# This file is part of the pyMOR project (http://www.pymor.org).
# Copyright 2013-2020 pyMOR developers and contributors. All rights reserved.
# License: BSD 2-Clause License (http://opensource.org/licenses/BSD-2-Clause)
import weakref
import numpy as np
from scipy.linalg import solve, solve_triangular
from pymor.operators.constructions import VectorArrayOperator, Concatenation, ComponentProjection, ZeroOperator
from pymor.operators.interface import Operator
from pymor.operators.numpy import NumpyMatrixOperator
from pymor.vectorarrays.interface import VectorArray
from pymor.vectorarrays.numpy import NumpyVectorSpace
[docs]class EmpiricalInterpolatedOperator(Operator):
"""Interpolate an |Operator| using Empirical Operator Interpolation.
Let `L` be an |Operator|, `0 <= c_1, ..., c_M < L.range.dim` indices
of interpolation DOFs and let `b_1, ..., b_M in R^(L.range.dim)` be collateral
basis vectors. If moreover `ψ_j(U)` denotes the j-th component of `U`, the
empirical interpolation `L_EI` of `L` w.r.t. the given data is given by ::
| M
| L_EI(U, μ) = ∑ b_i⋅λ_i such that
| i=1
|
| ψ_(c_i)(L_EI(U, μ)) = ψ_(c_i)(L(U, μ)) for i=0,...,M
Since the original operator only has to be evaluated at the given interpolation
DOFs, |EmpiricalInterpolatedOperator| calls
:meth:`~pymor.operators.interface.Operator.restricted`
to obtain a restricted version of the operator which is used
to quickly obtain the required evaluations. If the `restricted` method, is not
implemented, the full operator will be evaluated (which will lead to
the same result, but without any speedup).
The interpolation DOFs and the collateral basis can be generated using
the algorithms provided in the :mod:`pymor.algorithms.ei` module.
Parameters
----------
operator
The |Operator| to interpolate.
interpolation_dofs
List or 1D |NumPy array| of the interpolation DOFs `c_1, ..., c_M`.
collateral_basis
|VectorArray| containing the collateral basis `b_1, ..., b_M`.
triangular
If `True`, assume that ψ_(c_i)(b_j) = 0 for i < j, which means
that the interpolation matrix is triangular.
solver_options
The |solver_options| for the operator.
name
Name of the operator.
"""
def __init__(self, operator, interpolation_dofs, collateral_basis, triangular,
solver_options=None, name=None):
assert isinstance(operator, Operator)
assert isinstance(collateral_basis, VectorArray)
assert collateral_basis in operator.range
assert len(interpolation_dofs) == len(collateral_basis)
self.source = operator.source
self.range = operator.range
self.linear = operator.linear
self.solver_options = solver_options
self.name = name or f'{operator.name}_interpolated'
self._operator = weakref.ref(operator)
interpolation_dofs = np.array(interpolation_dofs, dtype=np.int32)
self.interpolation_dofs = interpolation_dofs
self.triangular = triangular
if len(interpolation_dofs) > 0:
try:
self.restricted_operator, self.source_dofs = operator.restricted(interpolation_dofs)
except NotImplementedError:
self.logger.warning('Operator has no "restricted" method. The full operator will be evaluated.')
self._operator = operator
interpolation_matrix = collateral_basis.dofs(interpolation_dofs).T
self.interpolation_matrix = interpolation_matrix
self.collateral_basis = collateral_basis.copy()
@property
def operator(self):
if hasattr(self, 'restricted_operator'):
return self._operator()
else:
return self._operator
[docs] def apply(self, U, mu=None):
assert self.parameters.assert_compatible(mu)
if len(self.interpolation_dofs) == 0:
return self.range.zeros(len(U))
if hasattr(self, 'restricted_operator'):
U_dofs = NumpyVectorSpace.make_array(U.dofs(self.source_dofs))
AU = self.restricted_operator.apply(U_dofs, mu=mu)
else:
AU = NumpyVectorSpace.make_array(self.operator.apply(U, mu=mu).dofs(self.interpolation_dofs))
try:
if self.triangular:
interpolation_coefficients = solve_triangular(self.interpolation_matrix, AU.to_numpy().T,
lower=True, unit_diagonal=True).T
else:
interpolation_coefficients = solve(self.interpolation_matrix, AU.to_numpy().T).T
except ValueError: # this exception occurs when AU contains NaNs ...
