def generator(inp):
global reuse_gen, bn_training
with tf.variable_scope('generator', reuse=reuse_gen):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.reshape([-1, 2, 2, 512])
mod.deconvLayer(3,
256,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #8
mod.deconvLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #16
mod.SelfAttention(32)
mod.deconvLayer(5,
64,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #32
mod.deconvLayer(5,
32,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #64
mod.deconvLayer(5,
16,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #128
mod.deconvLayer(5, 3, activation=M.PARAM_TANH) #output
reuse_gen = True
return mod.get_current_layer()
def feat_encoder(inp):
global reuse_enc, bn_training
with tf.variable_scope('encoder', reuse=reuse_enc):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.convLayer(7,
16,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #128
mod.maxpoolLayer(3, stride=2) # 64
mod.convLayer(5,
64,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) # 32
block(mod, 256, 2) # 16
block(mod, 256, 1)
block(mod, 256, 1)
mod.SelfAttention(32)
block(mod, 256, 2) # 8
block(mod, 256, 1)
block(mod, 256, 1)
mod.SelfAttention(32)
block(mod, 512, 2) #4
block(mod, 512, 1)
block(mod, 512, 1)
reuse_enc = True
return mod.get_current_layer()
def discriminator_feature(inp):
global reuse_dis2, bn_training
with tf.variable_scope('discriminator_feature', reuse=reuse_dis2):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.fcLayer(256, activation=M.PARAM_LRELU, batch_norm=True)
mod.fcLayer(64, activation=M.PARAM_LRELU, batch_norm=True)
mod.fcLayer(1)
reuse_dis2 = True
return mod.get_current_layer()
def age_classify(inp, age_size):
global reuse_agecls, bn_training
with tf.variable_scope('age_cls', reuse=reuse_agecls):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.flatten()
mod.fcLayer(512, activation=M.PARAM_LRELU)
mod.fcLayer(256, activation=M.PARAM_LRELU)
mod.fcLayer(age_size)
reuse_agecls = True
return mod.get_current_layer()
def discriminator(inp):
global reuse_dis, bn_training, blknum
blknum = 0
with tf.variable_scope('discriminator', reuse=reuse_dis):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.convLayer(7,
16,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) # 64
mod.convLayer(5,
32,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) # 32
mod.SelfAttention(4)
feat = mod.convLayer(5, 64, stride=2, activation=M.PARAM_LRELU) # 16
mod.batch_norm()
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) # 8
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True)
adv = mod.convLayer(3, 1)
mod.set_current_layer(feat)
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True)
block(mod, 128, 1) # 8
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True)
block(mod, 256, 1) # 4
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True)
block(mod, 256, 1) # 2
mod.flatten()
age = mod.fcLayer(1)
reuse_dis = True
return adv, age
def age_encoder(inp, ind):
global reuse_age_enc
name = 'decoder' + str(ind)
if not name in reuse_age_enc:
reuse = False
else:
reuse = True
with tf.variable_scope(name, reuse=reuse):
mod = Model(inp)
mod.fcLayer(2 * 2 * 512, activation=M.PARAM_RELU)
mod.SelfAttention(is_fc=True, residual=True)
reuse_age_enc[name] = 1
return mod.get_current_layer()
def discriminator_f(inp, id_num):
global reuse_dis_f, bn_training
with tf.variable_scope('dis_f', reuse=reuse_dis_f):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.flatten()
feat = mod.get_current_layer()
mod.fcLayer(512, activation=M.PARAM_LRELU, batch_norm=True)
mod.fcLayer(256, activation=M.PARAM_LRELU, batch_norm=True)
adv = mod.fcLayer(1)
mod.set_current_layer(feat)
ip = mod.fcLayer(id_num)
reuse_dis_f = True
return adv, ip
def attention_blk(features):
global reuse_att
with tf.variable_scope('attention_blk', reuse=reuse_att):
# get q0
f_dim = features.get_shape().as_list()[-1]
q0 = tf.get_variable('q0', [1, f_dim],
initializer=tf.random_normal_initializer(),
dtype=tf.float32)
BSIZE = tf.shape(features[0])
q0 = tf.tile(q0, [BSIZE, 1])
mod = Model(q0)
mod.QAttention(features)
mod.fcLayer(f_dim, activation=M.PARAM_TANH)
mod.QAttention(features)
reuse_att = True
return mod.get_current_layer()
def generator_att(inp):
global reuse_genatt, bn_training, blknum
blknum = 0
with tf.variable_scope('gen_att', reuse=reuse_genatt):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.convLayer(5,
32,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #64
block(mod, 64, 1)
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #32
block(mod, 128, 1)
# block(mod,256,1)
# block(mod,256,1)
block(mod, 256, 1)
mod.SelfAttention(64, residual=True)
# block(mod,512,1)
block(mod, 256, 1)
# block(mod,256,1)
block(mod, 128, 1)
mod.deconvLayer(3,
64,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) # 64
block(mod, 64, 1)
feat = mod.deconvLayer(5,
64,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #128
A = mod.convLayer(5, 1, activation=M.PARAM_SIGMOID) #output_attention
mod.set_current_layer(feat)
C = mod.convLayer(5, 3, activation=M.PARAM_TANH)
reuse_genatt = True
return A, C
def generator_att(inp):
global reuse_genatt, bn_training
with tf.variable_scope('gen_att', reuse=reuse_genatt):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.deconvLayer(3,
512,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #4
mod.deconvLayer(3,
256,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #8
mod.deconvLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #16
mod.SelfAttention(32)
mod.deconvLayer(5,
64,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #32
mod.deconvLayer(5,
32,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #64
feat = mod.deconvLayer(5,
16,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #128
A = mod.convLayer(5, 3, activation=M.PARAM_SIGMOID) #output_attention
mod.set_current_layer(feat)
C = mod.convLayer(5, 3, activation=M.PARAM_TANH)
reuse_genatt = True
return A, C