# Copyright 2021 IRT Saint Exupéry, https://www.irt-saintexupery.com
#
# This program is free software; you can redistribute it and/or
# modify it under the terms of the GNU Lesser General Public
# License version 3 as published by the Free Software Foundation.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
# Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with this program; if not, write to the Free Software Foundation,
# Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
# Contributors:
# INITIAL AUTHORS - initial API and implementation and/or initial
# documentation
# :author: Francois Gallard
# OTHER AUTHORS - MACROSCOPIC CHANGES
# :author: Pierre-Jean Barjhoux, Benoit Pauwels - MDOScenarioAdapter
# Jacobian computation
"""A discipline running a scenario."""
from __future__ import annotations
import logging
from copy import copy
from copy import deepcopy
from typing import Iterable
from typing import Sequence
from numpy import atleast_1d
from numpy import zeros
from numpy.core.multiarray import ndarray
from numpy.linalg import norm
from gemseo.algos.lagrange_multipliers import LagrangeMultipliers
from gemseo.algos.post_optimal_analysis import PostOptimalAnalysis
from gemseo.core.discipline import MDODiscipline
from gemseo.core.execution_sequence import LoopExecSequence
from gemseo.core.grammars.json_grammar import JSONGrammar
from gemseo.core.parallel_execution import DiscParallelLinearization
from gemseo.core.scenario import Scenario
LOGGER = logging.getLogger(__name__)
[docs]class MDOScenarioAdapter(MDODiscipline):
"""An adapter class for MDO Scenario.
The specified input variables update the default input data of the top level
discipline while the output ones filter the output data from the top level
discipline outputs.
"""
scenario: Scenario
"""The scenario to be adapted."""
post_optimal_analysis: PostOptimalAnalysis
"""The post-optimal analysis."""
LOWER_BND_SUFFIX = "_lower_bnd"
UPPER_BND_SUFFIX = "_upper_bnd"
MULTIPLIER_SUFFIX = "_multiplier"
_ATTR_TO_SERIALIZE = MDODiscipline._ATTR_TO_SERIALIZE + (
"scenario",
"_input_names",
"_reset_x0_before_opt",
"_initial_x",
"_set_x0_before_opt",
"_set_bounds_before_opt",
"_output_names",
"_output_multipliers",
)
def __init__(
self,
scenario: Scenario,
input_names: Sequence[str],
output_names: Sequence[str],
reset_x0_before_opt: bool = False,
set_x0_before_opt: bool = False,
set_bounds_before_opt: bool = False,
cache_type: str = MDODiscipline.SIMPLE_CACHE,
output_multipliers: bool = False,
grammar_type: str = MDODiscipline.JSON_GRAMMAR_TYPE,
name: str | None = None,
) -> None:
""".. # noqa: D205,D212,D415
Args:
scenario: The scenario to adapt.
input_names: The inputs to overload at sub-scenario execution.
output_names: The outputs to get from the sub-scenario execution.
reset_x0_before_opt: If True, reset the initial guess
before running the sub optimization.
set_x0_before_opt: If True, set the initial guess of the sub-scenario.
This is useful for multi-start optimization.
set_bounds_before_opt: If True, set the bounds of the design space.
This is useful for trust regions.
cache_type: The type of cache policy.
output_multipliers: If True,
the Lagrange multipliers of the scenario optimal solution are computed
and added to the outputs.
name: The name of the scenario adapter.
If None, use ``"{}_adapter"``.
Raises:
ValueError: If both `reset_x0_before_opt` and `set_x0_before_opt` are True.
"""
if reset_x0_before_opt and set_x0_before_opt:
raise ValueError("Inconsistent options for MDOScenarioAdapter.")
