firedrake.adjoint package¶
Submodules¶
firedrake.adjoint.assembly module¶
firedrake.adjoint.blocks module¶
- class firedrake.adjoint.blocks.AssembleBlock(form, ad_block_tag=None)[source]¶
Bases:
AssembleBlock
,Backend
- block_helper¶
- recompute_component(inputs, block_variable, idx, prepared)[source]¶
This method must be overridden.
The method should implement a routine for recomputing one output of the block in the forward computations. The output to recompute is determined by the idx argument, which corresponds to the index of the output in the outputs list. If the block only has a single output, then idx will always be 0.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the output to compute. prepared (object): Anything returned by the prepare_recompute_component method. Default is None.
- Returns:
An object of the same type as block_variable.checkpoint which is determined by OverloadedType._ad_create_checkpoint (often the same as block_variable.saved_output): The new output.
- class firedrake.adjoint.blocks.ConstantAssignBlock(other, ad_block_tag=None)[source]¶
Bases:
ConstantAssignBlock
,Backend
- block_helper¶
- class firedrake.adjoint.blocks.DirichletBCBlock(*args, **kwargs)[source]¶
Bases:
DirichletBCBlock
,Backend
- block_helper¶
- class firedrake.adjoint.blocks.FunctionAssignBlock(func, other, ad_block_tag=None)[source]¶
Bases:
FunctionAssignBlock
,Backend
- block_helper¶
- recompute_component(inputs, block_variable, idx, prepared)[source]¶
This method must be overridden.
The method should implement a routine for recomputing one output of the block in the forward computations. The output to recompute is determined by the idx argument, which corresponds to the index of the output in the outputs list. If the block only has a single output, then idx will always be 0.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the output to compute. prepared (object): Anything returned by the prepare_recompute_component method. Default is None.
- Returns:
An object of the same type as block_variable.checkpoint which is determined by OverloadedType._ad_create_checkpoint (often the same as block_variable.saved_output): The new output.
- class firedrake.adjoint.blocks.FunctionMergeBlock(func, idx, ad_block_tag=None)[source]¶
-
- block_helper¶
- evaluate_adj_component(inputs, adj_inputs, block_variable, idx, prepared=None)[source]¶
This method should be overridden.
The method should implement a routine for evaluating the adjoint of the block that corresponds to one dependency. If one considers the adjoint action a vector right multiplied with the Jacobian matrix, then this method should return one entry in the resulting product, where the entry returned is decided by the argument idx.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. adj_inputs (list): A list of the adjoint input values, determined by the outputs list. block_variable (BlockVariable): The block variable of the dependency corresponding to index idx. idx (int): The index of the component to compute. prepared (object): Anything returned by the prepare_evaluate_adj method. Default is None.
- Returns:
An object of a type consistent with the adj_value type of block_variable: The resulting product.
- evaluate_hessian_component(inputs, hessian_inputs, adj_inputs, block_variable, idx, relevant_dependencies, prepared=None)[source]¶
This method must be overridden.
The method should implement a routine for evaluating the hessian of the block. It is preferable that a “Forward-over-Reverse” scheme is used. Thus the hessians are evaluated in reverse (starting with the last block on the tape).
- evaluate_tlm()[source]¶
Computes the tangent linear action and stores the result in the tlm_value attribute of the outputs.
This method will by default call the evaluate_tlm_component method for each output.
- Args:
- markings (bool): If True, then each block_variable will have set marked_in_path attribute indicating
whether their tlm components are relevant for computing the final target tlm values. Default is False.
- recompute_component(inputs, block_variable, idx, prepared)[source]¶
This method must be overridden.
The method should implement a routine for recomputing one output of the block in the forward computations. The output to recompute is determined by the idx argument, which corresponds to the index of the output in the outputs list. If the block only has a single output, then idx will always be 0.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the output to compute. prepared (object): Anything returned by the prepare_recompute_component method. Default is None.
- Returns:
An object of the same type as block_variable.checkpoint which is determined by OverloadedType._ad_create_checkpoint (often the same as block_variable.saved_output): The new output.
- class firedrake.adjoint.blocks.FunctionSplitBlock(func, idx, ad_block_tag=None)[source]¶
-
- block_helper¶
- evaluate_adj_component(inputs, adj_inputs, block_variable, idx, prepared=None)[source]¶
This method should be overridden.
