def test_interpolate_tlm():
from firedrake_adjoint import ReducedFunctional, Control, taylor_test
mesh = UnitSquareMesh(10, 10)
V1 = VectorFunctionSpace(mesh, "CG", 1)
V2 = VectorFunctionSpace(mesh, "DG", 0)
V3 = VectorFunctionSpace(mesh, "CG", 2)
x = SpatialCoordinate(mesh)
f = interpolate(as_vector((x[0] * x[1], x[0] + x[1])), V1)
g = interpolate(as_vector((sin(x[1]) + x[0], cos(x[0]) * x[1])), V2)
u = Function(V3)
u.interpolate(f - 0.5 * g + f / (1 + dot(f, g)))
J = assemble(inner(f, g) * u**2 * dx)
rf = ReducedFunctional(J, Control(f))
h = Function(V1)
h.vector()[:] = 1
f.block_variable.tlm_value = h
tape = get_working_tape()
tape.evaluate_tlm()
assert J.block_variable.tlm_value is not None
assert taylor_test(rf, f, h, dJdm=J.block_variable.tlm_value) > 1.9
def test_interpolate_tlm_wit_constant():
from firedrake_adjoint import ReducedFunctional, Control, taylor_test
mesh = IntervalMesh(10, 0, 1)
V1 = FunctionSpace(mesh, "CG", 2)
V2 = FunctionSpace(mesh, "DG", 1)
x = SpatialCoordinate(mesh)
f = interpolate(x[0], V1)
g = interpolate(sin(x[0]), V1)
c = Constant(5.0)
u = Function(V2)
u.interpolate(c * f**2)
# test tlm w.r.t constant only:
c.block_variable.tlm_value = Constant(1.0)
J = assemble(u**2 * dx)
rf = ReducedFunctional(J, Control(c))
h = Constant(1.0)
tape = get_working_tape()
tape.evaluate_tlm()
assert abs(J.block_variable.tlm_value - 2.0) < 1e-5
assert taylor_test(rf, c, h, dJdm=J.block_variable.tlm_value) > 1.9
# test tlm w.r.t constant c and function f:
tape.reset_tlm_values()
c.block_variable.tlm_value = Constant(0.4)
f.block_variable.tlm_value = g
rf(c) # replay to reset checkpoint values based on c=5
tape.evaluate_tlm()
assert abs(J.block_variable.tlm_value -
(0.8 + 100. * (5 * cos(1.) - 3 * sin(1.)))) < 1e-4
def assemble(*args, **kwargs):
"""When a form is assembled, the information about its nonlinear dependencies is lost,
and it is no longer easy to manipulate. Therefore, fenics_adjoint overloads the :py:func:`dolfin.assemble`
function to *attach the form to the assembled object*. This lets the automatic annotation work,
even when the user calls the lower-level :py:data:`solve(A, x, b)`.
"""
annotate = annotate_tape(kwargs)
with stop_annotating():
output = backend.assemble(*args, **kwargs)
form = args[0]
if isinstance(output, float):
output = create_overloaded_object(output)
if annotate:
block = AssembleBlock(form)
tape = get_working_tape()
tape.add_block(block)
block.add_output(output.block_variable)
else:
# Assembled a vector or matrix
output.form = form
return output
def test_supermesh_project_hessian(vector):
from firedrake_adjoint import ReducedFunctional, Control, taylor_test
source, target_space = supermesh_setup()
control = Control(source)
target = project(source, target_space)
J = assemble(inner(target, target)**2 * dx)
rf = ReducedFunctional(J, control)
source_space = source.function_space()
h = Function(source_space)
h.vector()[:] = 10 * rand(source_space.dim())
J.block_variable.adj_value = 1.0
source.block_variable.tlm_value = h
tape = get_working_tape()
tape.evaluate_adj()
tape.evaluate_tlm()
J.block_variable.hessian_value = 0.0
tape.evaluate_hessian()
dJdm = J.block_variable.tlm_value
assert isinstance(source.block_variable.adj_value, Vector)
assert isinstance(source.block_variable.hessian_value, Vector)
Hm = source.block_variable.hessian_value.inner(h.vector())
assert taylor_test(rf, source, h, dJdm=dJdm, Hm=Hm) > 2.9
def solve(self, *args, **kwargs):
annotate = annotate_tape(kwargs)
if annotate:
tape = get_working_tape()
factory = args[0]
vec = args[1]
b = backend.as_backend_type(vec).__class__()
factory.F(b=b, x=vec)
F = b.form
bcs = b.bcs
u = vec.function
sb_kwargs = SolveVarFormBlock.pop_kwargs(kwargs)
block = SolveVarFormBlock(
F == 0,
u,
bcs,
solver_parameters={"newton_solver": self.parameters.copy()},
**sb_kwargs)
tape.add_block(block)
newargs = [self] + list(args)
out = backend.NewtonSolver.solve(*newargs, **kwargs)
if annotate:
block.add_output(u.create_block_variable())
return out
def wrapper(*args, **kwargs):
"""The project call performs an equation solve, and so it too must be annotated so that the
adjoint and tangent linear models may be constructed automatically by pyadjoint.
