def create_generators(self):
cnf = Config()
fn_prefix = './file_names/'
x_trains_path = fn_prefix + 'annotations_fns.npy'
train_filenames = self.create_annotation_name(
annotation_path=cnf.annotation_path)
save(x_trains_path, train_filenames)
return train_filenames
def get_discriminator_model(self):
cnf = Config()
inputs = tf.keras.Input(shape=(cnf.num_of_landmarks))
x = Dense(cnf.num_of_landmarks)(inputs)
x = BatchNormalization()(x)
x_1 = ReLU()(x)
x = Dense(cnf.num_of_landmarks)(x_1)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Add()([x, x_1])
x = Dense(128)(x)
x = BatchNormalization()(x)
x_1 = ReLU()(x)
x = Dense(128)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Add()([x, x_1])
x = Dense(256)(x)
x = BatchNormalization()(x)
x_2 = ReLU()(x)
x = Dense(256)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Add()([x, x_2])
x = Dense(256)(x)
x = BatchNormalization()(x)
x_2 = ReLU()(x)
x = Dense(256)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Add()([x, x_2])
# outputs = Dense(1, activation='softmax')(x)
# outputs = Dense(1, activation='sigmoid')(x)
outputs = Dense(1)(x)
model = tf.keras.Model(inputs=inputs,
outputs=outputs,
name="disc_model")
model.summary()
model_json = model.to_json()
with open("./model_arch/Disc_model.json", "w") as json_file:
json_file.write(model_json)
return model
def test(self):
cnf = Config()
'''making models'''
net_model = NetworkModels()
model_gen = net_model.get_generator_model()
model_disc = net_model.get_discriminator_model()
model_disc.load_weights('./models/last_we_model_disc_.h5')
# model_disc = tf.keras.models.load_model('./models/model_disc1999_.h5')
'''noise'''
test_sample = tf.random.normal([9, cnf.num_of_landmarks])
out_fake = model_disc(test_sample)
print('------------')
def get_batch_sample(self, batch_index, x_train_filenames):
cnf = Config()
pn_tr_path = cnf.annotation_path
batch_x = x_train_filenames[batch_index *
cnf.batch_size:(batch_index + 1) *
cnf.batch_size]
# pn_batch = np.array([self._load_and_create_hm(pn_tr_path + file_name) for file_name in batch_x])
pn_batch = np.array([
self._load_and_normalize(pn_tr_path + file_name)
for file_name in batch_x
])
pn_batch = tf.cast(pn_batch, tf.float32)
return pn_batch
def create_landmarks(self, normal_lnd):
cnf = Config()
normal_lnd = np.array(normal_lnd)
# landmarks_splited = _landmarks.split(';')
landmark_arr_x = []
landmark_arr_y = []
for j in range(0, len(normal_lnd), 2):
x = float(normal_lnd[j] *
cnf.image_input_size) + cnf.image_input_size / 2.0
y = float(normal_lnd[j + 1] *
cnf.image_input_size) + cnf.image_input_size / 2.0
landmark_arr_x.append(x)
landmark_arr_y.append(y)
return landmark_arr_x, landmark_arr_y
def _load_and_normalize(self, point_path):
cnf = Config()
annotation = load(point_path)
'''test print '''
# self.test_image_print(img_name='1', img=np.zeros([224,224,3]), landmarks=annotation)
'''normalize landmarks based on hyperface method'''
width = 1 # cnf.image_input_size
height = 1 # cnf.image_input_size
x_center = width / 2
y_center = height / 2
annotation_norm = []
for p in range(0, len(annotation), 2):
annotation_norm.append((x_center - annotation[p]) / width)
annotation_norm.append((y_center - annotation[p + 1]) / height)
'''denormalize for test'''
# landmarks_x, landmarks_y = self.create_landmarks(annotation_norm)
# plt.figure()
# plt.scatter(x=landmarks_x[:], y=landmarks_y[:], c='#000000', s=15)
# plt.savefig('11.png')
''''''
return annotation_norm
def train(self, train_gen, train_disc):
""""""
cnf = Config()
'''create loss obj'''
c_loss = CustomLosses()
'''create summary writer'''
summary_writer = tf.summary.create_file_writer(
"./train_logs/fit/" + datetime.now().strftime("%Y%m%d-%H%M%S"))
'''making models'''
net_model = NetworkModels()
model_gen = net_model.get_generator_model()
model_gen.trainable = train_gen
# model_gen.load_weights('./models/last_we_model_gen39_.h5')
model_disc = net_model.get_discriminator_model()
model_disc.trainable = train_disc
# model_disc.load_weights('./models/last_we_model_disc_.h5')
'''optimizer'''