The method should implement a routine for evaluating the adjoint of the block that corresponds to one dependency. If one considers the adjoint action a vector right multiplied with the Jacobian matrix, then this method should return one entry in the resulting product, where the entry returned is decided by the argument idx.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. adj_inputs (list): A list of the adjoint input values, determined by the outputs list. block_variable (BlockVariable): The block variable of the dependency corresponding to index idx. idx (int): The index of the component to compute. prepared (object): Anything returned by the prepare_evaluate_adj method. Default is None.
- Returns:
An object of a type consistent with the adj_value type of block_variable: The resulting product.
- evaluate_hessian_component(inputs, hessian_inputs, adj_inputs, block_variable, idx, relevant_dependencies, prepared=None)[source]¶
This method must be overridden.
The method should implement a routine for evaluating the hessian of the block. It is preferable that a “Forward-over-Reverse” scheme is used. Thus the hessians are evaluated in reverse (starting with the last block on the tape).
- evaluate_tlm_component(inputs, tlm_inputs, block_variable, idx, prepared=None)[source]¶
This method should be overridden.
The method should implement a routine for computing the tangent linear model of the block that corresponds to one output. If one considers the tangent linear action as a Jacobian matrix multiplied with a vector, then this method should return one entry in the resulting product, where the entry returned is decided by the argument idx.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. tlm_inputs (list): A list of the tlm input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the component to compute. prepared (object): Anything returned by the prepare_evaluate_tlm method. Default is None.
- Returns:
An object of the same type as block_variable.saved_output: The resulting product.
- recompute_component(inputs, block_variable, idx, prepared)[source]¶
This method must be overridden.
The method should implement a routine for recomputing one output of the block in the forward computations. The output to recompute is determined by the idx argument, which corresponds to the index of the output in the outputs list. If the block only has a single output, then idx will always be 0.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the output to compute. prepared (object): Anything returned by the prepare_recompute_component method. Default is None.
- Returns:
An object of the same type as block_variable.checkpoint which is determined by OverloadedType._ad_create_checkpoint (often the same as block_variable.saved_output): The new output.
- class firedrake.adjoint.blocks.GenericSolveBlock(lhs, rhs, func, bcs, *args, **kwargs)[source]¶
Bases:
GenericSolveBlock
,Backend
- block_helper¶
- recompute_component(inputs, block_variable, idx, prepared)[source]¶
This method must be overridden.
The method should implement a routine for recomputing one output of the block in the forward computations. The output to recompute is determined by the idx argument, which corresponds to the index of the output in the outputs list. If the block only has a single output, then idx will always be 0.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the output to compute. prepared (object): Anything returned by the prepare_recompute_component method. Default is None.
- Returns:
An object of the same type as block_variable.checkpoint which is determined by OverloadedType._ad_create_checkpoint (often the same as block_variable.saved_output): The new output.
- class firedrake.adjoint.blocks.InterpolateBlock(interpolator, *functions, **kwargs)[source]¶
-
Annotates an interpolator.
Consider the block as f with 1 forward model output
v
, and inputsu
andg
(there can, in principle, be any number of outputs). The adjoint input isvhat
(uhat
andghat
are adjoints tou
andv
respectively and are shown for completeness). The downstream block isJ
which has inputv
._ _ |J|--<--v--<--|f|--<--u--<--... ¯ | ¯ | vhat | uhat | ---<--g--<--... | ghat
(Arrows indicate forward model direction)
J : V ⟶ R i.e. J(v) ∈ R ∀ v ∈ V
Interpolation can operate on an expression which may not be linear in its arguments.
f : W × G ⟶ V i.e. f(u, g) ∈ V ∀ u ∈ W and g ∈ G. f = I ∘ expr I : X ⟶ V i.e. I(;x) ∈ V ∀ x ∈ X. X is infinite dimensional. expr: W × G ⟶ X i.e. expr(u, g) ∈ X ∀ u ∈ W and g ∈ G.
