"""
OpenMDAO Wrapper for pyoptsparse.
pyoptsparse is based on pyOpt, which is an object-oriented framework for
formulating and solving nonlinear constrained optimization problems, with
additional MPI capability.
"""
from collections import OrderedDict
import json
import signal
from distutils.version import LooseVersion
import numpy as np
from scipy.sparse import coo_matrix
try:
import pyoptsparse
Optimization = pyoptsparse.Optimization
except ImportError:
Optimization = None
pyoptsparse = None
from openmdao.core.constants import INT_DTYPE
from openmdao.core.analysis_error import AnalysisError
from openmdao.core.driver import Driver, RecordingDebugging
import openmdao.utils.coloring as c_mod
from openmdao.utils.class_util import weak_method_wrapper
from openmdao.utils.mpi import FakeComm
from openmdao.utils.om_warnings import issue_warning, DerivativesWarning
# names of optimizers that use gradients
grad_drivers = {'CONMIN', 'FSQP', 'IPOPT', 'NLPQLP',
'PSQP', 'SLSQP', 'SNOPT', 'NLPY_AUGLAG', 'ParOpt'}
# names of optimizers that allow multiple objectives
multi_obj_drivers = {'NSGA2'}
# All optimizers in pyoptsparse
optlist = ['ALPSO', 'CONMIN', 'FSQP', 'IPOPT', 'NLPQLP',
'NSGA2', 'PSQP', 'SLSQP', 'SNOPT', 'NLPY_AUGLAG', 'NOMAD', 'ParOpt']
# All optimizers that require an initial run
run_required = ['NSGA2', 'ALPSO']
DEFAULT_OPT_SETTINGS = {}
DEFAULT_OPT_SETTINGS['IPOPT'] = {
'hessian_approximation': 'limited-memory',
'nlp_scaling_method': 'user-scaling',
'linear_solver': 'mumps'
}
CITATIONS = """@article{Wu_pyoptsparse_2020,
author = {Neil Wu and Gaetan Kenway and Charles A. Mader and John Jasa and
Joaquim R. R. A. Martins},
title = {{pyOptSparse:} A {Python} framework for large-scale constrained
nonlinear optimization of sparse systems},
journal = {Journal of Open Source Software},
volume = {5},
number = {54},
month = {October},
year = {2020},
pages = {2564},
doi = {10.21105/joss.02564},
publisher = {The Open Journal},
}
@article{Hwang_maud_2018
author = {Hwang, John T. and Martins, Joaquim R.R.A.},
title = "{A Computational Architecture for Coupling Heterogeneous
Numerical Models and Computing Coupled Derivatives}",
journal = "{ACM Trans. Math. Softw.}",
volume = {44},
number = {4},
month = jun,
year = {2018},
pages = {37:1--37:39},
articleno = {37},
numpages = {39},
doi = {10.1145/3182393},
publisher = {ACM},
}
"""
DEFAULT_SIGNAL = None
[docs]class UserRequestedException(Exception):
"""
User Requested Exception.
This exception indicates that the user has requested that SNOPT/pyoptsparse ceases
model execution and reports to SNOPT that execution should be terminated.
"""
pass
[docs]class pyOptSparseDriver(Driver):
"""
Driver wrapper for pyoptsparse.
Pyoptsparse is based on pyOpt, which
is an object-oriented framework for formulating and solving nonlinear
constrained optimization problems, with additional MPI capability.
pypptsparse has interfaces to the following optimizers:
ALPSO, CONMIN, FSQP, IPOPT, NLPQLP, NSGA2, PSQP, SLSQP,
SNOPT, NLPY_AUGLAG, NOMAD, ParOpt.
Note that some of these are not open source and therefore not included
in the pyoptsparse source code.
pyOptSparseDriver supports the following:
equality_constraints
inequality_constraints
two_sided_constraints
Attributes
----------
fail : bool
Flag that indicates failure of most recent optimization.
hist_file : str or None
File location for saving pyopt_sparse optimization history.
Default is None for no output.
hotstart_file : str
Optional file to hot start the optimization.
opt_settings : dict
Dictionary for setting optimizer-specific options.
pyopt_solution : Solution
Pyopt_sparse solution object.
