def get_model(input_height, input_width, channels=3, input_count=3):
# X, Y, Mask
model_input = tf.keras.Input((input_count, input_height, input_width, channels))
input_noise, input_real, input_mask = model_input
encoded = mm.encoder(cat(input_noise, input_mask))
decoded = mm.decoder(encoded)
image_result = decoded * (1 - input_mask) + input_real * input_mask
probabilities = mm.discriminator_red(cat(image_result, input_real))
image_loss, input_loss = probabilities
hinge_loss = tf.keras.activations.relu(cat(1-image_loss, 1+input_loss))
def inference(from_file_path, args, batch_size=1, device_t='/gpu:0'):
imgs, style_name = args[0:2]
# check if the specified model exists
checkpoint_dir = os.path.join(root_models_dir, style_name,
"checkpoint_long")
assert os.path.exists(checkpoint_dir), 'Checkpoint not found!'
ngf = 32 # args.ngf
OPTIONS = namedtuple('OPTIONS', 'gf_dim is_training')
options = OPTIONS._make((ngf, False))
tfconfig = tf.ConfigProto(
allow_soft_placement=False) # THIS IS DIFFERENT FROM IMPL_1
tfconfig.gpu_options.allow_growth = True
tf.reset_default_graph()
with tf.Session(config=tfconfig) as sess:
# ==================== Define placeholders. ===================== #
with tf.name_scope('placeholder'):
input_photo = tf.placeholder(dtype=tf.float32,
shape=[batch_size, None, None, 3],
name='photo')
# ===================== Wire the graph. ========================= #
# Encode input images.
input_photo_features = encoder(image=input_photo,
options=options,
reuse=False)
# Decode obtained features.
output_photo = decoder(features=input_photo_features,
options=options,
reuse=False)
init_op = tf.global_variables_initializer()
sess.run(init_op)
saver = tf.train.Saver()
if os.path.isdir(checkpoint_dir):
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
raise Exception("No checkpoint found...")
else:
saver.restore(sess, checkpoint_dir)
model_args = (sess, input_photo, output_photo)
if from_file_path:
run_from_file_paths(model_args, args)
else:
return run_from_layers(model_args, args[0])
def setup(opts):
sess = tf.Session()
init_op = tf.global_variables_initializer()
sess.run(init_op)
with tf.name_scope('placeholder'):
input_photo = tf.placeholder(dtype=tf.float32,
shape=[1, None, None, 3],
name='photo')
input_photo_features = encoder(image=input_photo,
options={'gf_dim': 32},
reuse=False)
output_photo = decoder(features=input_photo_features,
options={'gf_dim': 32},
reuse=False)
saver = tf.train.Saver()
path = opts['styleCheckpoint']
model_name = [
p for p in os.listdir(path) if os.path.isdir(os.path.join(path, p))
][0]
checkpoint_dir = os.path.join(path, model_name, 'checkpoint_long')
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
ckpt_name = os.path.basename(ckpt.model_checkpoint_path)
saver.restore(sess, os.path.join(checkpoint_dir, ckpt_name))
return dict(sess=sess, input_photo=input_photo, output_photo=output_photo)
def __init__(self, input_height=256, input_width=256, batch_size=1,
bar_model_name=None, bar_checkpoint_name=None, mosaic_model_name=None,
mosaic_checkpoint_name=None, is_mosaic=False):
self.bar_model_name = bar_model_name
self.bar_checkpoint_name = bar_checkpoint_name
self.mosaic_model_name = mosaic_model_name
self.mosaic_checkpoint_name = mosaic_checkpoint_name
self.is_mosaic = is_mosaic
self.input_height = input_height
self.input_width = input_width
self.batch_size = batch_size
self.check_model_file()
self.build_model()
self.discriminator = mm.discriminator_red(tf.keras.Input(batch_size,
input_height,
input_width,
3)
)
self.encoder = mm.encoder(tf.keras.Input(batch_size,
input_height,
input_width,
6))
def build_model(self):
# ------- variables
tf.compat.v1.disable_eager_execution()
self.X = tf.compat.v1.placeholder(
tf.float32,
[self.batch_size, self.input_height, self.input_width, 3])
self.Y = tf.compat.v1.placeholder(
tf.float32,
[self.batch_size, self.input_height, self.input_width, 3])
self.MASK = tf.compat.v1.placeholder(
tf.float32,
[self.batch_size, self.input_height, self.input_width, 3])
IT = tf.compat.v1.placeholder(tf.float32)
# ------- structure
input = tf.concat([self.X, self.MASK], 3)