Arguments after a semicolon are linear (i.e. operation I is linear)
- block_helper¶
- evaluate_adj_component(inputs, adj_inputs, block_variable, idx, prepared=None)[source]¶
Denote
d_u[A]
as the gateaux derivative in theu
direction. Arguments after a semicolon are linear.This calculates
uhat = vhat ⋅ d_u[f](u, g; ⋅) (for inputs[idx] ∈ W) or ghat = vhat ⋅ d_g[f](u, g; ⋅) (for inputs[idx] ∈ G)
where
inputs[idx]
specifies the derivative direction,vhat
isadj_inputs[0]
(since we assume only one block output) and⋅
denotes an unspecified operand ofu'
(forinputs[idx]
∈W
) org'
(forinputs[idx]
∈G
) size (vhat
left multiplies the derivative).f = I ∘ expr : W × G ⟶ V i.e. I(expr|_{u, g}) ∈ V ∀ u ∈ W, g ∈ G.
Since
I
is linear we get thatd_u[I ∘ expr](u, g; u') = I ∘ d_u[expr](u|_u, g|_g; u') d_g[I ∘ expr](u, g; g') = I ∘ d_u[expr](u|_u, g|_g; g').
In tensor notation
uhat_q^T = vhat_p^T I([dexpr/du|_u]_q)_p or ghat_q^T = vhat_p^T I([dexpr/dg|_u]_q)_p
the output is then
uhat_q = I^T([dexpr/du|_u]_q)_p vhat_p or ghat_q = I^T([dexpr/dg|_u]_q)_p vhat_p.
- evaluate_hessian_component(inputs, hessian_inputs, adj_inputs, block_variable, idx, relevant_dependencies, prepared=None)[source]¶
Denote
d_u[A]
as the gateaux derivative in theu
direction. Arguments after a semicolon are linear.hessian_input is
d_v[d_v[J]](v; v', ⋅)
where the direction⋅
is left unspecified so it can be operated upon.Warning
NOTE: This comment describes the implementation of 1 block input
u
. (e.g. interpolating from an expression with 1 coefficient). Explaining how this works for multiple block inputs (e.g.u
andg
) is currently too complicated for the author to do succinctly!This function needs to output
d_u[d_u[J ∘ f]](u; u', ⋅)
where the direction⋅
will be specified in another function and multiplied on the right with the output of this function. We will calculate this using the chain rule.J : V ⟶ R i.e. J(v) ∈ R ∀ v ∈ V f = I ∘ expr : W ⟶ V J ∘ f : W ⟶ R i.e. J(f|u) ∈ R ∀ u ∈ V. d_u[J ∘ f] : W × W ⟶ R i.e. d_u[J ∘ f](u; u') d_u[d_u[J ∘ f]] : W × W × W ⟶ R i.e. d_u[d_u[J ∘ f]](u; u', u'') d_v[J] : V × V ⟶ R i.e. d_v[J](v; v') d_v[d_v[J]] : V × V × V ⟶ R i.e. d_v[d_v[J]](v; v', v'')
Chain rule:
d_u[J ∘ f](u; u') = d_v[J](v = f|u; v' = d_u[f](u; u'))
Multivariable chain rule:
d_u[d_u[J ∘ f]](u; u', u'') = d_v[d_v[J]](v = f|u; v' = d_u[f](u; u'), v'' = d_u[f](u; u'')) + d_v'[d_v[J]](v = f|u; v' = d_u[f](u; u'), v'' = d_u[d_u[f]](u; u', u'')) = d_v[d_v[J]](v = f|u; v' = d_u[f](u; u'), v''=d_u[f](u; u'')) + d_v[J](v = f|u; v' = v'' = d_u[d_u[f]](u; u', u''))
since
d_v[d_v[J]]
is linear inv'
so differentiating wrt to it leaves its coefficient, the bare d_v[J] operator which acts on thev''
term that remains.The
d_u[d_u[f]](u; u', u'')
term can be simplified further:f = I ∘ expr : W ⟶ V i.e. I(expr|u) ∈ V ∀ u ∈ W d_u[I ∘ expr] : W × W ⟶ V i.e. d_u[I ∘ expr](u; u') d_u[d_u[I ∘ expr]] : W × W × W ⟶ V i.e. d_u[I ∘ expr](u; u', u'') d_x[I] : X × X ⟶ V i.e. d_x[I](x; x') d_x[d_x[I]] : X × X × X ⟶ V i.e. d_x[d_x[I]](x; x', x'') d_u[expr] : W × W ⟶ X i.e. d_u[expr](u; u') d_u[d_u[expr]] : W × W × W ⟶ X i.e. d_u[d_u[expr]](u; u', u'')
Since
I
is linear we get thatd_u[d_u[I ∘ expr]](u; u', u'') = I ∘ d_u[d_u[expr]](u; u', u'').
