def __build_graph(self):
tf.set_random_seed(SEED)
np.random.seed(SEED)
self.is_training = tf.placeholder(tf.bool)
self.x1 = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
self.x2 = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
# Normalize + reshape 'real' input data
norm_x1 = 2*(tf.cast(self.x1, tf.float32)-.5)
norm_x2 = 2 * (tf.cast(self.x2, tf.float32) - .5)
# norm_img_black=2*(tf.cast(self.img_black, tf.float32)-.5)
# Set Encoder and Decoder archs
self.Encoder, self.Decoder,self.Classifier,self.gan_discriminator = NetsRetreiverWithClassifier(self.arch)
# Encode
self.z1 = self.__Enc(norm_x1)
# original stage
# Decode
self.x_out1 = self.__Dec(self.z1)
self.z2=self.__Enc(norm_x2)
z1_part1,z1_part2=tf.split(self.z1,2,axis=1)
z2_part1, z2_part2 = tf.split(self.z2, 2, axis=1)
x2fg_x1bg=tf.concat([z2_part1,z1_part2],axis=1)
x1fg_x2bg=tf.concat([z1_part1,z2_part2],axis=1)
self.x2fg_x1bg_out=self.__Dec(x2fg_x1bg)
self.x1fg_x2bg_out = self.__Dec(x1fg_x2bg)
# Loss and optimizer
self.__prep_loss_optimizer(norm_x1)
def __build_graph(self):
tf.set_random_seed(SEED)
np.random.seed(SEED)
self.is_training = tf.placeholder(tf.bool)
self.x1 = tf.placeholder(tf.float32,
shape=[None] + list(self.image_shape))
self.gt0 = tf.placeholder(tf.float32,
shape=[None] +
list([self.latent_num + 9])) # label
# Normalize + reshape 'real' input data
norm_x1 = 2 * (tf.cast(self.x1, tf.float32) - .5)
# Set Encoder and Decoder archs
self.Encoder, self.Decoder, self.Classifier, self.gan_discriminator = NetsRetreiverWithClassifier(
self.arch)
# Encode
self.z1 = self.__Enc(norm_x1)
# original stage
# Decode
self.x_out1 = self.__Dec(self.z1)
# split latent representation into 2 part and classification
r_part1, r_part2 = tf.split(self.z1, 2, axis=1)
c_p0 = self.__Classifier(r_part1)
c_p1 = self.__Classifier(r_part2)
print("c_p1.shape")
print(c_p1.shape)
# Loss and optimizer
self.__prep_loss_optimizer(norm_x1, c_p0, c_p1)
def __build_graph(self):
tf.set_random_seed(SEED)
np.random.seed(SEED)
self.is_training = tf.placeholder(tf.bool)
self.x1 = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
self.label=tf.placeholder(tf.float32, shape=[None] + list([self.latent_num*(self.latent_num+1)]))
self.mask= tf.placeholder(tf.float32, shape=[None] + list([self.latent_dim]))
self.mask_zero=tf.placeholder(tf.float32, shape=[None] + list([self.latent_dim]))
self.img_black=tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
self.gt0=tf.placeholder(tf.float32, shape=[None] + list([self.latent_num+1]))
gts0=self.gt0
# classification gt
#self.vec_one=tf.placeholder(tf.float32, shape=[None] + list([1])) # which used in the classification loss
#self.class_gt0=tf.placeholder(tf.float32, shape=[None] + list([self.latent_num+1]))
# Normalize + reshape 'real' input data
norm_x1 = 2*(tf.cast(self.x1, tf.float32)-.5)
norm_img_black=2*(tf.cast(self.img_black, tf.float32)-.5)
# Set Encoder and Decoder archs
self.Encoder, self.Decoder,self.Classifier,self.gan_discriminator = NetsRetreiverWithClassifier(self.arch)
# Encode
self.z1 = self.__Enc(norm_x1)
# original stage
# Decode
self.x_out1 = self.__Dec(self.z1)
# random set 0
self.r1=tf.multiply(self.z1,self.mask)
self.x_out_r0 = self.__Dec(self.r1)
# set representation all 000
self.r_zero=tf.multiply(self.z1,self.mask_zero)
self.img_out_zero=self.__Dec(self.r_zero)
# split latent representation into 3 part and classification
r_all0,r_all0,r_all0=tf.split(self.r_zero,3,axis=1)
