def __init__(self, height, width, channel, num_class, \
ksize, radix=4, kpaths=4):
super(Neuralnet, self).__init__()
self.height, self.width, self.channel, self.num_class = height, width, channel, num_class
self.ksize, self.radix, self.kpaths = ksize, radix, kpaths
self.customlayers = lay.Layers()
self.forward = tf.function(self.__call__)
def __init__(self, \
height, width, channel, ksize, zdim, \
learning_rate=1e-3, path='', verbose=True):
print("\nInitializing Neural Network...")
self.height, self.width, self.channel, self.ksize, self.zdim = \
height, width, channel, ksize, zdim
self.learning_rate = learning_rate
self.path_ckpt = path
self.x = tf.compat.v1.placeholder(tf.float32, [None, self.height, self.width, self.channel], \
name="x")
self.z = tf.compat.v1.placeholder(tf.float32, [None, self.zdim], \
name="z")
self.batch_size = tf.compat.v1.placeholder(tf.int32, shape=[], \
name="batch_size")
self.training = tf.compat.v1.placeholder(tf.bool, shape=[], \
name="training")
self.layer = lay.Layers()
self.variables, self.losses = {}, {}
self.__build_model(x_real=self.x,
z=self.z,
ksize=self.ksize,
verbose=verbose)
self.__build_loss()
with tf.control_dependencies(self.variables['ops_d']):
self.optimizer_d = tf.compat.v1.train.AdamOptimizer( \
self.learning_rate/5, name='Adam_d').minimize(\
self.losses['loss_d'], var_list=self.variables['params_d'])
with tf.control_dependencies(self.variables['ops_g']):
self.optimizer_g = tf.compat.v1.train.AdamOptimizer( \
self.learning_rate, name='Adam_g').minimize(\
self.losses['loss_g'], var_list=self.variables['params_g'])
with tf.control_dependencies(self.variables['ops_e']):
self.optimizer_e = tf.compat.v1.train.AdamOptimizer( \
self.learning_rate, name='Adam_e').minimize(\
self.losses['loss_e'], var_list=self.variables['params_e'])
tf.compat.v1.summary.scalar('f-AnoGAN/mean_real',
self.losses['mean_real'])
tf.compat.v1.summary.scalar('f-AnoGAN/mean_fake',
self.losses['mean_fake'])
tf.compat.v1.summary.scalar('f-AnoGAN/mean_izi', self.losses['izi'])
tf.compat.v1.summary.scalar('f-AnoGAN/mean_ziz', self.losses['ziz'])
tf.compat.v1.summary.scalar('f-AnoGAN/loss_d', self.losses['loss_d'])
tf.compat.v1.summary.scalar('f-AnoGAN/loss_g', self.losses['loss_g'])
tf.compat.v1.summary.scalar('f-AnoGAN/loss_e', self.losses['loss_e'])
self.summaries = tf.compat.v1.summary.merge_all()
self.__init_session(path=self.path_ckpt)
def __init__(self, height, width, channel, num_class, ksize, learning_rate=1e-3, ckpt_dir='./Checkpoint'):
print("\nInitializing Short-ResNet...")
self.height, self.width, self.channel, self.num_class = height, width, channel, num_class
self.ksize, self.learning_rate = ksize, learning_rate
self.ckpt_dir = ckpt_dir
self.customlayers = lay.Layers()
self.model(tf.zeros([1, self.height, self.width, self.channel]), verbose=True)
self.optimizer = tf.optimizers.Adam(self.learning_rate)
self.summary_writer = tf.summary.create_file_writer(self.ckpt_dir)
def __init__(self, \
height, width, channel, ksize, \
learning_rate=1e-3, path='', verbose=True):
print("\nInitializing Neural Network...")
self.height, self.width, self.channel, self.ksize = \
height, width, channel, ksize
self.learning_rate = learning_rate
self.path_ckpt = path
self.x = tf.compat.v1.placeholder(tf.float32, [None, self.height, self.width, self.channel], \
name="x")
self.batch_size = tf.compat.v1.placeholder(tf.int32, shape=[], \
name="batch_size")
self.training = tf.compat.v1.placeholder(tf.bool, shape=[], \
name="training")
self.layer = lay.Layers()
self.variables, self.losses = {}, {}
self.__build_model(ksize=self.ksize, verbose=verbose)
self.__build_loss()
with tf.control_dependencies(self.variables['ops_d']):
self.optimizer_d = tf.compat.v1.train.AdamOptimizer( \
self.learning_rate, name='Adam_d').minimize(\
self.losses['loss_d'], var_list=self.variables['params_d'])
with tf.control_dependencies(self.variables['ops_g']):
self.optimizer_g = tf.compat.v1.train.AdamOptimizer( \
self.learning_rate, name='Adam_g').minimize(\
self.losses['loss_g'], var_list=self.variables['params_g'])
# Lπ· = βπ β π·(π)β1 β βπΊ(π) β π·(πΊ(π))β1
tf.compat.v1.summary.scalar('ADAE/D/loss_d_term1', \
tf.compat.v1.reduce_mean(self.losses['loss_d_term1']))
tf.compat.v1.summary.scalar('ADAE/D/loss_d_term2', \
tf.compat.v1.reduce_mean(self.losses['loss_d_term2']))
tf.compat.v1.summary.scalar('ADAE/D/loss_d', self.losses['loss_d'])
# LπΊ = βπ β πΊ(π)β1+βπΊ(π) β π·(πΊ(π))β1
tf.compat.v1.summary.scalar('ADAE/G/loss_g_term1', \
tf.compat.v1.reduce_mean(self.losses['loss_g_term1']))
tf.compat.v1.summary.scalar('ADAE/G/loss_g_term2', \
tf.compat.v1.reduce_mean(self.losses['loss_g_term2']))
tf.compat.v1.summary.scalar('ADAE/G/loss_g', self.losses['loss_g'])
self.summaries = tf.compat.v1.summary.merge_all()
self.__init_session(path=self.path_ckpt)
def __init__(self, \
height, width, channel, ksize, \
w_enc=1, w_con=50, w_adv=1, \
learning_rate=1e-3, path='', verbose=True):
print("\nInitializing Neural Network...")
self.height, self.width, self.channel, self.ksize = height, width, channel, ksize
self.w_enc, self.w_con, self.w_adv = w_enc, w_con, w_adv
self.learning_rate = learning_rate
self.path_ckpt = path
self.x = tf.compat.v1.placeholder(
tf.float32, [None, self.height, self.width, self.channel])
self.batch_size = tf.compat.v1.placeholder(tf.int32, shape=[])
self.training = tf.compat.v1.placeholder(tf.bool, shape=[])
self.layer = lay.Layers()
self.conv_shapes = []
self.variables, self.losses = {}, {}
self.__build_model(x_real=self.x, ksize=self.ksize, verbose=verbose)
self.__build_loss()
with tf.control_dependencies(
tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.UPDATE_OPS)):
self.optimizer = tf.compat.v1.train.AdamOptimizer( \
self.learning_rate).minimize(self.losses['target'])
tf.compat.v1.summary.scalar('GANomaly/loss_enc',
self.losses['mean_enc'])
tf.compat.v1.summary.scalar('GANomaly/loss_con',
self.losses['mean_con'])
tf.compat.v1.summary.scalar('GANomaly/loss_adv',
self.losses['mean_adv'])
tf.compat.v1.summary.scalar('GANomaly/loss_target',
self.losses['target'])
self.summaries = tf.compat.v1.summary.merge_all()
self.__init_session(path=self.path_ckpt)