So our full hessian is:
d_u[d_u[J ∘ f]](u; u', u'') = d_v[d_v[J]](v = f|u; v' = d_u[f](u; u'), v''=d_u[f](u; u'')) + d_v[J](v = f|u; v' = v'' = d_u[d_u[f]](u; u', u''))
In tensor notation
[d^2[J ∘ f]/du^2|_u]_{lk} u'_k u''_k = [d^2J/dv^2|_{v=f|_u}]_{ij} [df/du|_u]_{jk} u'_k [df/du|_u]_{il} u''_l + [dJ/dv|_{v=f_u}]_i I([d^2expr/du^2|_u]_{lk} u'_k)_i u''_l
In the first term:
[df/du|_u]_{jk} u'_k = v'_j => [d^2J/dv^2|_{v=f|_u}]_{ij} [df/du|_u]_{jk} u'_k = [d^2J/dv^2|_{v=f|_u}]_{ij} v'_j = hessian_input_i => [d^2J/dv^2|_{v=f|_u}]_{ij} [df/du|_u]_{jk} u'_k [df/du|_u]_{il} = hessian_input_i [df/du|_u]_{il} = self.evaluate_adj_component(inputs, hessian_inputs, ...)_l
In the second term we calculate everything explicitly though note
[dJ/dv|_{v=f_u}]_i = adj_inputs[0]_i
Also, the second term is 0 if
expr
is linear.
- evaluate_tlm_component(inputs, tlm_inputs, block_variable, idx, prepared=None)[source]¶
Denote
d_u[A]
as the gateaux derivative in theu
direction. Arguments after a semicolon are linear.For a block with two inputs this calculates
v' = d_u[f](u, g; u') + d_g[f](u, g; g')
where
u' = tlm_inputs[0]
andg = tlm_inputs[1]
.f = I ∘ expr : W × G ⟶ V i.e. I(expr|_{u, g}) ∈ V ∀ u ∈ W, g ∈ G.
Since
I
is linear we get thatd_u[I ∘ expr](u, g; u') = I ∘ d_u[expr](u|_u, g|_g; u') d_g[I ∘ expr](u, g; g') = I ∘ d_u[expr](u|_u, g|_g; g').
In tensor notation the output is then
v'_l = I([dexpr/du|_{u,g}]_k u'_k)_l + I([dexpr/du|_{u,g}]_k g'_k)_l = I([dexpr/du|_{u,g}]_k u'_k + [dexpr/du|_{u,g}]_k g'_k)_l
since
I
is linear.
- prepare_evaluate_adj(inputs, adj_inputs, relevant_outputs)[source]¶
Runs preparations before evalute_adj_component is ran.
The return value is supplied to each of the subsequent evaluate_adj_component calls. This method is intended to be overridden for blocks that require such preparations, by default there is none.
- Args:
inputs: The values of the inputs adj_inputs: The adjoint inputs relevant_dependencies: A list of the relevant block variables for evaluate_adj_component.
- Returns:
Anything. The returned value is supplied to evaluate_adj_component
- prepare_evaluate_hessian(inputs, hessian_inputs, adj_inputs, relevant_dependencies)[source]¶
Runs preparations before evalute_hessian_component is ran for each relevant dependency.
The return value is supplied to each of the subsequent evaluate_hessian_component calls. This method is intended to be overridden for blocks that require such preparations, by default there is none.
- Args:
inputs: The values of the inputs hessian_inputs: The hessian inputs adj_inputs: The adjoint inputs relevant_dependencies: A list of the relevant block variables for evaluate_hessian_component.
- Returns:
Anything. The returned value is supplied to evaluate_hessian_component
- prepare_evaluate_tlm(inputs, tlm_inputs, relevant_outputs)[source]¶
Runs preparations before evalute_tlm_component is ran.
The return value is supplied to each of the subsequent evaluate_tlm_component calls. This method is intended to be overridden for blocks that require such preparations, by default there is none.
- Args:
inputs: The values of the inputs tlm_inputs: The tlm inputs relevant_outputs: A list of the relevant block variables for evaluate_tlm_component.
- Returns:
Anything. The returned value is supplied to evaluate_tlm_component
- prepare_recompute_component(inputs, relevant_outputs)[source]¶
Runs preparations before recompute_component is ran.
The return value is supplied to each of the subsequent recompute_component calls. This method is intended to be overridden for blocks that require such preparations, by default there is none.
