gemseo / core / derivatives

Hide inherited members

jacobian_assembly module

Coupled derivatives calculations.

class gemseo.core.derivatives.jacobian_assembly.AssembledJacobianOperator(*args, **kwargs)[source]

Bases: LinearOperator

Representation of the assembled Jacobian as a SciPy LinearOperator.

Parameters:

  • functions (Iterable[str]) – The functions to differentiate.

  • variables (Iterable[str]) – The differentiation variables.

  • n_functions (int) – The number of functions components.

  • n_variables (int) – The number of variables components.

  • get_jacobian_generator (Callable[[Iterable[str], Iterable[str], bool], Iterator]) – The method to iterate over the Jacobians associated with the provided functions and variables.

  • is_residual (bool) –

    Whether the functions are residuals.

    By default it is set to False.

adjoint()

Hermitian adjoint.

Returns the Hermitian adjoint of self, aka the Hermitian conjugate or Hermitian transpose. For a complex matrix, the Hermitian adjoint is equal to the conjugate transpose.

Can be abbreviated self.H instead of self.adjoint().

Returns:

A_H – Hermitian adjoint of self.

Return type:

LinearOperator

dot(x)

Matrix-matrix or matrix-vector multiplication.

Parameters:

x (array_like) – 1-d or 2-d array, representing a vector or matrix.

Returns:

Ax – 1-d or 2-d array (depending on the shape of x) that represents the result of applying this linear operator on x.

Return type:

array

matmat(X)

Matrix-matrix multiplication.

Performs the operation y=A*X where A is an MxN linear operator and X dense N*K matrix or ndarray.

Parameters:

X ({matrix, ndarray}) – An array with shape (N,K).

Returns:

Y – A matrix or ndarray with shape (M,K) depending on the type of the X argument.

Return type:

{matrix, ndarray}

Notes

This matmat wraps any user-specified matmat routine or overridden _matmat method to ensure that y has the correct type.

matvec(x)

Matrix-vector multiplication.

Performs the operation y=A*x where A is an MxN linear operator and x is a column vector or 1-d array.

Parameters:

x ({matrix, ndarray}) – An array with shape (N,) or (N,1).

Returns:

y – A matrix or ndarray with shape (M,) or (M,1) depending on the type and shape of the x argument.

Return type:

{matrix, ndarray}

Notes

This matvec wraps the user-specified matvec routine or overridden _matvec method to ensure that y has the correct shape and type.

rmatmat(X)

Adjoint matrix-matrix multiplication.

Performs the operation y = A^H * x where A is an MxN linear operator and x is a column vector or 1-d array, or 2-d array. The default implementation defers to the adjoint.

Parameters:

X ({matrix, ndarray}) – A matrix or 2D array.

Returns:

Y – A matrix or 2D array depending on the type of the input.

Return type:

{matrix, ndarray}

Notes

This rmatmat wraps the user-specified rmatmat routine.

rmatvec(x)

Adjoint matrix-vector multiplication.

Performs the operation y = A^H * x where A is an MxN linear operator and x is a column vector or 1-d array.

Parameters:

x ({matrix, ndarray}) – An array with shape (M,) or (M,1).

Returns:

y – A matrix or ndarray with shape (N,) or (N,1) depending on the type and shape of the x argument.

Return type:

{matrix, ndarray}

Notes

This rmatvec wraps the user-specified rmatvec routine or overridden _rmatvec method to ensure that y has the correct shape and type.

transpose()

Transpose this linear operator.

Returns a LinearOperator that represents the transpose of this one. Can be abbreviated self.T instead of self.transpose().

property H

Hermitian adjoint.

Returns the Hermitian adjoint of self, aka the Hermitian conjugate or Hermitian transpose. For a complex matrix, the Hermitian adjoint is equal to the conjugate transpose.

Can be abbreviated self.H instead of self.adjoint().

Returns:

A_H – Hermitian adjoint of self.

Return type:

LinearOperator

property T

Transpose this linear operator.

