def _componentwise_distance(self, dx, small=False, opt=0):
"""
Compute the componentwise distance with respect to the correlation kernel
Parameters
----------
dx : numpy.ndarray
Distance matrix.
small : bool, optional
Compute the componentwise distance in small (n_components) dimension
or in initial dimension. The default is False.
opt : int, optional
useless for MGP
Returns
-------
d : numpy.ndarray
Component wise distance.
"""
if small:
d = componentwise_distance(dx, self.options["corr"], self.options["n_comp"])
else:
d = componentwise_distance(dx, self.options["corr"], self.nx)
return d
def test_corr_derivatives(self):
for ind, corr in enumerate(self.corr_def): # For every kernel
self.corr_str[ind] = self.corr_def[ind]
D = componentwise_distance(self.D, self.corr_str, self.X.shape[1])
k = corr(self.theta, D)
K = np.eye(self.X.shape[0])
K[self.ij[:, 0], self.ij[:, 1]] = k[:, 0]
K[self.ij[:, 1], self.ij[:, 0]] = k[:, 0]
grad_norm_all = []
diff_norm_all = []
ind_theta = []
for i, theta_i in enumerate(self.theta):
eps_theta = np.zeros(self.theta.shape)
eps_theta[i] = self.eps
k_dk = corr(self.theta + eps_theta, D)
K_dk = np.eye(self.X.shape[0])
K_dk[self.ij[:, 0], self.ij[:, 1]] = k_dk[:, 0]
K_dk[self.ij[:, 1], self.ij[:, 0]] = k_dk[:, 0]
grad_eps = (K_dk - K) / self.eps
dk = corr(self.theta, D, grad_ind=i)
dK = np.zeros((self.X.shape[0], self.X.shape[0]))
dK[self.ij[:, 0], self.ij[:, 1]] = dk[:, 0]
dK[self.ij[:, 1], self.ij[:, 0]] = dk[:, 0]
grad_norm_all.append(np.linalg.norm(dK))
diff_norm_all.append(np.linalg.norm(grad_eps))
ind_theta.append(r"$x_%d$" % i)
self.assert_error(np.array(grad_norm_all), np.array(diff_norm_all),
1e-5, 1e-5) # from utils/smt_test_case.py
def _componentwise_distance(self,
dx,
opt=0,
theta=None,
return_derivative=False):
if opt == 0:
# Kriging step
d = componentwise_distance(
dx,
self.options["corr"],
self.nx,
theta=theta,
return_derivative=return_derivative,
)
else:
# KPLS step
d = componentwise_distance_PLS(
dx,
self.options["corr"],
self.options["n_comp"],
self.coeff_pls,
theta=theta,
return_derivative=return_derivative,
)
return d
def _componentwise_distance(self,dx,opt=0):
if opt == 0:
# Kriging step
d = componentwise_distance(dx,self.options['corr'],self.nx)
else:
# KPLS step
d = componentwise_distance_PLS(dx,self.options['corr'],
self.options['n_comp'],self.coeff_pls)
return d
def _componentwise_distance(self, dx, opt=0):
# Modif for KPLSK model
if opt == 0:
# Kriging step
d = componentwise_distance(dx, self.options["corr"], self.nx)
else:
# KPLS step
d = super(MFKPLSK, self)._componentwise_distance(dx, opt)
return d
def _componentwise_distance(self,
dx,
opt=0,
theta=None,
return_derivative=False):
d = componentwise_distance(
dx,
self.options["corr"],
self.nx,
theta=theta,
return_derivative=return_derivative,
)
return d
def _componentwise_distance(self, dx, opt=0):
d = componentwise_distance(dx, self.options["corr"], self.nx)
return d
def _componentwise_distance(self,dx,opt=0):
d = componentwise_distance(dx,self.options['corr'].__name__,
self.nx)
return d