- Args:
inputs: The values of the inputs relevant_outputs: A list of the relevant block variables for recompute_component.
- Returns:
Anything. The returned value is supplied to recompute_component
- recompute_component(inputs, block_variable, idx, prepared)[source]¶
This method must be overridden.
The method should implement a routine for recomputing one output of the block in the forward computations. The output to recompute is determined by the idx argument, which corresponds to the index of the output in the outputs list. If the block only has a single output, then idx will always be 0.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the output to compute. prepared (object): Anything returned by the prepare_recompute_component method. Default is None.
- Returns:
An object of the same type as block_variable.checkpoint which is determined by OverloadedType._ad_create_checkpoint (often the same as block_variable.saved_output): The new output.
- class firedrake.adjoint.blocks.MeshInputBlock(mesh, ad_block_tag=None)[source]¶
Bases:
Block
Block which links a MeshGeometry to its coordinates, which is a firedrake function.
- block_helper¶
- evaluate_adj_component(inputs, adj_inputs, block_variable, idx, prepared=None)[source]¶
This method should be overridden.
The method should implement a routine for evaluating the adjoint of the block that corresponds to one dependency. If one considers the adjoint action a vector right multiplied with the Jacobian matrix, then this method should return one entry in the resulting product, where the entry returned is decided by the argument idx.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. adj_inputs (list): A list of the adjoint input values, determined by the outputs list. block_variable (BlockVariable): The block variable of the dependency corresponding to index idx. idx (int): The index of the component to compute. prepared (object): Anything returned by the prepare_evaluate_adj method. Default is None.
- Returns:
An object of a type consistent with the adj_value type of block_variable: The resulting product.
- evaluate_hessian_component(inputs, hessian_inputs, adj_inputs, idx, block_variable, relevant_dependencies, prepared=None)[source]¶
This method must be overridden.
The method should implement a routine for evaluating the hessian of the block. It is preferable that a “Forward-over-Reverse” scheme is used. Thus the hessians are evaluated in reverse (starting with the last block on the tape).
- evaluate_tlm_component(inputs, tlm_inputs, block_variable, idx, prepared=None)[source]¶
This method should be overridden.
The method should implement a routine for computing the tangent linear model of the block that corresponds to one output. If one considers the tangent linear action as a Jacobian matrix multiplied with a vector, then this method should return one entry in the resulting product, where the entry returned is decided by the argument idx.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. tlm_inputs (list): A list of the tlm input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the component to compute. prepared (object): Anything returned by the prepare_evaluate_tlm method. Default is None.
- Returns:
An object of the same type as block_variable.saved_output: The resulting product.
- recompute_component(inputs, block_variable, idx, prepared)[source]¶
This method must be overridden.
The method should implement a routine for recomputing one output of the block in the forward computations. The output to recompute is determined by the idx argument, which corresponds to the index of the output in the outputs list. If the block only has a single output, then idx will always be 0.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the output to compute. prepared (object): Anything returned by the prepare_recompute_component method. Default is None.
- Returns:
An object of the same type as block_variable.checkpoint which is determined by OverloadedType._ad_create_checkpoint (often the same as block_variable.saved_output): The new output.
- class firedrake.adjoint.blocks.MeshOutputBlock(func, mesh, ad_block_tag=None)[source]¶
Bases:
Block
Block which is called when the coordinates of a mesh are changed.
- block_helper¶
- evaluate_adj_component(inputs, adj_inputs, block_variable, idx, prepared=None)[source]¶
This method should be overridden.
The method should implement a routine for evaluating the adjoint of the block that corresponds to one dependency. If one considers the adjoint action a vector right multiplied with the Jacobian matrix, then this method should return one entry in the resulting product, where the entry returned is decided by the argument idx.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. adj_inputs (list): A list of the adjoint input values, determined by the outputs list. block_variable (BlockVariable): The block variable of the dependency corresponding to index idx. idx (int): The index of the component to compute. prepared (object): Anything returned by the prepare_evaluate_adj method. Default is None.
- Returns:
An object of a type consistent with the adj_value type of block_variable: The resulting product.
- evaluate_hessian_component(inputs, hessian_inputs, adj_inputs, idx, block_variable, relevant_dependencies, prepared=None)[source]¶
This method must be overridden.