opti_gen = tf.keras.optimizers.Adam(lr=1e-4)
opti_disc = tf.keras.optimizers.Adam(lr=1e-8, beta_1=0.5)
'''create sample generator'''
dhp = DataHelper()
x_train_filenames = dhp.create_generators()
'''create train configuration'''
step_per_epoch = len(x_train_filenames) // cnf.batch_size
# step_per_epoch = 5
'''start train process'''
train_gen = True
train_disc = True
model_gen.trainable = train_gen
model_disc.trainable = train_disc
for epoch in range(cnf.epochs):
# if epoch < 500 and (epoch + 1) % 30 == 0:
# train_gen = not train_gen
# train_disc = not train_disc
#
# model_gen.trainable = train_gen
# model_disc.trainable = train_disc
#
# print('=================================')
# print(' Generator is :' + str(train_gen))
# print(' Discriminator is :' + str(train_disc))
# print('=================================')
# else:
# model_gen.trainable = True
# model_disc.trainable = True
#
# print('=================================')
# print(' Generator is :' + str(train_gen))
# print(' Discriminator is :' + str(train_disc))
# print('=================================')
for batch_index in range(step_per_epoch):
'''creating noises'''
noise = tf.random.normal(
[cnf.batch_size, cnf.noise_input_size])
'''load annotation and images'''
real_data = dhp.get_batch_sample(
batch_index=batch_index,
x_train_filenames=x_train_filenames)
''''''
self.train_step(epoch=epoch,
step=batch_index,
real_data=real_data,
model_gen=model_gen,
model_disc=model_disc,
opti_gen=opti_gen,
opti_disc=opti_disc,
cnf=cnf,
c_loss=c_loss,
noise=noise)
'''save sample images:'''
if (epoch + 1) % 10 == 0:
test_sample = tf.random.normal([9, cnf.noise_input_size])
self.save_sample_images(model=model_gen,
epoch=epoch,
test_input=test_sample,
dhp=dhp)
'''save weights'''
if (epoch + 1) % 1000 == 0:
model_gen.save('./models/model_gen' + str(epoch) + '_.h5')
model_disc.save('./models/model_disc' + str(epoch) + '_.h5')
'''save last weights'''
model_gen.save('./models/model_gen_LAST.h5')
model_disc.save('./models/model_disc_LAST.h5')
def get_generator_model(self):
cnf = Config()
inputs = tf.keras.Input(shape=(cnf.noise_input_size, ))
x = Dense(cnf.noise_input_size)(inputs)
x_1 = BatchNormalization()(x)
x = Dense(cnf.noise_input_size)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(cnf.noise_input_size)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(cnf.noise_input_size)(x)
x = BatchNormalization()(x)
x = Add()([x, x_1])
x = ReLU()(x)
#
x = Dense(128)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(128)(x)
x_1 = BatchNormalization()(x)
x = Dense(128)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(128)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(128)(x)
x = BatchNormalization()(x)
x = Add()([x, x_1])
x = ReLU()(x)
#
x = Dense(128)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(128)(x)
x_1 = BatchNormalization()(x)
x = Dense(128)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(128)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(128)(x)
x = BatchNormalization()(x)
x = Add()([x, x_1])
x = ReLU()(x)
####
x = Dense(256)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(256)(x)
x_2 = BatchNormalization()(x)
x = Dense(256)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(256)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(256)(x)
x = BatchNormalization()(x)
x = Add()([x, x_2])
x = ReLU()(x)
##
x = Dense(256)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(256)(x)
x_2 = BatchNormalization()(x)
x = Dense(256)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(256)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Dense(256)(x)
x = BatchNormalization()(x)
x = Add()([x, x_2])
x = ReLU()(x)
#######################
x = Dropout(.2)(x)
outputs = Dense(cnf.num_of_landmarks, activation='linear')(x)
model = tf.keras.Model(inputs=inputs,
outputs=outputs,
name="gen_model")
model.summary()
model_json = model.to_json()
with open("./model_arch/Gen_model.json", "w") as json_file:
json_file.write(model_json)
return model