self.scenario = scenario
self._set_x0_before_opt = set_x0_before_opt
self._set_bounds_before_opt = set_bounds_before_opt
self._input_names = input_names
self._output_names = output_names
self._reset_x0_before_opt = reset_x0_before_opt
self._output_multipliers = output_multipliers
name = name or f"{scenario.name}_adapter"
super().__init__(name, cache_type=cache_type, grammar_type=grammar_type)
self._update_grammars()
self._dv_in_names = None
if set_x0_before_opt:
dv_names = set(self.scenario.formulation.design_space.variables_names)
self._dv_in_names = list(dv_names & set(self._input_names))
# Set the initial bounds as default bounds
self._bounds_names = []
if set_bounds_before_opt:
dspace = scenario.design_space
lower_bounds = dspace.array_to_dict(dspace.get_lower_bounds())
lower_suffix = MDOScenarioAdapter.LOWER_BND_SUFFIX
upper_bounds = dspace.array_to_dict(dspace.get_upper_bounds())
upper_suffix = MDOScenarioAdapter.UPPER_BND_SUFFIX
for bounds, suffix in [
(lower_bounds, lower_suffix),
(upper_bounds, upper_suffix),
]:
bounds = {name + suffix: val for name, val in bounds.items()}
self.default_inputs.update(bounds)
self._bounds_names.extend(bounds.keys())
# Optimization functions are redefined at each run
# since default inputs of top
# level discipline change
# History must be erased otherwise the wrong values are retrieved
# between two runs
scenario.clear_history_before_run = True
self._initial_x = deepcopy(
scenario.design_space.get_current_value(as_dict=True)
)
self.post_optimal_analysis = None
def _update_grammars(self) -> None:
"""Update the input and output grammars.
Raises:
ValueError: Either if a specified input is missing from the input grammar
or if a specified output is missing from the output grammar.
"""
formulation = self.scenario.formulation
opt_problem = formulation.opt_problem
top_leveld = formulation.get_top_level_disc()
for disc in top_leveld:
self.input_grammar.update(disc.input_grammar)
self.output_grammar.update(disc.output_grammar)
# The output may also be the optimum value of the design
# variables, so the output grammar may contain inputs
# of the disciplines. All grammars are filtered just after
# this loop
self.output_grammar.update(disc.input_grammar)
self.default_inputs.update(disc.default_inputs)
try:
self.input_grammar.restrict_to(self._input_names)
except KeyError:
missing_inputs = set(self._input_names) - set(self.input_grammar.keys())
if missing_inputs:
raise ValueError(
"Can't compute inputs from scenarios: {}.".format(
", ".join(sorted(missing_inputs))
)
)
# Add the design variables bounds to the input grammar
if self._set_bounds_before_opt:
current_x = self.scenario.design_space.get_current_value(as_dict=True)
names_to_values = dict()
for suffix in [
MDOScenarioAdapter.LOWER_BND_SUFFIX,
MDOScenarioAdapter.UPPER_BND_SUFFIX,
]:
names_to_values.update({k + suffix: v for k, v in current_x.items()})
bounds_grammar = JSONGrammar("bounds")
bounds_grammar.update_from_data(names_to_values)
self.input_grammar.update(bounds_grammar)
# If a DV is not an input of the top level disciplines:
missing_outputs = set(self._output_names) - set(self.output_grammar.keys())
if missing_outputs:
dv_names = opt_problem.design_space.variables_names
miss_dvs = set(dv_names) & set(missing_outputs)
if miss_dvs:
dv_gram = JSONGrammar("dvs")
dv_gram.update(miss_dvs)
self.output_grammar.update(dv_gram)
try:
self.output_grammar.restrict_to(self._output_names)
except KeyError:
missing_outputs = set(self._output_names) - set(self.output_grammar.keys())
if missing_outputs:
raise ValueError(
"Can't compute outputs from scenarios: {}.".format(
", ".join(sorted(missing_outputs))
)
)
# Add the Lagrange multipliers to the output grammar
if self._output_multipliers:
self._add_output_multipliers()
def _add_output_multipliers(self) -> None:
"""Add the Lagrange multipliers of the scenario optimal solution as outputs."""
# Fill a dictionary with data of typical shapes
base_dict = dict()
problem = self.scenario.formulation.opt_problem
# bound-constraints multipliers
current_x = problem.design_space.get_current_value(as_dict=True)
base_dict.update(
{
self.get_bnd_mult_name(var_name, False): val
for var_name, val in current_x.items()
}
)
base_dict.update(
{
self.get_bnd_mult_name(var_name, True): val
for var_name, val in current_x.items()
}
)
# equality- and inequality-constraints multipliers
base_dict.update(
{
self.get_cstr_mult_name(cstr_name): zeros(1)
for cstr_name in problem.get_constraints_names()
}
)
# Update the output grammar
multipliers_grammar = JSONGrammar("multipliers")
multipliers_grammar.update_from_data(base_dict)
self.output_grammar.update(multipliers_grammar)
[docs] @staticmethod
def get_bnd_mult_name(
variable_name: str,
is_upper: bool,
) -> str:
"""Return the name of the lower bound-constraint multiplier of a variable.