_check_jac : bool
Used internally to control when to perform singular checks on computed total derivs.
_exc_info : None or <Exception>
Cached exception that was raised in the _objfunc or _gradfunc callbacks.
_in_user_function :bool
This is set to True at the start of a pyoptsparse callback to _objfunc and _gradfunc, and
restored to False at the finish of each callback.
_indep_list : list
List of design variables.
_quantities : list
Contains the objectives plus nonlinear constraints.
_signal_cache : <Function>
Cached function pointer that was assigned as handler for signal defined in option
user_terminate_signal.
_user_termination_flag : bool
This is set to True when the user sends a signal to terminate the job.
"""
[docs] def __init__(self, **kwargs):
"""
Initialize pyopt.
Parameters
----------
**kwargs : dict of keyword arguments
Keyword arguments that will be mapped into the Driver options.
"""
if Optimization is None:
raise RuntimeError('pyOptSparseDriver is not available, pyOptsparse is not installed.')
super().__init__(**kwargs)
# What we support
self.supports['inequality_constraints'] = True
self.supports['equality_constraints'] = True
self.supports['multiple_objectives'] = True
self.supports['two_sided_constraints'] = True
self.supports['linear_constraints'] = True
self.supports['simultaneous_derivatives'] = True
self.supports['total_jac_sparsity'] = True
# What we don't support yet
self.supports['active_set'] = False
self.supports['integer_design_vars'] = False
self.supports['distributed_design_vars'] = False
self.supports._read_only = True
# The user places optimizer-specific settings in here.
self.opt_settings = {}
# The user can set a file name here to store history
self.hist_file = None
# The user can set a file here to hot start the optimization
# with a history file
self.hotstart_file = None
# We save the pyopt_solution so that it can be queried later on.
self.pyopt_solution = None
self._indep_list = []
self._quantities = []
self.fail = False
self._signal_cache = None
self._user_termination_flag = False
self._in_user_function = False
self._check_jac = False
self._exc_info = None
self.cite = CITATIONS
def _declare_options(self):
"""
Declare options before kwargs are processed in the init method.
"""
self.options.declare('optimizer', default='SLSQP', values=optlist,
desc='Name of optimizers to use')
self.options.declare('title', default='Optimization using pyOpt_sparse',
desc='Title of this optimization run')
self.options.declare('print_results', types=bool, default=True,
desc='Print pyOpt results if True')
self.options.declare('gradient method', default='openmdao',
values={'openmdao', 'pyopt_fd', 'snopt_fd'},
desc='Finite difference implementation to use')
self.options.declare('user_terminate_signal', default=DEFAULT_SIGNAL, allow_none=True,
desc='OS signal that triggers a clean user-termination. Only SNOPT'
'supports this option.')
self.options.declare('singular_jac_behavior', default='warn',
values=['error', 'warn', 'ignore'],
desc='Defines behavior of a zero row/col check after first call to'
'compute_totals:'
'error - raise an error.'
'warn - raise a warning.'
"ignore - don't perform check.")
self.options.declare('singular_jac_tol', default=1e-16,
desc='Tolerance for zero row/column check.')
# Deprecated option
self.options.declare('user_teriminate_signal', default=None, allow_none=True,
desc='OS signal that triggers a clean user-termination. Only SNOPT'
'supports this option.',
deprecation="The option 'user_teriminate_signal' was misspelled and "
"will be deprecated. Please use 'user_terminate_signal' instead.")
def _setup_driver(self, problem):
"""
Prepare the driver for execution.
This is the final thing to run during setup.
Parameters
----------
problem : <Problem>
Pointer to the containing problem.
"""
super()._setup_driver(problem)
self.supports._read_only = False
self.supports['gradients'] = self.options['optimizer'] in grad_drivers
self.supports._read_only = True
if len(self._objs) > 1 and self.options['optimizer'] not in multi_obj_drivers:
raise RuntimeError('Multiple objectives have been added to pyOptSparseDriver'
' but the selected optimizer ({0}) does not support'
' multiple objectives.'.format(self.options['optimizer']))
self._setup_tot_jac_sparsity()
# Handle deprecated option.
if self.options._dict['user_teriminate_signal']['val'] is not None:
self.options['user_terminate_signal'] = \
self.options._dict['user_teriminate_signal']['val']
[docs] def run(self):
"""
Excute pyOptsparse.
