + args.dataset+ '/train-'+training_run_id(), sess.graph)
# Data loading
if not singleTestOnly:
a_data_pool, b_data_pool, c_data_pool, a_test_pool, b_test_pool, c_test_pool = load_data(args)
else:
single_test_input_pool = data.ImageData(sess, glob(args.singletestdir+'/*.png'), 1, load_size=args.load_size, crop_size=args.crop_size, shuffle = False, random_flip = True) #Fix the random flip problem, see data.py, then make the flip False.
b2c_pool = utils.ItemPool()
c2b_pool = utils.ItemPool()
a2b_pool = utils.ItemPool()
b2a_pool = utils.ItemPool()
# Checkpoint management.
saver = tf.train.Saver(max_to_keep=5)
# If the triplet mode is enabled, we try to load the existing checkpoint for that first.
# Otherwise, we try to load the regular checkpoint only.
subnet_maybe = ('/'+args.subnet) if len(args.subnet) > 0 else ''
subnet_ext_maybe = (subnet_maybe + ('-transitive2')) if args.transform_twice else subnet_maybe
d_a_train = tf.train.AdamOptimizer(lr, beta1=0.5).minimize(d_loss_a,
var_list=d_a_var)
d_b_train = tf.train.AdamOptimizer(lr, beta1=0.5).minimize(d_loss_b,
var_list=d_b_var)
g_train = tf.train.AdamOptimizer(lr, beta1=0.5).minimize(g_loss,
var_list=g_var)
''' Train '''
sess = tf.Session()
cnt, update_cnt = util.counter()
trainA_path = glob('./datasets/' + dataset + '/trainA/*.jpg')
trainB_path = glob('./datasets/' + dataset + '/trainB/*.jpg')
trainA_pool = data.ImageData(sess,
trainA_path,
batch_size,
load_size=load_size,
crop_size=crop_size)
trainB_pool = data.ImageData(sess,
trainB_path,
batch_size,
load_size=load_size,
crop_size=crop_size)
testA_path = glob('./datasets/' + dataset + '/testA/*.jpg')
testB_path = glob('./datasets/' + dataset + '/testB/*.jpg')
testA_pool = data.ImageData(sess,
testA_path,
batch_size,
load_size=load_size,
crop_size=crop_size)
