Dualspace update

demos/linear-wave-equation/linear_wave_equation.py.rst

@@ -107,7 +107,7 @@ options at this point, we may either `lump` the mass, which reduces
 the inversion to a pointwise division::
 
       if lump_mass:
-          p += interpolate(assemble(dt * inner(nabla_grad(v), nabla_grad(phi))*dx) / assemble(v*dx), V)
+          p.dat.data[:] += assemble(dt * inner(nabla_grad(v), nabla_grad(phi))*dx).dat.data_ro / assemble(v*dx).dat.data_ro
 
 In the mass lumped case, we must now ensure that the resulting
 solution for :math:`p` satisfies the boundary conditions::

docs/notebooks/11-extract-adjoint-solutions.ipynb

@@ -1227,6 +1227,8 @@
     "    t = i*timesteps_per_export*dt\n",
     "    tricontourf(forward_solutions[i], axes=axs[i, 0])\n",
     "    adjoint_solution = dJdu if i == num_exports else solve_blocks[timesteps_per_export*i].adj_sol\n",
+    "    # Get the Riesz representer\n",
+    "    adjoint_solution = dJdu.riesz_representation(riesz_map=\"H1\")\n",
     "    tricontourf(adjoint_solution, axes=axs[i, 1])\n",
     "    axs[i, 0].annotate('t={:.2f}'.format(t), (0.05, 0.05), color='white');\n",
     "    axs[i, 1].annotate('t={:.2f}'.format(t), (0.05, 0.05), color='white');\n",

firedrake/__init__.py

@@ -72,6 +72,7 @@
 from firedrake.assemble import *
 from firedrake.bcs import *
 from firedrake.checkpointing import *
+from firedrake.cofunction import *
 from firedrake.constant import *
 from firedrake.exceptions import *
 from firedrake.function import *

firedrake/adjoint_utils/assembly.py

@@ -19,12 +19,14 @@ def wrapper(*args, **kwargs):
             output = assemble(*args, **kwargs)
 
         from firedrake.function import Function
+        from firedrake.cofunction import Cofunction
         form = args[0]
-        if isinstance(output, (numbers.Complex, Function)):
+        if isinstance(output, (numbers.Complex, Function, Cofunction)):
+            # Assembling a 0-form or 1-form (e.g. Form)
             if not annotate:
                 return output
 
-            if not isinstance(output, (float, Function)):
+            if not isinstance(output, (float, Function, Cofunction)):
                 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)
@@ -34,7 +36,7 @@ def wrapper(*args, **kwargs):
 
             block.add_output(output.block_variable)
         else:
-            # Assembled a matrix
+            # Assembled a 2-form
             output.form = form
 
         return output

firedrake/adjoint_utils/blocks/assembly.py

@@ -1,7 +1,7 @@
 import ufl
 import firedrake
 from ufl.formatting.ufl2unicode import ufl2unicode
-from pyadjoint import Block, create_overloaded_object
+from pyadjoint import Block, AdjFloat, create_overloaded_object
 from .backend import Backend
 from firedrake.adjoint_utils.checkpointing import maybe_disk_checkpoint
 
@@ -13,8 +13,11 @@ def __init__(self, form, ad_block_tag=None):
         if self.backend.__name__ != "firedrake":
             mesh = self.form.ufl_domain().ufl_cargo()
         else:
-            mesh = self.form.ufl_domain()
-        self.add_dependency(mesh)
+            mesh = self.form.ufl_domain() if hasattr(self.form, 'ufl_domain') else None
+
+        if mesh:
+            self.add_dependency(mesh)
+
         for c in self.form.coefficients():
             self.add_dependency(c, no_duplicates=True)
 
@@ -28,7 +31,7 @@ def compute_action_adjoint(self, adj_input, arity_form, form=None,
            `<(dform/dc_rep)*, adj_input>`
 
            - If `form` has arity 0 => `dform/dc_rep` is a 1-form and
-             `adj_input` a foat, we can simply use the `*` operator.
+             `adj_input` a float, we can simply use the `*` operator.
 