r_part1,r_part2,r_part3=tf.split(self.z1,3,axis=1)
gts_part1,gts_part2,gts_part3=tf.split(self.label,3,axis=1)
c_p0=self.__Classifier(r_all0)
c_p1=self.__Classifier(r_part1)
c_p2=self.__Classifier(r_part2)
c_p3=self.__Classifier(r_part3)
print("gts_part1.shape")
print(gts_part1.shape)
print(c_p1.shape)
# Loss and optimizer
self.__prep_loss_optimizer(norm_x1,norm_img_black,c_p0,c_p1,c_p2,c_p3,gts0,gts_part1,gts_part2,gts_part3)
def __build_graph(self):
tf.set_random_seed(SEED)
np.random.seed(SEED)
self.is_training = tf.placeholder(tf.bool)
self.x1 = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
# Normalize + reshape 'real' input data
norm_x1 = 2*(tf.cast(self.x1, tf.float32)-.5)
# norm_img_black=2*(tf.cast(self.img_black, tf.float32)-.5)
# Set Encoder and Decoder archs
self.Encoder, self.Decoder,self.Classifier,self.gan_discriminator = NetsRetreiverWithClassifier(self.arch)
# Encode
self.z1 = self.__Enc(norm_x1)
# original stage
# Decode
self.x_out1 = self.__Dec(self.z1)
# Loss and optimizer
self.__prep_loss_optimizer(norm_x1)
def __build_graph(self):
tf.set_random_seed(SEED)
np.random.seed(SEED)
self.is_training = tf.placeholder(tf.bool)
self.x1 = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
self.aux1_mask = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
# auxilary dataset
self.aux1 = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
self.aux2 = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
self.aux_GT1 = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
self.aux_GT2 = tf.placeholder(tf.float32, shape=[None] + list(self.image_shape))
self.class_gt0 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num+9]))
self.class_gt1 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num+9]))
self.class_gt2 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num+9]))
self.class_gt3 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num+9]))
self.class_gt4 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num + 9]))
self.class_gt5 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num + 9]))
self.class_gt6 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num + 9]))
self.class_gt7 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num + 9]))
self.class_gt8 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num + 9]))
self.class_gt9 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num + 9]))
self.class_gt10 = tf.placeholder(tf.float32, shape=[None] + list([self.latent_num + 9]))
# onesample labels
self.aux_class_gt=tf.placeholder(tf.float32,shape=[None]+list([self.latent_num+9]))
# Normalize + reshape 'real' input data
norm_x1 = 2*(tf.cast(self.x1, tf.float32)-.5)
norm_aux1 = 2*(tf.cast(self.aux1, tf.float32)-.5)
norm_aux2 = 2*(tf.cast(self.aux2, tf.float32)-.5)
norm_aux_GT1 = 2*(tf.cast(self.aux_GT1, tf.float32)-.5)
norm_aux_GT2 = 2*(tf.cast(self.aux_GT2, tf.float32)-.5)
# Set Encoder and Decoder archs
self.Encoder, self.Decoder,self.Classifier,self.gan_discriminator = NetsRetreiverWithClassifier(self.arch)
# Encode and decode
self.z1 = self.__Enc(norm_x1)
self.x1_out = self.__Dec(self.z1)
# aux data
self.aux_z1 = self.__Enc(norm_aux1)
self.aux1_out = self.__Dec(self.aux_z1)
self.aux_z2 = self.__Enc(norm_aux2)
self.aux2_out = self.__Dec(self.aux_z2)
aux1_head,aux1_bg=tf.split(self.aux_z1,2,axis=1)
aux2_head,aux2_bg=tf.split(self.aux_z2,2,axis=1)
GT1_z=tf.concat([aux2_head,aux1_bg],axis=1)
GT2_z=tf.concat([aux1_head,aux2_bg],axis=1)
self.GT1_out = self.__Dec(GT1_z)