Returns a LinearOperator that represents the transpose of this one. Can be abbreviated self.T instead of self.transpose().

ndim = 2
class gemseo.core.derivatives.jacobian_assembly.CoupledSystem[source]

Bases: object

Compute coupled (total) derivatives of a system of residuals.

Use several methods:

  • direct or adjoint

  • factorized for multiple RHS

adjoint_mode(functions, dres_dx, dres_dy_t, dfun_dx, dfun_dy, linear_solver='DEFAULT', use_lu_fact=False, **linear_solver_options)[source]

Compute the total derivative Jacobian in adjoint mode.

Parameters:

  • functions (Iterable[str]) – The functions to differentiate.

  • dres_dx (dok_matrix | LinearOperator) – The Jacobian of the residuals wrt the design variables.

  • dres_dy_t (dok_matrix | LinearOperator) – The Jacobian of the residuals wrt the coupling variables.

  • dfun_dx (Mapping[str, dok_matrix]) – The Jacobian of the functions wrt the design variables.

  • dfun_dy (Mapping[str, dok_matrix]) – The Jacobian of the functions wrt the coupling variables.

  • linear_solver (str) –

    The name of the linear solver.

    By default it is set to “DEFAULT”.

  • use_lu_fact (bool) –

    Whether to factorize dres_dy_t once.

    By default it is set to False.

  • **linear_solver_options (Any) – The optional parameters.

Returns:

The Jacobian of total coupled derivatives.

Return type:

dict[str, ndarray]

direct_mode(functions, n_variables, n_couplings, dres_dx, dres_dy, dfun_dx, dfun_dy, linear_solver='DEFAULT', use_lu_fact=False, **linear_solver_options)[source]

Compute the total derivative Jacobian in direct mode.

Parameters:

  • functions (Iterable[str]) – The functions to differentiate.

  • n_variables (int) – The number of variables.

  • n_couplings (int) – The number of couplings.

  • dres_dx (dok_matrix | LinearOperator) – The Jacobian of the residuals wrt the design variables.

  • dres_dy (dok_matrix | LinearOperator) – The Jacobian of the residuals wrt the coupling variables.

  • dfun_dx (Mapping[str, dok_matrix]) – The Jacobian of the functions wrt the design variables.

  • dfun_dy (Mapping[str, dok_matrix]) – The Jacobian of the functions wrt the coupling variables.

  • linear_solver (str) –

    The name of the linear solver.

    By default it is set to “DEFAULT”.

  • use_lu_fact (bool) –

    Whether to factorize dres_dy once.

    By default it is set to False.

  • **linear_solver_options (Any) – The optional parameters.

Returns:

The Jacobian of the total coupled derivatives.

Return type:

dict[str, dok_matrix]

DEFAULT_LINEAR_SOLVER: ClassVar[str] = 'DEFAULT'

The default linear solver.

linear_problem: LinearProblem | None

The considered linear problem.

lu_fact: int

The number of LU mode calls (adjoint or direct).

n_adjoint_modes: int

The number of adjoint mode calls.

n_direct_modes: int

The number of direct mode calls.

n_linear_resolutions: int

The number of linear resolutions.

class gemseo.core.derivatives.jacobian_assembly.JacobianAssembly(coupling_structure)[source]

Bases: object

Assembly of Jacobians.

Typically, assemble discipline’s Jacobians into a system Jacobian.

Parameters:

coupling_structure (MDOCouplingStructure) – The MDOCouplingStructure associated disciplines that form the coupled system.

class DerivationMode(value)

Bases: StrEnum

The derivation modes.

ADJOINT = 'adjoint'

The adjoint resolution mode for MDAs, solves one system per output.

AUTO = 'auto'

Automatic switch between direct, reverse or adjoint depending on data sizes.

DIRECT = 'direct'

The direct Jacobian accumulation, chain rule from inputs to outputs, or derivation of an MDA that solves one system per input.

REVERSE = 'reverse'

The reverse Jacobian accumulation, chain rule from outputs to inputs.

class JacobianPosition(row_slice, column_slice, row_index, column_index)[source]

Bases: NamedTuple

The position of the discipline’s Jacobians within the assembled Jacobian.