To disable the annotation of this function, just pass :py:data:`annotate=False`. This is useful in
cases where the solve is known to be irrelevant or diagnostic for the purposes of the adjoint
computation (such as projecting fields to other function spaces for the purposes of
visualisation)."""
annotate = annotate_tape(kwargs)
if annotate:
bcs = kwargs.get("bcs", [])
sb_kwargs = ProjectBlock.pop_kwargs(kwargs)
if isinstance(args[1], function.Function):
# block should be created before project because output might also be an input that needs checkpointing
output = args[1]
V = output.function_space()
block = ProjectBlock(args[0], V, output, bcs, **sb_kwargs)
with stop_annotating():
output = project(*args, **kwargs)
if annotate:
tape = get_working_tape()
if not isinstance(args[1], function.Function):
block = ProjectBlock(args[0], args[1], output, bcs,
**sb_kwargs)
tape.add_block(block)
block.add_output(output.create_block_variable())
return output
def solve(self, **kwargs):
annotate = annotate_tape()
if annotate:
block_helper = BlockSolveBlockHelper()
tape = get_working_tape()
problem = self._ad_problem
# sb_kwargs = SolveBlock.pop_kwargs(kwargs)
block = NonlinearBlockSolveBlock(
problem._ad_b == 0,
problem._ad_u,
problem._ad_bcs,
block_helper=block_helper,
problem_J=problem._ad_A,
block_field=self._ad_problem.block_field,
block_split=self._ad_problem.block_split)
tape.add_block(block)
with stop_annotating():
out = super(NonlinearBlockSolver, self).solve()
if annotate:
block.add_output(self._ad_problem._ad_u.create_block_variable())
return out
def wrapper(*args, **kwargs):
ad_block_tag = kwargs.pop("ad_block_tag", None)
annotate = annotate_tape(kwargs)
if annotate:
tape = get_working_tape()
solve_block_type = SolveVarFormBlock
if not isinstance(args[0], ufl.equation.Equation):
solve_block_type = SolveLinearSystemBlock
sb_kwargs = solve_block_type.pop_kwargs(kwargs)
sb_kwargs.update(kwargs)
block = solve_block_type(*args,
ad_block_tag=ad_block_tag,
**sb_kwargs)
tape.add_block(block)
with stop_annotating():
output = solve(*args, **kwargs)
if annotate:
if hasattr(args[1], "create_block_variable"):
block_variable = args[1].create_block_variable()
else:
block_variable = args[1].function.create_block_variable()
block.add_output(block_variable)
return output
def wrapper(self, **kwargs):
"""To disable the annotation, just pass :py:data:`annotate=False` to this routine, and it acts exactly like the
Firedrake solve call. This is useful in cases where the solve is known to be irrelevant or diagnostic
for the purposes of the adjoint computation (such as projecting fields to other function spaces
for the purposes of visualisation)."""