Args:
variable_name: The name of the variable.
is_upper: If True, return name of the upper bound-constraint multiplier.
Otherwise, return the name of the lower bound-constraint multiplier.
Returns:
The name of a bound-constraint multiplier.
"""
mult_name = variable_name
mult_name += "_upp-bnd" if is_upper else "_low-bnd"
mult_name += MDOScenarioAdapter.MULTIPLIER_SUFFIX
return mult_name
[docs] @staticmethod
def get_cstr_mult_name(
constraint_name: str,
) -> str:
"""Return the name of the multiplier of a constraint.
Args:
constraint_name: The name of the constraint.
Returns:
The name of the multiplier.
"""
return constraint_name + MDOScenarioAdapter.MULTIPLIER_SUFFIX
def _run(self) -> None:
self._pre_run()
self.scenario.execute()
self._post_run()
def _pre_run(self) -> None:
"""Pre-run the scenario."""
formulation = self.scenario.formulation
design_space = formulation.opt_problem.design_space
top_leveld = formulation.get_top_level_disc()
# Update the top level discipline default inputs with adapter inputs
# This is the key role of the adapter
for indata in self._input_names:
for disc in top_leveld:
if disc.is_input_existing(indata):
disc.default_inputs[indata] = self.local_data[indata]
if self.scenario.cache is not None:
# Default inputs have changed, therefore caches shall be cleared
self.scenario.cache.clear()
self.scenario.reset_statuses_for_run()
self._reset_optimization_problem()
# Set the starting point of the sub scenario with current dv names
if self._set_x0_before_opt:
dv_values = {dv_n: self.local_data[dv_n] for dv_n in self._dv_in_names}
self.scenario.formulation.design_space.set_current_value(dv_values)
# Set the bounds of the sub-scenario
if self._set_bounds_before_opt:
for name in design_space.variables_names:
# Set the lower bound
lower_suffix = MDOScenarioAdapter.LOWER_BND_SUFFIX
lower_bound = self.local_data[name + lower_suffix]
design_space.set_lower_bound(name, lower_bound)
# Set the upper bound
upper_suffix = MDOScenarioAdapter.UPPER_BND_SUFFIX
upper_bound = self.local_data[name + upper_suffix]
design_space.set_upper_bound(name, upper_bound)
def _reset_optimization_problem(self) -> None:
"""Reset the optimization problem."""
self.scenario.formulation.opt_problem.reset(
design_space=self._reset_x0_before_opt, database=False, preprocessing=False
)
def _post_run(self) -> None:
"""Post-process the scenario."""
formulation = self.scenario.formulation
opt_problem = formulation.opt_problem
design_space = opt_problem.design_space
# Test if the last evaluation is the optimum
x_opt = design_space.get_current_value()
last_x = opt_problem.database.get_x_by_iter(-1)
last_eval_not_opt = norm(x_opt - last_x) / (1.0 + norm(last_x)) > 1e-14
if last_eval_not_opt:
# Revaluate all functions at optimum
# To re execute all disciplines and get the right data
opt_problem.evaluate_functions(
x_opt,
eval_jac=False,
eval_obj=True,
normalize=False,
# Force call without database
no_db_no_norm=True,
)
# Retrieves top-level discipline outputs
self._retrieve_top_level_outputs()
# Compute the Lagrange multipliers and store them in the local data
if self._output_multipliers:
self._compute_lagrange_multipliers()
def _retrieve_top_level_outputs(self) -> None:
"""Retrieve the top-level outputs.
This methods overwrites the adapter outputs with the top-level discipline
outputs and the optimal design parameters.
"""
formulation = self.scenario.formulation
top_level_disciplines = formulation.get_top_level_disc()
current_x = formulation.opt_problem.design_space.get_current_value(as_dict=True)
for name in self._output_names:
for discipline in top_level_disciplines:
if discipline.is_output_existing(name) and name not in current_x:
self.local_data[name] = discipline.local_data[name]
output_value_in_current_x = current_x.get(name)
if output_value_in_current_x is not None:
self.local_data[name] = output_value_in_current_x
def _compute_lagrange_multipliers(self) -> None:
"""Compute the Lagrange multipliers for the optimal solution of the scenario.