Create new instance of JacobianPosition(row_slice, column_slice, row_index, column_index)

Parameters:

  • row_slice (slice) –

  • column_slice (slice) –

  • row_index (int) –

  • column_index (int) –

count(value, /)

Return number of occurrences of value.

index(value, start=0, stop=9223372036854775807, /)

Return first index of value.

Raises ValueError if the value is not present.

column_index: int

The column index of the disciplinary Jacobian within the assembled Jacobian when defined blockwise.

column_slice: slice

The column slice indicating where to position the disciplinary Jacobian within the assembled Jacobian when defined as an array.

row_index: int

The row index of the disciplinary Jacobian within the assembled Jacobian when defined blockwise.

row_slice: slice

The row slice indicating where to position the disciplinary Jacobian within the assembled Jacobian when defined as an array.

class JacobianType(value)[source]

Bases: StrEnum

The available types for the Jacobian matrix.

LINEAR_OPERATOR = 'linear_operator'

Jacobian as a SciPy LinearOperator implementing the appropriate method to perform matrix-vector products.

MATRIX = 'matrix'

Jacobian matrix in Compressed Sparse Row (CSR) format.

assemble_jacobian(functions, variables, is_residual=False, jacobian_type=JacobianType.MATRIX)[source]

Form the Jacobian as a SciPy LinearOperator.

Parameters:

  • functions (Collection[str]) – The functions to differentiate.

  • variables (Collection[str]) – The differentiation variables.

  • is_residual (bool) –

    Whether the functions are residuals.

    By default it is set to False.

  • jacobian_type (JacobianType) –

    The type of representation for the Jacobian ∂f/∂v.

    By default it is set to “matrix”.

Returns:

The Jacobian ∂f/∂v in the specified type.

Return type:

csr_matrix | AssembledJacobianOperator

compute_dimension(names)[source]

Compute the total number of functions/variables/couplings of the full system.

Parameters:

names (Iterable[str]) – The names of the inputs or the outputs.

Returns:

The dimension if the system.

Return type:

int

compute_newton_step(in_data, couplings, linear_solver='DEFAULT', matrix_type=JacobianType.MATRIX, residuals=None, resolved_residual_names=(), **linear_solver_options)[source]

Compute the Newton step for the coupled system of disciplines residuals.

Parameters:

  • in_data (Mapping[str, NDArray[float]]) – The input data.

  • couplings (Collection[str]) – The coupling variables.

  • linear_solver (str) –

    The name of the linear solver.

    By default it is set to “DEFAULT”.

  • matrix_type (JacobianType) –

    The representation of the matrix ∂R/∂y (sparse or linear operator).

    By default it is set to “matrix”.

  • residuals (ndarray | None) – The residuals vector, if None use residuals.

  • resolved_residual_names (Collection[str]) –

    The names of residual variables.

    By default it is set to ().

  • **linear_solver_options (Any) – The options passed to the linear solver factory.

Returns:

The Newton step - relax_factor . [∂R/∂y]^-1 . R as an array of steps for which the order is given by the couplings argument. Whether the linear solver converged.

Return type:

tuple[ndarray, bool]

compute_sizes(functions, variables, couplings, residual_variables=None)[source]

Compute the number of scalar functions, variables and couplings.

Parameters:

  • functions (Iterable[str]) – The functions to differentiate.

  • variables (Iterable[str]) – The differentiation variables.

  • couplings (Iterable[str]) – The coupling variables.

  • residual_variables (Mapping[str, str] | None) – The mapping of residuals of disciplines to their respective state variables.

Raises:

ValueError – When the size of some variables could not be determined.

Return type:

None

plot_dependency_jacobian(functions, variables, save=True, show=False, filepath=None, markersize=None)[source]

Plot the Jacobian matrix.

Nonzero elements of the sparse matrix are represented by blue squares.

Parameters:

  • functions (Collection[str]) – The functions to plot.

  • variables (Collection[str]) – The variables.

  • show (bool) –

    Whether the plot is displayed.

    By default it is set to False.

  • save (bool) –

    Whether the plot is saved in a PDF file.

    By default it is set to True.