The method should implement a routine for evaluating the hessian of the block. It is preferable that a “Forward-over-Reverse” scheme is used. Thus the hessians are evaluated in reverse (starting with the last block on the tape).
- evaluate_tlm_component(inputs, tlm_inputs, block_variable, idx, prepared=None)[source]¶
This method should be overridden.
The method should implement a routine for computing the tangent linear model of the block that corresponds to one output. If one considers the tangent linear action as a Jacobian matrix multiplied with a vector, then this method should return one entry in the resulting product, where the entry returned is decided by the argument idx.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. tlm_inputs (list): A list of the tlm input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the component to compute. prepared (object): Anything returned by the prepare_evaluate_tlm method. Default is None.
- Returns:
An object of the same type as block_variable.saved_output: The resulting product.
- recompute_component(inputs, block_variable, idx, prepared)[source]¶
This method must be overridden.
The method should implement a routine for recomputing one output of the block in the forward computations. The output to recompute is determined by the idx argument, which corresponds to the index of the output in the outputs list. If the block only has a single output, then idx will always be 0.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the output to compute. prepared (object): Anything returned by the prepare_recompute_component method. Default is None.
- Returns:
An object of the same type as block_variable.checkpoint which is determined by OverloadedType._ad_create_checkpoint (often the same as block_variable.saved_output): The new output.
- class firedrake.adjoint.blocks.NonlinearVariationalSolveBlock(equation, func, bcs, adj_F, dFdm_cache, problem_J, solver_params, solver_kwargs, **kwargs)[source]¶
Bases:
GenericSolveBlock
- block_helper¶
- evaluate_adj_component(inputs, adj_inputs, block_variable, idx, prepared=None)[source]¶
This method should be overridden.
The method should implement a routine for evaluating the adjoint of the block that corresponds to one dependency. If one considers the adjoint action a vector right multiplied with the Jacobian matrix, then this method should return one entry in the resulting product, where the entry returned is decided by the argument idx.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. adj_inputs (list): A list of the adjoint input values, determined by the outputs list. block_variable (BlockVariable): The block variable of the dependency corresponding to index idx. idx (int): The index of the component to compute. prepared (object): Anything returned by the prepare_evaluate_adj method. Default is None.
- Returns:
An object of a type consistent with the adj_value type of block_variable: The resulting product.
- prepare_evaluate_adj(inputs, adj_inputs, relevant_dependencies)[source]¶
Runs preparations before evalute_adj_component is ran.
The return value is supplied to each of the subsequent evaluate_adj_component calls. This method is intended to be overridden for blocks that require such preparations, by default there is none.
- Args:
inputs: The values of the inputs adj_inputs: The adjoint inputs relevant_dependencies: A list of the relevant block variables for evaluate_adj_component.
- Returns:
Anything. The returned value is supplied to evaluate_adj_component
- class firedrake.adjoint.blocks.ProjectBlock(v, V, output, bcs=[], *args, **kwargs)[source]¶
Bases:
SolveVarFormBlock
- block_helper¶
- class firedrake.adjoint.blocks.SolveLinearSystemBlock(A, u, b, *args, **kwargs)[source]¶
Bases:
GenericSolveBlock
- block_helper¶
- class firedrake.adjoint.blocks.SolveVarFormBlock(equation, func, bcs=[], *args, **kwargs)[source]¶
Bases:
GenericSolveBlock
- block_helper¶
- class firedrake.adjoint.blocks.SupermeshProjectBlock(source, target_space, target, bcs=[], **kwargs)[source]¶
-
Annotates supermesh projection.
Suppose we have a source space, \(V_A\), and a target space, \(V_B\). Projecting a source from \(V_A\) to \(V_B\) amounts to solving the linear system
\[ \begin{align}\begin{aligned}M_B * v_B = M_{AB} * v_A,\\where :math:`M_B` is the mass matrix on :math:`V_B`, :math:`M_{AB}` is the mixed mass matrix for :math:`V_A` and :math:`V_B` and :math:`v_A` and :math:`v_B` are vector representations of the source and target :class:`.Function`s.\\This can be broken into two steps: Step 1. form RHS, multiplying the source with the mixed mass matrix; Step 2. solve linear system.\end{aligned}\end{align} \]- block_helper¶
- evaluate_adj_component(inputs, adj_inputs, block_variable, idx, prepared=None)[source]¶
- Recall that the forward propagation can be broken down as
Step 1. multiply \(w := M_{AB} * v_A\); Step 2. solve \(M_B * v_B = w\).