annotate = annotate_tape(kwargs)
if annotate:
tape = get_working_tape()
problem = self._ad_problem
sb_kwargs = NonlinearVariationalSolveBlock.pop_kwargs(kwargs)
sb_kwargs.update(kwargs)
block = NonlinearVariationalSolveBlock(
problem._ad_F == 0,
problem._ad_u,
problem._ad_bcs,
problem_J=problem._ad_J,
solver_params=self.parameters,
solver_kwargs=self._ad_kwargs,
**sb_kwargs)
tape.add_block(block)
with stop_annotating():
out = solve(self, **kwargs)
if annotate:
block.add_output(
self._ad_problem._ad_u.create_block_variable())
return out
def solve(self, *args, **kwargs):
annotate = annotate_tape(kwargs)
if annotate:
if len(args) == 3:
block_helper = LUSolveBlockHelper()
A = args[0]
x = args[1]
b = args[2]
elif len(args) == 2:
block_helper = self.block_helper
A = self.operator
x = args[0]
b = args[1]
u = x.function
parameters = self.parameters.copy()
tape = get_working_tape()
sb_kwargs = LUSolveBlock.pop_kwargs(kwargs)
block = LUSolveBlock(A,
x,
b,
lu_solver_parameters=parameters,
block_helper=block_helper,
lu_solver_method=self.method,
**sb_kwargs)
tape.add_block(block)
out = backend.LUSolver.solve(self, *args, **kwargs)
if annotate:
block.add_output(u.create_block_variable())
return out
def wrapper(self, b, *args, **kwargs):
ad_block_tag = kwargs.pop("ad_block_tag", None)
annotate = annotate_tape(kwargs)
if annotate:
bcs = kwargs.get("bcs", [])
if isinstance(
b, firedrake.Function
) and b.ufl_domain() != self.function_space().mesh():
block = SupermeshProjectBlock(b,
self.function_space(),
self,
bcs,
ad_block_tag=ad_block_tag)
else:
block = ProjectBlock(b,
self.function_space(),
self,
bcs,
ad_block_tag=ad_block_tag)
tape = get_working_tape()
tape.add_block(block)
with stop_annotating():
output = project(self, b, *args, **kwargs)
if annotate:
block.add_output(output.create_block_variable())
return output
def wrapper(*args, **kwargs):
"""The project call performs an equation solve, and so it too must be annotated so that the
adjoint and tangent linear models may be constructed automatically by pyadjoint.
To disable the annotation of this function, just pass :py:data:`annotate=False`. This is useful in
cases where the solve is known to be irrelevant or diagnostic for the purposes of the adjoint
computation (such as projecting fields to other function spaces for the purposes of
visualisation)."""
annotate = annotate_tape(kwargs)
with stop_annotating():
output = project(*args, **kwargs)
output = create_overloaded_object(output)
if annotate:
bcs = kwargs.pop("bcs", [])
sb_kwargs = ProjectBlock.pop_kwargs(kwargs)
block = ProjectBlock(args[0], args[1], output, bcs, **sb_kwargs)
tape = get_working_tape()
tape.add_block(block)
block.add_output(output.block_variable)
return output
def wrapper(interpolator, *function, **kwargs):
"""To disable the annotation, just pass :py:data:`annotate=False` to this routine, and it acts exactly like the
Firedrake interpolate call."""