  • filepath (str | None) – The file name to save to. If None, coupled_jacobian.pdf is used, otherwise coupled_jacobian_ + filepath + .pdf.

  • markersize (float | None) – size of the markers

Returns:

The file name.

Return type:

str

residuals(in_data, var_names)[source]

Form the matrix of residuals wrt coupling variables.

Given disciplinary explicit calculations Yi(Y0_t,…Yn_t), fill the residual matrix:

[Y0(Y0_t,...Yn_t) - Y0_t]
[                       ]
[Yn(Y0_t,...Yn_t) - Yn_t]
Parameters:

  • in_data (Mapping[str, Any]) – The values prescribed for the calculation of the residuals (Y0_t,…Yn_t).

  • var_names (Iterable[str]) – The names of variables associated with the residuals (R).

Returns:

The residuals array.

Return type:

ndarray

set_newton_differentiated_ios(couplings)[source]

Set the differentiated inputs and outputs for the Newton algorithm.

Also ensures that JacobianAssembly.sizes contains the sizes of all the coupling sizes needed for Newton.

Parameters:

couplings (Collection[str]) – The coupling variables.

Return type:

None

split_jac(coupled_system, variables)[source]

Split a Jacobian dict into a dict of dict.

Parameters:

  • coupled_system (Mapping[str, ndarray | dok_matrix]) – The derivatives to split.

  • variables (Iterable[str]) – The variables wrt which the differentiation is performed.

Returns:

The Jacobian.

Return type:

dict[str, ndarray | dok_matrix]

total_derivatives(in_data, functions, variables, couplings, linear_solver='DEFAULT', mode=DerivationMode.AUTO, matrix_type=JacobianType.MATRIX, use_lu_fact=False, exec_cache_tol=None, execute=True, residual_variables=None, **linear_solver_options)[source]

Compute the Jacobian of total derivatives of the coupled system.

Parameters:

  • in_data – The input data dict.

  • functions (Collection[str]) – The functions to differentiate.

  • variables (Collection[str]) – The differentiation variables.

  • couplings (Iterable[str]) – The coupling variables.

  • linear_solver (str) –

    The name of the linear solver.

    By default it is set to “DEFAULT”.

  • mode (DerivationMode) –

    The linearization mode (auto, direct or adjoint).

    By default it is set to “auto”.

  • matrix_type (JacobianType) –

    The representation of the matrix ∂R/∂y (sparse or linear operator).

    By default it is set to “matrix”.

  • use_lu_fact (bool) –

    Whether to factorize dres_dy once, unsupported for linear operator mode.

    By default it is set to False.

  • exec_cache_tol (float | None) – The discipline cache tolerance to when calling the linearize method. If None, no tolerance is set (equivalent to tol=0.0).

  • execute (bool) –

    Whether to start by executing the discipline with the input data for which to compute the Jacobian; this allows to ensure that the discipline was executed with the right input data; it can be almost free if the corresponding output data have been stored in the MDODiscipline.cache.

    By default it is set to True.

  • linear_solver_options – The options passed to the linear solver factory.

  • residual_variables (Mapping[str, str] | None) – a mapping of residuals of disciplines to their respective state variables.

  • **linear_solver_options (Any) – The options passed to the linear solver factory.

Returns:

The total coupled derivatives.

Raises:

ValueError – When the linearization_mode is incorrect.

Return type:

dict[str, dict[str, ndarray]] | dict[Any, dict[Any, None]]

N_CPUS: Final[int] = 2

The number of available CPUs.

coupled_system: CoupledSystem

The coupled derivative system of residuals.

coupling_structure: MDOCouplingStructure

The considered coupling structure.

disciplines: dict[str, MDODiscipline]

The MDODisciplines, stored using their name.

sizes: dict[str, int]

The number of elements of a given str.

gemseo.core.derivatives.jacobian_assembly.default_dict_factory()[source]

Instantiates a defaultdict(None) object.

Return type:

dict[Any, None]

gemseo.core.derivatives.jacobian_assembly.none_factory()[source]

Returns None…

To be used for defaultdict

Return type:

None