def covariance_matrix(self, x, y):
with _tf.name_scope("Cosine_cov"):
x = self.transform.__call__(x)
y = self.transform.__call__(y)
d = _pairwise_dist(x, y)
k = _tf.cos(2.0 * _np.pi * d)
return k
def covariance_matrix(self, x, y):
with _tf.name_scope("Exponential_cov"):
x = self.transform.__call__(x)
y = self.transform.__call__(y)
d = _pairwise_dist(x, y)
k = _tf.exp(-3 * d)
return k
def covariance_matrix(self, x, y):
with _tf.name_scope("Gaussian_cov"):
x = self.transform.__call__(x)
y = self.transform.__call__(y)
d2 = _tf.pow(_pairwise_dist(x, y), 2)
k = _tf.exp(-3 * d2)
return k
def covariance_matrix(self, x, y):
with _tf.name_scope("Spherical_cov"):
x = self.transform.__call__(x)
y = self.transform.__call__(y)
d = _pairwise_dist(x, y)
k = 1 - 1.5 * d + 0.5 * _tf.pow(d, 3)
k = _tf.where(_tf.less(d, 1.0), k, _tf.zeros_like(k))
return k
def covariance_matrix(self, x, y):
with _tf.name_scope("Cubic_cov"):
x = self.transform.__call__(x)
y = self.transform.__call__(y)
d = _pairwise_dist(x, y)
k = 1 - 7 * _tf.pow(d, 2) + 35 / 4 * _tf.pow(d, 3) \
- 7 / 2 * _tf.pow(d, 5) + 3 / 4 * _tf.pow(d, 7)
k = _tf.where(_tf.less(d, 1.0), k, _tf.zeros_like(k))
return k
def covariance_matrix_d1(self, x, y, dir_y):
with _tf.name_scope("Cosine_point_dir_cov"):
x = self.transform.__call__(x)
y = self.transform.__call__(y)
dir_y = self.transform.__call__(dir_y)
sx = _tf.shape(x, name="shape_x")
x_prod = _tf.matmul(x, dir_y, False, True)
y_prod = _tf.reduce_sum(y * dir_y, axis=1, keepdims=True)
y_prod = _tf.transpose(y_prod)
y_prod = _tf.tile(y_prod, [sx[0], 1])
dif = x_prod - y_prod
d = _pairwise_dist(x, y)
k = -_tf.sin(2.0 * _np.pi * d)
k = k * dif * 2.0 * _np.pi / (d + 1e-9)
return k
def covariance_matrix(self, x, y):
x = self.transform(x)
y = self.transform(y)
d = _pairwise_dist(x, y)
cov = (1 + 6 * d + 12 * d**2) * _tf.math.exp(-6 * d)
return cov