self.GT2_out = self.__Dec(GT2_z)
#dual swap
x1_head,x1_bg=tf.split(self.z1,2,axis=1)
self.mix_head_out=self.__Dec(tf.concat([aux1_head,x1_bg],axis=1))
mix_head,mix_bg=tf.split(self.__Enc(self.mix_head_out),2,axis=1)
x1_dual_out=self.__Dec(tf.concat([x1_head,mix_bg],axis=1))
self.aux1_mix_head_out=self.__Dec(tf.concat([x1_head,aux1_bg],axis=1))
# classification loss
## for x1
r_part1, r_part2 = tf.split(self.z1, 2, axis=1)
c_p0 = self.__Classifier(r_part1)
c_p1 = self.__Classifier(r_part2)
## for aux1
aux1_r_part1, aux1_r_part2 = tf.split(self.aux_z1, 2, axis=1)
aux1_c_p0 = self.__Classifier(aux1_r_part1)
aux1_c_p1 = self.__Classifier(aux1_r_part2)
## for aux2
aux2_r_part1, aux2_r_part2 = tf.split(self.aux_z2, 2, axis=1)
aux2_c_p0 = self.__Classifier(aux2_r_part1)
aux2_c_p1 = self.__Classifier(aux2_r_part2)
# Loss and optimizer
self.__prep_loss_optimizer(norm_x1,norm_aux1,norm_aux2,norm_aux_GT1,norm_aux_GT2,x1_dual_out,c_p0,c_p1,aux1_c_p0,aux1_c_p1,aux2_c_p0,aux2_c_p1)
def __build_graph(self):
tf.set_random_seed(SEED)
np.random.seed(SEED)
self.is_training = tf.placeholder(tf.bool)
self.x1 = tf.placeholder(tf.float32,
shape=[None] + list(self.image_shape))
self.x_gan = tf.placeholder(tf.float32,
shape=[None] + list(self.image_shape))
self.mask = tf.placeholder(tf.float32,
shape=[None] + list([self.latent_dim]))
self.mask_zero = tf.placeholder(tf.float32,
shape=[None] + list([self.latent_dim]))
self.img_white = tf.placeholder(tf.float32,
shape=[None] + list(self.image_shape))
# classification gt
self.vec_one = tf.placeholder(
tf.float32,
shape=[None] + list([1])) # which used in the classification loss
self.class_gt0 = tf.placeholder(tf.float32,
shape=[None] +
list([self.latent_num + 1]))
self.class_gt1 = tf.placeholder(tf.float32,
shape=[None] +
list([self.latent_num + 1]))
self.class_gt2 = tf.placeholder(tf.float32,
shape=[None] +
list([self.latent_num + 1]))
# Normalize + reshape 'real' input data
norm_x1 = 2 * (tf.cast(self.x1, tf.float32) - .5)
norm_img_white = 2 * (tf.cast(self.img_white, tf.float32) - .5)
norm_x_gan = 2 * (tf.cast(self.x_gan, tf.float32) - .5)
# Set Encoder and Decoder archs
#self.Encoder, self.Decoder = NetsRetreiver(self.arch)
#self.Encoder, self.Decoder,self.Classifier = NetsRetreiverWithClassifier(self.arch)
self.Encoder, self.Decoder, self.Classifier, self.gan_discriminator = NetsRetreiverWithClassifier(
self.arch)
# Encode
self.z1 = self.__Enc(norm_x1)
# original stage
# Decode
self.x_out1 = self.__Dec(self.z1)
# random set 0
self.r1 = tf.multiply(self.z1, self.mask)
self.x_out_r0 = self.__Dec(self.r1)
r11 = self.__Enc(self.x_out_r0)
# set representation all 000
self.r_zero = tf.multiply(self.z1, self.mask_zero)
self.img_out_zero = self.__Dec(self.r_zero)
# split latent representation into 2 part and classification
r_part1, r_part2 = tf.split(self.r1, 2, axis=1)
r_all0, r_all0 = tf.split(self.r_zero, 2, axis=1)
c_p0 = self.__Classifier(r_all0)
c_p1 = self.__Classifier(r_part1)
c_p2 = self.__Classifier(r_part2)
# GAN Discriminator
self.fake_data = self.x_out_r0
self.real_data = tf.reshape(norm_x_gan, [
-1, self.image_shape[0] * self.image_shape[1] * self.image_shape[2]
])
#self.real_data=norm_x1
disc_real = self.__GAN_discriminator(
self.real_data) # input real image data
disc_fake = self.__GAN_discriminator(
self.fake_data) # input fake image data
# Loss and optimizer
self.__prep_loss_optimizer(norm_x1, norm_img_white, r11, c_p0, c_p1,
c_p2, disc_real, disc_fake)