ad_block_tag = kwargs.pop("ad_block_tag", None)
annotate = annotate_tape(kwargs)
if annotate:
sb_kwargs = InterpolateBlock.pop_kwargs(kwargs)
sb_kwargs.update(kwargs)
block = InterpolateBlock(interpolator,
*function,
ad_block_tag=ad_block_tag,
**sb_kwargs)
tape = get_working_tape()
tape.add_block(block)
with stop_annotating():
output = interpolate(interpolator, *function, **kwargs)
if annotate:
from firedrake import Function
if isinstance(interpolator.V, Function):
block.add_output(output.create_block_variable())
else:
block.add_output(output.block_variable)
return output
def assign(self, *args, **kwargs):
annotate = annotate_tape(kwargs)
outputs = Enlist(args[0])
inputs = Enlist(args[1])
if annotate:
for i, o in enumerate(outputs):
if not isinstance(o, OverloadedType):
outputs[i] = create_overloaded_object(o)
for j, i in enumerate(outputs):
if not isinstance(i, OverloadedType):
inputs[j] = create_overloaded_object(i)
block = FunctionAssignerBlock(self, inputs)
tape = get_working_tape()
tape.add_block(block)
with stop_annotating():
ret = backend.FunctionAssigner.assign(self, outputs.delist(), inputs.delist(), **kwargs)
if annotate:
for output in outputs:
block.add_output(output.block_variable)
return ret
def sub(self, i, deepcopy=False, **kwargs):
from .function_assigner import FunctionAssigner, FunctionAssignerBlock
annotate = annotate_tape(kwargs)
if deepcopy:
ret = create_overloaded_object(
backend.Function.sub(self, i, deepcopy, **kwargs))
if annotate:
fa = FunctionAssigner(ret.function_space(),
self.function_space())
block = FunctionAssignerBlock(fa, Enlist(self))
tape = get_working_tape()
tape.add_block(block)
block.add_output(ret.block_variable)
else:
extra_kwargs = {}
if annotate:
extra_kwargs = {
"block_class": FunctionSplitBlock,
"_ad_floating_active": True,
"_ad_args": [self, i],
"_ad_output_args": [i],
"output_block_class": FunctionMergeBlock,
"_ad_outputs": [self],
}
ret = compat.create_function(self, i, **extra_kwargs)
return ret
def __init__(
self,
functional,
controls,
level_set,
scale=1.0,
tape=None,
eval_cb_pre=lambda *args: None,
eval_cb_post=lambda *args: None,
derivative_cb_pre=lambda *args: None,
derivative_cb_post=lambda *args: None,
hessian_cb_pre=lambda *args: None,
hessian_cb_post=lambda *args: None,
):
self.functional = functional
self.cost_function = self.functional
self.tape = get_working_tape() if tape is None else tape
self.controls = Enlist(controls)
self.level_set = Enlist(level_set)
self.scale = scale
self.eval_cb_pre = eval_cb_pre
self.eval_cb_post = eval_cb_post
self.derivative_cb_pre = derivative_cb_pre
self.derivative_cb_post = derivative_cb_post
self.hessian_cb_pre = hessian_cb_pre
self.hessian_cb_post = hessian_cb_post
def wrapper(*args, **kwargs):
"""When a form is assembled, the information about its nonlinear dependencies is lost,
and it is no longer easy to manipulate. Therefore, we decorate :func:`.assemble`
to *attach the form to the assembled object*. This lets the automatic annotation work,
even when the user calls the lower-level :py:data:`solve(A, x, b)`.
"""
ad_block_tag = kwargs.pop("ad_block_tag", None)
annotate = annotate_tape(kwargs)
with stop_annotating():
output = assemble(*args, **kwargs)
form = args[0]
if isinstance(output, numbers.Complex):
if not annotate:
return output
if not isinstance(output, float):
raise NotImplementedError(
"Taping for complex-valued 0-forms not yet done!")
output = create_overloaded_object(output)
block = AssembleBlock(form, ad_block_tag=ad_block_tag)
tape = get_working_tape()
tape.add_block(block)
block.add_output(output.block_variable)
else:
# Assembled a vector or matrix
output.form = form
return output
def test_interpolate_hessian_nonlinear_expr_multi():
# Note this is a direct copy of
# pyadjoint/tests/firedrake_adjoint/test_hessian.py::test_nonlinear
# with modifications where indicated.
from firedrake_adjoint import ReducedFunctional, Control, taylor_test, get_working_tape
# Get tape instead of creating a new one for consistency with other tests
tape = get_working_tape()
mesh = UnitSquareMesh(10, 10)
V = FunctionSpace(mesh, "Lagrange", 1)
# Interpolate from f in another function space to force hessian evaluation
# of interpolation. Functions in W form our control space c, our expansion
# space h and perterbation direction g.
W = FunctionSpace(mesh, "Lagrange", 2)
f = Function(W)
f.vector()[:] = 5
w = Function(W)
w.vector()[:] = 4
c = Constant(2.)
