def connect_zx(self, z):
# first concatenate
x_scrambled = keras.layers.Concatenate()(z)
# unscramble x
self.output_x = IndexLayer(self.indices_merge)(
x_scrambled) # , name='output_x'
return self.output_x
def connect_xz(self, x):
# first concatenate
z_scrambled = keras.layers.Concatenate()(x)
# unscramble x
self.output_z = IndexLayer(self.indices_merge)(
z_scrambled) # , name='output_z'
return self.output_z
def connect_zx(self, z):
# if we have reduced the dimension, we ignore the last dimensions from the z-direction.
if self.keepdims < self.dim:
z = IndexLayer(np.arange(0, self.keepdims))(z)
self.output_x = keras.layers.Lambda(lambda z: tf.matmul(z, self.Tblacken) + self.X0mean)(z)
# Jacobian
self.log_det_zx = keras.layers.Lambda(lambda z: -self.jacobian_xz * keras.backend.ones((tf.shape(z)[0], 1)))(z)
return self.output_x
def connect_zx(self, z):
self.input_z = z
z = IndexLayer(np.arange(0, self.dim-6))(z)
self.output_x = keras.layers.Lambda(lambda z: self.z2x(z)[0])(z)
self.angle_loss = keras.layers.Lambda(lambda z: self.z2x(z)[1])(z)
# self.log_det_zx = keras.layers.Lambda(lambda z: log_det_jacobian(self.z2x(z), z))(z)
self.log_det_zx = keras.layers.Lambda(lambda z: self.z2x_jacobian(z))(z)
return self.output_x
def connect_xz(self, x):
# split X into different coordinate channels
self.output_z = [
IndexLayer(isplit)(x) for isplit in self.indices_split
]
return self.output_z
def connect_zx(self, z):
self.output_x = IndexLayer(self.reverse)(z)
return self.output_x
def connect_xz(self, x):
self.output_z = IndexLayer(self.order)(x)
return self.output_z