def create_generate_model(self, b_size):
""" create samping model in fig 1 and 3 in the paper """
self.set_is_training(False)
with tf.variable_scope('AE', reuse=tf.AUTO_REUSE):
self._create_generate_input()
# if self.z is not fed in, just sample from diagonal Gaussian
self.z = ops.tf_sample_standard_diag_guassian(b_size, self.n_code)
decoder_in = self.z
self.layers['generate'] = (self.decoder(decoder_in) + 1.) / 2.
def create_generate_style_model(self, n_sample):
""" create samping model in fig 6 in the paper """
self.set_is_training(False)
with tf.variable_scope('AE', reuse=tf.AUTO_REUSE):
self._create_generate_input()
label = []
for i in range(self.n_class):
label.extend([i for k in range(n_sample)])
label = tf.convert_to_tensor(label) # [n_class]
one_hot_label = tf.one_hot(label, self.n_class) # [n_class*n_sample, n_class]
z = ops.tf_sample_standard_diag_guassian(n_sample, self.n_code)
z = tf.tile(z, [self.n_class, 1]) # [n_class*n_sample, n_code]
decoder_in = tf.concat((z, one_hot_label), axis=-1)
self.layers['generate'] = (self.decoder(decoder_in) + 1. ) / 2.
def create_generate_model(self, b_size):
self.set_is_training(False)
self._create_generate_input()
self.z = ops.tf_sample_standard_diag_guassian(b_size, self._n_code)
self.layers['generate'] = tf.nn.sigmoid(self.decoder(self.z))