# Note that we interpolate from a nonlinear expression with 3 coefficients
expr_interped = Function(V).interpolate(f**2 + w**2 + c**2)
u = Function(V)
v = TestFunction(V)
bc = DirichletBC(V, Constant(1), "on_boundary")
F = inner(grad(u), grad(v)) * dx - u**2 * v * dx - expr_interped * v * dx
solve(F == 0, u, bc)
J = assemble(u**4 * dx)
Jhat = ReducedFunctional(J, Control(f))
# Note functions are in W, not V.
h = Function(W)
h.vector()[:] = 10 * rand(W.dim())
J.block_variable.adj_value = 1.0
# Note only the tlm_value of f is set here - unclear why.
f.block_variable.tlm_value = h
tape.evaluate_adj()
tape.evaluate_tlm()
J.block_variable.hessian_value = 0
tape.evaluate_hessian()
g = f.copy(deepcopy=True)
dJdm = J.block_variable.tlm_value
assert isinstance(f.block_variable.adj_value, Vector)
assert isinstance(f.block_variable.hessian_value, Vector)
Hm = f.block_variable.hessian_value.inner(h.vector())
# If the new interpolate block has the right hessian, taylor test
# convergence rate should be as for the unmodified test.
assert taylor_test(Jhat, g, h, dJdm=dJdm, Hm=Hm) > 2.9
def wrapper(self, other, **kwargs):
annotate = annotate_tape(kwargs)
func = __idiv__(self, other, **kwargs)
if annotate:
block = FunctionAssignBlock(func, self / other)
tape = get_working_tape()
tape.add_block(block)
block.add_output(func.create_block_variable())
return func
def get_solve_blocks():
"""
Extract all blocks of the tape which correspond
to PDE solves, except for those which correspond
to calls of the ``project`` operator.
"""
return [
block for block in get_working_tape().get_blocks()
if issubclass(type(block), GenericSolveBlock)
and not issubclass(type(block), ProjectBlock)
]
def __init__(self, *args, **kwargs):
ad_block_tag = kwargs.pop("ad_block_tag", None)
annotate = annotate_tape(kwargs)
super(Constant, self).__init__(*args, **kwargs)
backend.Constant.__init__(self, *args, **kwargs)
if annotate and len(args) > 0:
value = args[0]
if isinstance(value, OverloadedType):
block = ConstantAssignBlock(value, ad_block_tag=ad_block_tag)
tape = get_working_tape()
tape.add_block(block)
block.add_output(self.block_variable)
elif isinstance(value, (tuple, list)):
value = numpy.array(value, dtype="O")
if any(isinstance(v, OverloadedType) for v in value.flat):
block = ConstantAssignBlock(value,
ad_block_tag=ad_block_tag)
tape = get_working_tape()
tape.add_block(block)
block.add_output(self.block_variable)
def move(mesh, vector, **kwargs):
annotate = annotate_tape(kwargs)
if annotate:
assert isinstance(mesh, OverloadedType)
assert isinstance(vector, OverloadedType)
tape = get_working_tape()
block = ALEMoveBlock(mesh, vector, **kwargs)
tape.add_block(block)
with stop_annotating():
output = __backend_ALE_move(mesh, vector)
if annotate:
block.add_output(mesh.create_block_variable())
return output
def wrapper(self, other, **kwargs):
ad_block_tag = kwargs.pop("ad_block_tag", None)
annotate = annotate_tape(kwargs)
func = __imul__(self, other, **kwargs)
if annotate:
block = FunctionAssignBlock(func,
self * other,
ad_block_tag=ad_block_tag)
tape = get_working_tape()
tape.add_block(block)
block.add_output(func.create_block_variable())
return func
def __getitem__(self, item):
annotate = annotate_tape()
if annotate:
block = NumpyArraySliceBlock(self, item)
tape = get_working_tape()
tape.add_block(block)
with stop_annotating():
out = numpy.ndarray.__getitem__(self, item)
if annotate:
out = create_overloaded_object(out)
block.add_output(out.create_block_variable())
return out
def solve(*args, **kwargs):
"""This solve routine wraps the real Dolfin solve call. Its purpose is to annotate the model,
recording what solves occur and what forms are involved, so that the adjoint and tangent linear models may be
constructed automatically by pyadjoint.