vec_en = mm.encoder(input, reuse=False, name='G_en')
vec_con = mm.contextual_block(vec_en,
vec_en,
self.MASK,
3,
50.0,
'CB1',
stride=1)
I_co = mm.decoder(vec_en,
self.input_height,
self.input_height,
reuse=False,
name='G_de')
I_ge = mm.decoder(vec_con,
self.input_height,
self.input_height,
reuse=True,
name='G_de')
self.image_result = I_ge * (1 - self.MASK) + self.Y * self.MASK
D_real_red = mm.discriminator_red(self.Y, reuse=False, name='disc_red')
D_fake_red = mm.discriminator_red(self.image_result,
reuse=True,
name='disc_red')
# ------- Loss
Loss_D_red = tf.reduce_mean(
input_tensor=tf.nn.relu(1 + D_fake_red)) + tf.reduce_mean(
input_tensor=tf.nn.relu(1 - D_real_red))
Loss_D = Loss_D_red
Loss_gan_red = -tf.reduce_mean(input_tensor=D_fake_red)
Loss_gan = Loss_gan_red
Loss_s_re = tf.reduce_mean(input_tensor=tf.abs(I_ge - self.Y))
Loss_hat = tf.reduce_mean(input_tensor=tf.abs(I_co - self.Y))
A = tf.image.rgb_to_yuv((self.image_result + 1) / 2.0)
A_Y = tf.cast(A[:, :, :, 0:1] * 255.0, dtype=tf.int32)
B = tf.image.rgb_to_yuv((self.Y + 1) / 2.0)
B_Y = tf.cast(B[:, :, :, 0:1] * 255.0, dtype=tf.int32)
ssim = tf.reduce_mean(input_tensor=tf.image.ssim(A_Y, B_Y, 255.0))
alpha = IT / 1000000
Loss_G = 0.1 * Loss_gan + 10 * Loss_s_re + 5 * (1 - alpha) * Loss_hat
# --------------------- variable & optimizer
var_D = [
v for v in tf.compat.v1.global_variables()
if v.name.startswith('disc_red')
]
var_G = [
v for v in tf.compat.v1.global_variables()
if v.name.startswith('G_en') or v.name.startswith('G_de')
or v.name.startswith('CB1')
]
update_ops = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
optimize_D = tf.compat.v1.train.AdamOptimizer(learning_rate=0.0004,
beta1=0.5,
beta2=0.9).minimize(
Loss_D,
var_list=var_D)
optimize_G = tf.compat.v1.train.AdamOptimizer(learning_rate=0.0001,
beta1=0.5,
beta2=0.9).minimize(
Loss_G,
var_list=var_G)
config = tf.compat.v1.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.4
# config.gpu_options.allow_growth = False
self.sess = tf.compat.v1.Session(config=config)
init = tf.compat.v1.global_variables_initializer()
self.sess.run(init)
saver = tf.compat.v1.train.Saver()
if self.is_mosaic:
Restore = tf.compat.v1.train.import_meta_graph(
self.mosaic_model_name)
Restore.restore(
self.sess,
tf.train.latest_checkpoint(self.mosaic_checkpoint_name))
else:
Restore = tf.compat.v1.train.import_meta_graph(self.bar_model_name)
Restore.restore(
self.sess,
tf.train.latest_checkpoint(self.bar_checkpoint_name))
saving_iter = 10000
Max_iter = 1000000
# ------- variables
X = tf.placeholder(tf.float32, [BATCH_SIZE, HEIGHT, WIDTH, 3])
Y = tf.placeholder(tf.float32, [BATCH_SIZE, HEIGHT, WIDTH, 3])
MASK = tf.placeholder(tf.float32, [BATCH_SIZE, HEIGHT, WIDTH, 3])
IT = tf.placeholder(tf.float32)
# ------- structure
input = tf.concat([X, MASK], 3)
vec_en = mm.encoder(input, reuse=False)
vec_con = mm.contextual_block(vec_en, vec_en, MASK)
I_co = mm.decoder(vec_en, reuse=False)
I_ge = mm.decoder(vec_con, reuse=True)
image_result = I_ge * (1 - MASK) + Y * MASK
D_real_red = mm.discriminator_red(Y, reuse=False)
D_fake_red = mm.discriminator_red(image_result, reuse=True)
# ------- Loss
Loss_D_red = tf.reduce_mean(tf.nn.relu(1 + D_fake_red)) + tf.reduce_mean(
tf.nn.relu(1 - D_real_red))
saving_iter = 10000
Max_iter = 1000000
# ------- variables
X = tf.placeholder(tf.float32, [batch_size, Height, Width, 3])
Y = tf.placeholder(tf.float32, [batch_size, Height, Width, 3])
MASK = tf.placeholder(tf.float32, [batch_size, Height, Width, 3])
IT = tf.placeholder(tf.float32)
# ------- structure
input = tf.concat([X, MASK], 3)
vec_en = mm.encoder(input, reuse=False, name='G_en')
vec_con = mm.contextual_block(vec_en, vec_en, MASK, 3, 50.0, 'CB1', stride=1)
I_co = mm.decoder(vec_en, Height, reuse=False, name='G_de')
I_ge = mm.decoder(vec_con, Height, reuse=True, name='G_de')
image_result = I_ge * (1 - MASK) + Y * MASK
D_real_red = mm.discriminator_red(Y, reuse=False, name='disc_red')
D_fake_red = mm.discriminator_red(image_result, reuse=True, name='disc_red')
# ------- Loss
Loss_D_red = tf.reduce_mean(tf.nn.relu(1 + D_fake_red)) + tf.reduce_mean(
tf.nn.relu(1 - D_real_red))