This methods stores the multipliers in the local data.
"""
# Compute the Lagrange multipliers
problem = self.scenario.formulation.opt_problem
x_opt = problem.solution.x_opt
lagrange = LagrangeMultipliers(problem)
lagrange.compute(x_opt, problem.ineq_tolerance)
# Store the Lagrange multipliers in the local data
multipliers = lagrange.get_multipliers_arrays()
self.local_data.update(
{
self.get_bnd_mult_name(name, False): mult
for name, mult in multipliers[lagrange.LOWER_BOUNDS].items()
}
)
self.local_data.update(
{
self.get_bnd_mult_name(name, True): mult
for name, mult in multipliers[lagrange.UPPER_BOUNDS].items()
}
)
self.local_data.update(
{
self.get_cstr_mult_name(name): mult
for name, mult in multipliers[lagrange.EQUALITY].items()
}
)
self.local_data.update(
{
self.get_cstr_mult_name(name): mult
for name, mult in multipliers[lagrange.INEQUALITY].items()
}
)
[docs] def get_expected_workflow(self) -> LoopExecSequence: # noqa: D102
return self.scenario.get_expected_workflow()
[docs] def get_expected_dataflow( # noqa: D102
self,
) -> list[tuple[MDODiscipline, MDODiscipline, list[str]]]:
return self.scenario.get_expected_dataflow()
def _compute_jacobian(
self,
inputs: Sequence[str] | None = None,
outputs: Sequence[str] | None = None,
) -> None:
"""Compute the Jacobian of the adapted scenario outputs.
The Jacobian is stored as a dictionary of numpy arrays:
jac = {name: { input_name: ndarray(output_dim, input_dim) } }
The bound-constraints on the scenario optimization variables
are assumed independent of the other scenario inputs.
Args:
inputs: The linearization should be performed with respect to these inputs.
If None, the linearization should be performed w.r.t. all inputs.
outputs: The linearization should be performed on these outputs.
If None, the linearization should be performed on all outputs.
Raises:
ValueError: Either
if the dimension of the objective function is greater than 1,
if a specified input is not an input of the adapter,
if a specified output is not an output of the adapter,
or if there is non differentiable outputs.
"""
opt_problem = self.scenario.formulation.opt_problem
objective_names = self.scenario.formulation.opt_problem.objective.outvars
if len(objective_names) != 1:
raise ValueError("The objective must be single-valued.")
# Check the required inputs
if inputs is None:
inputs = set(self._input_names + self._bounds_names)
else:
not_inputs = set(inputs) - set(self._input_names) - set(self._bounds_names)
if not_inputs:
raise ValueError(
"The following are not inputs of the adapter: {}.".format(
", ".join(sorted(not_inputs))
)
)
# N.B the adapter is assumed constant w.r.t. bounds
bound_inputs = set(inputs) & set(self._bounds_names)
# Check the required outputs
if outputs is None:
outputs = objective_names
else:
not_outputs = sorted(set(outputs) - set(self._output_names))
if not_outputs:
raise ValueError(
"The following are not outputs of the adapter: {}.".format(
", ".join(not_outputs)
)
)
non_differentiable_outputs = sorted(set(outputs) - set(objective_names))
if non_differentiable_outputs:
raise ValueError(
"Post-optimal Jacobians of {} cannot be computed.".format(
", ".join(non_differentiable_outputs)
)
)
# Initialize the Jacobian
diff_inputs = [name for name in inputs if name not in bound_inputs]
# N.B. there may be only bound inputs
self._init_jacobian(diff_inputs, outputs)
# Compute the Jacobians of the optimization functions
jacobians = self._compute_auxiliary_jacobians(diff_inputs, use_threading=True)
# Perform the post-optimal analysis
ineq_tolerance = opt_problem.ineq_tolerance
self.post_optimal_analysis = PostOptimalAnalysis(opt_problem, ineq_tolerance)
post_opt_jac = self.post_optimal_analysis.execute(
outputs, diff_inputs, jacobians
)
self.jac.update(post_opt_jac)
# Fill the Jacobian blocks w.r.t. bounds with zeros
for output_derivatives in self.jac.values():
for bound_input_name in bound_inputs:
bound_input_size = self.default_inputs[bound_input_name].size
output_derivatives[bound_input_name] = zeros((1, bound_input_size))
def _compute_auxiliary_jacobians(
self,
inputs: Iterable[str],
func_names: Iterable[str] | None = None,
use_threading: bool = True,
) -> dict[str, dict[str, ndarray]]:
"""Compute the Jacobians of the optimization functions.