interpolation_coefficients = np.empty((len(AU), len(self.collateral_basis))) + np.nan
return self.collateral_basis.lincomb(interpolation_coefficients)
[docs] def jacobian(self, U, mu=None):
assert self.parameters.assert_compatible(mu)
options = self.solver_options.get('jacobian') if self.solver_options else None
if len(self.interpolation_dofs) == 0:
if isinstance(self.source, NumpyVectorSpace) and isinstance(self.range, NumpyVectorSpace):
return NumpyMatrixOperator(np.zeros((self.range.dim, self.source.dim)), solver_options=options,
source_id=self.source.id, range_id=self.range.id,
name=self.name + '_jacobian')
else:
return ZeroOperator(self.range, self.source, name=self.name + '_jacobian')
elif hasattr(self, 'operator'):
return EmpiricalInterpolatedOperator(self.operator.jacobian(U, mu=mu), self.interpolation_dofs,
self.collateral_basis, self.triangular,
solver_options=options, name=self.name + '_jacobian')
else:
restricted_source = self.restricted_operator.source
U_dofs = restricted_source.make_array(U.dofs(self.source_dofs))
JU = self.restricted_operator.jacobian(U_dofs, mu=mu) \
.apply(restricted_source.make_array(np.eye(len(self.source_dofs))))
try:
if self.triangular:
interpolation_coefficients = solve_triangular(self.interpolation_matrix, JU.to_numpy().T,
lower=True, unit_diagonal=True).T
else:
interpolation_coefficients = solve(self.interpolation_matrix, JU.to_numpy().T).T
except ValueError: # this exception occurs when AU contains NaNs ...
interpolation_coefficients = np.empty((len(JU), len(self.collateral_basis))) + np.nan
J = self.collateral_basis.lincomb(interpolation_coefficients)
if isinstance(J.space, NumpyVectorSpace):
J = NumpyMatrixOperator(J.to_numpy().T, range_id=self.range.id)
else:
J = VectorArrayOperator(J)
return Concatenation([J, ComponentProjection(self.source_dofs, self.source)],
solver_options=options, name=self.name + '_jacobian')
def __getstate__(self):
d = self.__dict__.copy()
del d['_operator']
return d
[docs]class ProjectedEmpiciralInterpolatedOperator(Operator):
"""A projected |EmpiricalInterpolatedOperator|."""
def __init__(self, restricted_operator, interpolation_matrix, source_basis_dofs,
projected_collateral_basis, triangular, solver_options=None, name=None):
name = name or f'{restricted_operator.name}_projected'
self.__auto_init(locals())
self.source = NumpyVectorSpace(len(source_basis_dofs))
self.range = projected_collateral_basis.space
self.linear = restricted_operator.linear
[docs] def apply(self, U, mu=None):
assert self.parameters.assert_compatible(mu)
U_dofs = self.source_basis_dofs.lincomb(U.to_numpy())
AU = self.restricted_operator.apply(U_dofs, mu=mu)
try:
if self.triangular:
interpolation_coefficients = solve_triangular(self.interpolation_matrix, AU.to_numpy().T,
lower=True, unit_diagonal=True).T
else:
interpolation_coefficients = solve(self.interpolation_matrix, AU.to_numpy().T).T
except ValueError: # this exception occurs when AU contains NaNs ...
interpolation_coefficients = np.empty((len(AU), len(self.projected_collateral_basis))) + np.nan
return self.projected_collateral_basis.lincomb(interpolation_coefficients)
[docs] def jacobian(self, U, mu=None):
assert len(U) == 1
assert self.parameters.assert_compatible(mu)
options = self.solver_options.get('jacobian') if self.solver_options else None
if self.interpolation_matrix.shape[0] == 0:
return NumpyMatrixOperator(np.zeros((self.range.dim, self.source.dim)), solver_options=options,
name=self.name + '_jacobian')
U_dofs = self.source_basis_dofs.lincomb(U.to_numpy()[0])
J = self.restricted_operator.jacobian(U_dofs, mu=mu).apply(self.source_basis_dofs)
try:
if self.triangular:
interpolation_coefficients = solve_triangular(self.interpolation_matrix, J.to_numpy().T,
lower=True, unit_diagonal=True).T
else:
interpolation_coefficients = solve(self.interpolation_matrix, J.to_numpy().T).T
except ValueError: # this exception occurs when J contains NaNs ...
interpolation_coefficients = (np.empty((len(self.source_basis_dofs),
len(self.projected_collateral_basis)))
+ np.nan)
M = self.projected_collateral_basis.lincomb(interpolation_coefficients)
if isinstance(M.space, NumpyVectorSpace):
return NumpyMatrixOperator(M.to_numpy().T, solver_options=options)
else:
assert not options
return VectorArrayOperator(M)
def with_cb_dim(self, dim):
assert dim <= self.restricted_operator.range.dim
interpolation_matrix = self.interpolation_matrix[:dim, :dim]
restricted_operator, source_dofs = self.restricted_operator.restricted(np.arange(dim))
old_pcb = self.projected_collateral_basis
projected_collateral_basis = NumpyVectorSpace.make_array(old_pcb.to_numpy()[:dim, :])
old_sbd = self.source_basis_dofs
source_basis_dofs = NumpyVectorSpace.make_array(old_sbd.to_numpy()[:, source_dofs])
return ProjectedEmpiciralInterpolatedOperator(restricted_operator, interpolation_matrix,
source_basis_dofs, projected_collateral_basis, self.triangular,
solver_options=self.solver_options, name=self.name)