To disable the annotation, just pass :py:data:`annotate=False` to this routine, and it acts exactly like the
Dolfin solve call. This is useful in cases where the solve is known to be irrelevant or diagnostic
for the purposes of the adjoint computation (such as projecting fields to other function spaces
for the purposes of visualisation).
The overloaded solve takes optional callback functions to extract adjoint solutions.
All of the callback functions follow the same signature, taking a single argument of type Function.
Keyword Args:
adj_cb (function, optional): callback function supplying the adjoint solution in the interior.
The boundary values are zero.
adj_bdy_cb (function, optional): callback function supplying the adjoint solution on the boundary.
The interior values are not guaranteed to be zero.
adj2_cb (function, optional): callback function supplying the second-order adjoint solution in the interior.
The boundary values are zero.
adj2_bdy_cb (function, optional): callback function supplying the second-order adjoint solution on
the boundary. The interior values are not guaranteed to be zero.
"""
ad_block_tag = kwargs.pop("ad_block_tag", None)
annotate = annotate_tape(kwargs)
if annotate:
tape = get_working_tape()
solve_block_type = SolveVarFormBlock
if not isinstance(args[0], ufl.equation.Equation):
solve_block_type = SolveLinearSystemBlock
sb_kwargs = solve_block_type.pop_kwargs(kwargs)
sb_kwargs.update(kwargs)
block = solve_block_type(*args, ad_block_tag=ad_block_tag, **sb_kwargs)
tape.add_block(block)
with stop_annotating():
output = backend.solve(*args, **kwargs)
if annotate:
if hasattr(args[1], "create_block_variable"):
block_variable = args[1].create_block_variable()
else:
block_variable = args[1].function.create_block_variable()
block.add_output(block_variable)
return output
def wrapper(self, *args, **kwargs):
annotate = annotate_tape(kwargs)
func = copy(self, *args, **kwargs)
if annotate:
if kwargs.pop("deepcopy", False):
block = FunctionAssignBlock(func, self)
tape = get_working_tape()
tape.add_block(block)
block.add_output(func.create_block_variable())
else:
# TODO: Implement. Here we would need to use floating types.
raise NotImplementedError("Currently kwargs['deepcopy'] must be set True")
return func
def project(self, b, *args, **kwargs):
annotate = annotate_tape(kwargs)
with stop_annotating():
output = super(Function, self).project(b, *args, **kwargs)
output = create_overloaded_object(output)
if annotate:
bcs = kwargs.pop("bcs", [])
block = ProjectBlock(b, self.function_space(), output, bcs)
tape = get_working_tape()
tape.add_block(block)
block.add_output(output.create_block_variable())
return output
def copy(self, *args, **kwargs):
annotate = annotate_tape(kwargs)
c = backend.Function.copy(self, *args, **kwargs)
func = create_overloaded_object(c)
if annotate:
if kwargs.pop("deepcopy", False):
block = FunctionAssignBlock(func, self)
tape = get_working_tape()
tape.add_block(block)
block.add_output(func.create_block_variable())
else:
# TODO: Implement. Here we would need to use floating types.
pass
return func
def assign(self, *args, **kwargs):
annotate_tape = kwargs.pop("annotate_tape", True)
if annotate_tape:
other = args[0]
if not isinstance(other, OverloadedType):
other = create_overloaded_object(other)
block = AssignBlock(self, other)
tape = get_working_tape()
tape.add_block(block)
ret = backend.Constant.assign(self, *args, **kwargs)
if annotate_tape:
block.add_output(self.create_block_variable())
return ret
def wrapper(self, *args, **kwargs):
annotate = annotate_tape(kwargs)
if annotate:
other = args[0]
if not isinstance(other, OverloadedType):
other = create_overloaded_object(other)
block = ConstantAssignBlock(other)
tape = get_working_tape()
tape.add_block(block)
ret = assign(self, *args, **kwargs)
if annotate:
block.add_output(self.create_block_variable())
return ret