- For a seed vector \(v_B^{seed}\) from the target space, the adjoint is given by
Adjoint of step 2. solve \(M_B^T * w = v_B^{seed}\) for w; Adjoint of step 1. multiply \(v_A^{adj} := M_{AB}^T * w\).
- evaluate_hessian_component(inputs, hessian_inputs, adj_inputs, block_variable, idx, relevant_dependencies, prepared=None)[source]¶
This method must be overridden.
The method should implement a routine for evaluating the hessian of the block. It is preferable that a “Forward-over-Reverse” scheme is used. Thus the hessians are evaluated in reverse (starting with the last block on the tape).
- evaluate_tlm_component(inputs, tlm_inputs, block_variable, idx, prepared=None)[source]¶
Given that the input is a Function, we just have a linear operation. As such, the tlm is just the sum of each tlm input projected into the target space.
- recompute_component(inputs, block_variable, idx, prepared)[source]¶
This method must be overridden.
The method should implement a routine for recomputing one output of the block in the forward computations. The output to recompute is determined by the idx argument, which corresponds to the index of the output in the outputs list. If the block only has a single output, then idx will always be 0.
- Args:
inputs (list): A list of the saved input values, determined by the dependencies list. block_variable (BlockVariable): The block variable of the output corresponding to index idx. idx (int): The index of the output to compute. prepared (object): Anything returned by the prepare_recompute_component method. Default is None.
- Returns:
An object of the same type as block_variable.checkpoint which is determined by OverloadedType._ad_create_checkpoint (often the same as block_variable.saved_output): The new output.
firedrake.adjoint.checkpointing module¶
A module providing support for disk checkpointing of the adjoint tape.
- firedrake.adjoint.checkpointing.checkpointable_mesh(mesh)[source]¶
Write a mesh to disk and read it back.
Since a mesh will be repartitioned by being written to disk and reread, only meshes read from a checkpoint file are safe to use with disk checkpointing.
The workflow for disk checkpointing is therefore to create the mesh(es) required, and then call this function on them. Only the mesh(es) returned by this function can be used in disk checkpointing.
- Parameters:
- meshfiredrake.Mesh
The mesh to be checkpointed.
- Returns:
- firedrake.Mesh
The checkpointed mesh to be used in the rest of the computation.
- firedrake.adjoint.checkpointing.disk_checkpointing()[source]¶
Return true if disk checkpointing of the adjoint tape is active.
- firedrake.adjoint.checkpointing.enable_disk_checkpointing(dirname=None, comm=<mpi4py.MPI.Intracomm object>, cleanup=True)[source]¶
Add a DiskCheckpointer to the current tape and switch on disk checkpointing.
- Parameters:
- dirnamestr
The directory in which the disk checkpoints should be stored. If not specified then the current working directory is used. Checkpoints are stored in a temporary subdirectory of this directory.
- commmpi4py.MPI.Intracomm
The MPI communicator over which the computation to be disk checkpointed is defined. This will usually match the communicator on which the mesh(es) are defined.
- cleanupbool
If set to False, checkpoint files will not be deleted when no longer required. This is usually only useful for debugging.
firedrake.adjoint.constant module¶
- class firedrake.adjoint.constant.ConstantMixin(*args, **kwargs)[source]¶
Bases:
OverloadedType
firedrake.adjoint.dirichletbc module¶
firedrake.adjoint.function module¶
- class firedrake.adjoint.function.DelegatedFunctionCheckpoint(other)[source]¶
Bases:
CheckpointBase
A wrapper which delegates the checkpoint of this Function to another Function.
This enables us to avoid checkpointing a Function twice when it is copied.
- Parameters:
- other: BlockVariable
The block variable to which we delegate checkpointing.
firedrake.adjoint.interpolate module¶
firedrake.adjoint.mesh module¶
- class firedrake.adjoint.mesh.MeshGeometryMixin(*args, **kwargs)[source]¶
Bases:
OverloadedType
firedrake.adjoint.projection module¶
firedrake.adjoint.solving module¶
- firedrake.adjoint.solving.annotate_solve(solve)[source]¶
This solve routine wraps the Firedrake
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
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).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 (string, options): tag used to label the resulting block on the Pyadjoint tape. This
is useful for identifying which block is associated with which equation in the forward model.