            - If `form` has arity 1 => `dform/dc_rep` is a 2-form and we can
              symbolically take its adjoint and then apply the action on
@@ -38,31 +41,38 @@ def compute_action_adjoint(self, adj_input, arity_form, form=None,
             if dform is None:
                 dc = self.backend.TestFunction(space)
                 dform = self.backend.derivative(form, c_rep, dc)
-            dform_vector = self.compat.assemble_adjoint_value(dform)
-            # Return a Vector scaled by the scalar `adj_input`
-            return dform_vector * adj_input, dform
+            dform_adj = self.compat.assemble_adjoint_value(dform)
+            if dform_adj == 0:
+                # `dform_adj` is a `ZeroBaseForm`
+                return AdjFloat(0.), dform
+            # Return the adjoint model of `form` scaled by the scalar `adj_input`
+            adj_output = dform_adj._ad_mul(adj_input)
+            return adj_output, dform
         elif arity_form == 1:
             if dform is None:
                 dc = self.backend.TrialFunction(space)
                 dform = self.backend.derivative(form, c_rep, dc)
-            # Get the Function
-            adj_input = adj_input.function
             # Symbolic operators such as action/adjoint require derivatives to
             # have been expanded beforehand. However, UFL doesn't support
             # expanding coordinate derivatives of Coefficients in physical
             # space, implying that we can't symbolically take the
-            # action/adjoint of the Jacobian for SpatialCoordinates. ->
-            # Workaround: Apply action/adjoint numerically (using PETSc).
+            # action/adjoint of the Jacobian for SpatialCoordinates.
+            # -> Workaround: Apply action/adjoint numerically (using PETSc).
             if not isinstance(c_rep, self.backend.SpatialCoordinate):
                 # Symbolically compute: (dform/dc_rep)^* * adj_input
                 adj_output = self.backend.action(self.backend.adjoint(dform),
                                                  adj_input)
                 adj_output = self.compat.assemble_adjoint_value(adj_output)
             else:
+                adj_output = self.backend.Cofunction(space.dual())
+                # Assemble `dform`: derivatives are expanded along the way
+                # which may lead to a ZeroBaseForm
+                assembled_dform = self.compat.assemble_adjoint_value(dform)
+                if assembled_dform == 0:
+                    return adj_output, dform
                 # Get PETSc matrix
-                dform_mat = self.compat.assemble_adjoint_value(dform).petscmat
+                dform_mat = assembled_dform.petscmat
                 # Action of the adjoint (Hermitian transpose)
-                adj_output = self.backend.Function(space)
                 with adj_input.dat.vec_ro as v_vec:
                     with adj_output.dat.vec as res_vec:
                         dform_mat.multHermitian(v_vec, res_vec)
@@ -105,7 +115,7 @@ def evaluate_adj_component(self, inputs, adj_inputs, block_variable, idx,
         if self.compat.isconstant(c):
             mesh = self.compat.extract_mesh_from_form(self.form)
             space = c._ad_function_space(mesh)
-        elif isinstance(c, self.backend.Function):
+        elif isinstance(c, (self.backend.Function, self.backend.Cofunction)):
             space = c.function_space()
         elif isinstance(c, self.compat.MeshType):
             c_rep = self.backend.SpatialCoordinate(c_rep)
@@ -123,8 +133,6 @@ def evaluate_tlm_component(self, inputs, tlm_inputs, block_variable, idx,
         form = prepared
         dform = 0.
 
-        from ufl.algorithms.analysis import extract_arguments
-        arity_form = len(extract_arguments(form))
         for bv in self.get_dependencies():
             c_rep = bv.saved_output
             tlm_value = bv.tlm_value
@@ -133,15 +141,14 @@ def evaluate_tlm_component(self, inputs, tlm_inputs, block_variable, idx,
                 continue
             if isinstance(c_rep, self.compat.MeshType):
                 X = self.backend.SpatialCoordinate(c_rep)
+                # Spatial coordinates derivatives cannot be expanded in the physical space,
+                # which is required by symbolic operators such as `action`.
                 dform += self.backend.derivative(form, X, tlm_value)
             else:
-                dform += self.backend.derivative(form, c_rep, tlm_value)
+                dform += self.backend.action(self.backend.derivative(form, c_rep), tlm_value)
         if not isinstance(dform, float):
             dform = ufl.algorithms.expand_derivatives(dform)
             dform = self.compat.assemble_adjoint_value(dform)
-            if arity_form == 1 and dform != 0:
-                # Then dform is a Vector
-                dform = dform.function
         return dform
 