Args:
inputs: The names of the inputs w.r.t. which differentiate.
func_names: The names of the functions to differentiate
If None, then all the optimizations functions are differentiated.
use_threading: Whether to use threads instead of processes
to parallelize the execution;
multiprocessing will copy (serialize) all the disciplines,
while threading will share all the memory.
This is important to note
if you want to execute the same discipline multiple times,
you shall use multiprocessing.
Returns:
The Jacobians of the optimization functions.
"""
# Gather the names of the functions to differentiate
opt_problem = self.scenario.formulation.opt_problem
if func_names is None:
func_names = (
opt_problem.objective.outvars + opt_problem.get_constraints_names()
)
# Identify the disciplines that compute the functions
disciplines = dict()
for func_name in func_names:
for discipline in self.scenario.formulation.get_top_level_disc():
if discipline.is_all_outputs_existing([func_name]):
disciplines[func_name] = discipline
break
# Linearize the required disciplines
unique_disciplines = list(set(disciplines.values()))
for discipline in unique_disciplines:
diff_inputs = set(discipline.get_input_data_names()) & set(inputs)
diff_outputs = set(discipline.get_output_data_names()) & set(func_names)
if diff_inputs and diff_outputs:
discipline.add_differentiated_inputs(list(diff_inputs))
discipline.add_differentiated_outputs(list(diff_outputs))
parallel_linearization = DiscParallelLinearization(
unique_disciplines, use_threading=use_threading
)
# Update the local data with the optimal design parameters
# [The adapted scenario is assumed to have been run beforehand.]
post_opt_data = copy(self.local_data)
post_opt_data.update(opt_problem.design_space.get_current_value(as_dict=True))
parallel_linearization.execute([post_opt_data] * len(unique_disciplines))
# Store the Jacobians
jacobians = dict()
for func_name in func_names:
jacobians[func_name] = dict()
func_jacobian = disciplines[func_name].jac[func_name]
for input_name in inputs:
jacobians[func_name][input_name] = func_jacobian[input_name]
return jacobians
[docs] def add_outputs(
self,
outputs_names: Iterable[str],
) -> None:
"""Add outputs to the scenario adapter.
Args:
outputs_names: The names of the outputs to be added.
"""
names_to_add = [
name for name in outputs_names if name not in self._output_names
]
self._output_names.extend(names_to_add)
self._update_grammars()
[docs]class MDOObjScenarioAdapter(MDOScenarioAdapter):
"""A scenario adapter overwriting the local data with the optimal objective."""
def _retrieve_top_level_outputs(self) -> None:
formulation = self.scenario.formulation
opt_problem = formulation.opt_problem
top_level_disciplines = formulation.get_top_level_disc()
# Get the optimal outputs
optimum = opt_problem.design_space.get_current_value(as_dict=True)
f_opt = opt_problem.get_optimum()[0]
if not opt_problem.minimize_objective:
f_opt = -f_opt
if not opt_problem.is_mono_objective:
raise ValueError("The objective function must be single-valued.")
# Overwrite the adapter local data
objective = opt_problem.objective.outvars[0]
if objective in self._output_names:
self.local_data[objective] = atleast_1d(f_opt)
for output in self._output_names:
if output != objective:
for discipline in top_level_disciplines:
if discipline.is_output_existing(output) and output not in optimum:
self.local_data[output] = discipline.local_data[output]
value = optimum.get(output)
if value is not None:
self.local_data[output] = value
def _compute_jacobian(
self,
inputs: Sequence[str] | None = None,
outputs: Sequence[str] | None = None,
) -> None:
MDOScenarioAdapter._compute_jacobian(self, inputs, outputs)
# The gradient of the objective function cannot be computed by the
# disciplines, but the gradients of the constraints can.
# The objective function is assumed independent of non-optimization
# variables.
obj_name = self.scenario.formulation.opt_problem.objective.outvars[0]
mult_cstr_jac_key = PostOptimalAnalysis.MULT_DOT_CONSTR_JAC
self.jac[obj_name] = dict(self.jac[mult_cstr_jac_key])