     def prepare_evaluate_hessian(self, inputs, hessian_inputs, adj_inputs,
@@ -165,7 +172,7 @@ def evaluate_hessian_component(self, inputs, hessian_inputs, adj_inputs,
         if self.compat.isconstant(c1):
             mesh = self.compat.extract_mesh_from_form(form)
             space = c1._ad_function_space(mesh)
-        elif isinstance(c1, self.backend.Function):
+        elif isinstance(c1, (self.backend.Function, self.backend.Cofunction)):
             space = c1.function_space()
         elif isinstance(c1, self.compat.ExpressionType):
             mesh = form.ufl_domain().ufl_cargo()
@@ -180,7 +187,7 @@ def evaluate_hessian_component(self, inputs, hessian_inputs, adj_inputs,
             hessian_input, arity_form, form, c1_rep, space
         )
 
-        ddform = 0
+        ddform = 0.
         for other_idx, bv in relevant_dependencies:
             c2_rep = bv.saved_output
             tlm_input = bv.tlm_value
@@ -196,10 +203,8 @@ def evaluate_hessian_component(self, inputs, hessian_inputs, adj_inputs,
 
         if not isinstance(ddform, float):
             ddform = ufl.algorithms.expand_derivatives(ddform)
-            if not ddform.empty():
-                hessian_outputs += self.compute_action_adjoint(
-                    adj_input, arity_form, dform=ddform
-                )[0]
+            if not (isinstance(ddform, ufl.ZeroBaseForm) or (isinstance(ddform, ufl.Form) and ddform.empty())):
+                hessian_outputs += self.compute_action_adjoint(adj_input, arity_form, dform=ddform)[0]
 
         if isinstance(c1, self.compat.ExpressionType):
             return [(hessian_outputs, space)]

firedrake/adjoint_utils/blocks/compat.py

@@ -77,7 +77,7 @@ def extract_bc_subvector(value, Vtarget, bc):
             for idx in bc._indices:
                 r = r.sub(idx)
             assert Vtarget == r.function_space()
-            return r.vector()
+            return r
         compat.extract_bc_subvector = extract_bc_subvector
 
         def extract_mesh_from_form(form):
@@ -115,15 +115,7 @@ def constant_function_firedrake_compat(value):
             return value.dat.data
         compat.constant_function_firedrake_compat = constant_function_firedrake_compat
 
-        def assemble_adjoint_value(*args, **kwargs):
-            """A wrapper around Firedrake's assemble that returns a Vector
-            instead of a Function when assembling a 1-form."""
-            result = backend.assemble(*args, **kwargs)
-            if isinstance(result, backend.Function):
-                return result.vector()
-            else:
-                return result
-        compat.assemble_adjoint_value = assemble_adjoint_value
+        compat.assemble_adjoint_value = backend.assemble
 
         def gather(vec):
             return vec.gather()

firedrake/adjoint_utils/blocks/dirichlet_bc.py

@@ -39,7 +39,7 @@ def evaluate_adj_component(self, inputs, adj_inputs, block_variable, idx,
         adj_output = None
         for adj_input in adj_inputs:
             if self.compat.isconstant(c):
-                adj_value = self.backend.Function(self.parent_space)
+                adj_value = self.backend.Function(self.parent_space.dual())
                 adj_input.apply(adj_value.vector())
                 if self.function_space != self.parent_space:
                     vec = self.compat.extract_bc_subvector(
@@ -77,13 +77,13 @@ def evaluate_adj_component(self, inputs, adj_inputs, block_variable, idx,
                 #       you can even use the Function outside its domain.
                 # For now we will just assume the FunctionSpace is the same for
                 # the BC and the Function.
-                adj_value = self.backend.Function(self.parent_space)
+                adj_value = self.backend.Function(self.parent_space.dual())
                 adj_input.apply(adj_value.vector())
                 r = self.compat.extract_bc_subvector(
                     adj_value, c.function_space(), bc
                 )
             elif isinstance(c, self.compat.Expression):
-                adj_value = self.backend.Function(self.parent_space)
+                adj_value = self.backend.Function(self.parent_space.dual())
                 adj_input.apply(adj_value.vector())
                 output = self.compat.extract_bc_subvector(
                     adj_value, self.collapsed_space, bc
@@ -93,6 +93,7 @@ def evaluate_adj_component(self, inputs, adj_inputs, block_variable, idx,
                 adj_output = r
             else:
                 adj_output += r
+
         return adj_output
 
     def evaluate_tlm_component(self, inputs, tlm_inputs, block_variable, idx,