Note that pyOpt controls the execution, and the individual optimizers
(e.g., SNOPT) control the iteration.
Returns
-------
boolean
Failure flag; True if failed to converge, False is successful.
"""
problem = self._problem()
model = problem.model
relevant = model._relevant
self.pyopt_solution = None
self._total_jac = None
self.iter_count = 0
fwd = problem._mode == 'fwd'
optimizer = self.options['optimizer']
self._quantities = []
self._check_for_missing_objective()
self._check_jac = self.options['singular_jac_behavior'] in ['error', 'warn']
# Only need initial run if we have linear constraints or if we are using an optimizer that
# doesn't perform one initially.
con_meta = self._cons
model_ran = False
if optimizer in run_required or np.any([con['linear'] for con in self._cons.values()]):
with RecordingDebugging(self._get_name(), self.iter_count, self) as rec:
# Initial Run
model.run_solve_nonlinear()
rec.abs = 0.0
rec.rel = 0.0
model_ran = True
self.iter_count += 1
# compute dynamic simul deriv coloring or just sparsity if option is set
if c_mod._use_total_sparsity:
coloring = None
if self._coloring_info['coloring'] is None and self._coloring_info['dynamic']:
coloring = c_mod.dynamic_total_coloring(self, run_model=not model_ran,
fname=self._get_total_coloring_fname())
if coloring is not None:
# if the improvement wasn't large enough, don't use coloring
pct = coloring._solves_info()[-1]
info = self._coloring_info
if info['min_improve_pct'] > pct:
info['coloring'] = info['static'] = None
msg = f"Coloring was deactivated. Improvement of {pct:.1f}% was less " \
f"than min allowed ({info['min_improve_pct']:.1f}%)."
issue_warning(msg, prefix=self.msginfo, category=DerivativesWarning)
comm = None if isinstance(problem.comm, FakeComm) else problem.comm
opt_prob = Optimization(self.options['title'], weak_method_wrapper(self, '_objfunc'),
comm=comm)
# Add all design variables
input_meta = self._designvars
self._indep_list = indep_list = list(input_meta)
input_vals = self.get_design_var_values()
for name, meta in input_meta.items():
size = meta['global_size'] if meta['distributed'] else meta['size']
opt_prob.addVarGroup(name, size, type='c',
value=input_vals[name],
lower=meta['lower'], upper=meta['upper'])
if not hasattr(pyoptsparse, '__version__') or \
LooseVersion(pyoptsparse.__version__) < LooseVersion('2.5.1'):
opt_prob.finalizeDesignVariables()
else:
opt_prob.finalize()
# Add all objectives
objs = self.get_objective_values()
for name in objs:
opt_prob.addObj(name)
self._quantities.append(name)
# Calculate and save derivatives for any linear constraints.
lcons = [key for (key, con) in con_meta.items() if con['linear']]
if len(lcons) > 0:
_lin_jacs = self._compute_totals(of=lcons, wrt=indep_list, return_format='dict')
# convert all of our linear constraint jacs to COO format. Otherwise pyoptsparse will
# do it for us and we'll end up with a fully dense COO matrix and very slow evaluation
# of linear constraints!
to_remove = []
for jacdct in _lin_jacs.values():
for n, subjac in jacdct.items():
if isinstance(subjac, np.ndarray):
# we can safely use coo_matrix to automatically convert the ndarray
# since our linear constraint jacs are constant, so zeros won't become
# nonzero during the optimization.
mat = coo_matrix(subjac)
if mat.row.size > 0:
# convert to 'coo' format here to avoid an emphatic warning
# by pyoptsparse.
jacdct[n] = {'coo': [mat.row, mat.col, mat.data], 'shape': mat.shape}
# Add all equality constraints
for name, meta in con_meta.items():
if meta['equals'] is None:
continue
size = meta['global_size'] if meta['distributed'] else meta['size']
lower = upper = meta['equals']
if fwd:
wrt = [v for v in indep_list if name in relevant[input_meta[v]['ivc_source']]]
else:
rels = relevant[name]
wrt = [v for v in indep_list if input_meta[v]['ivc_source'] in rels]
if meta['linear']:
jac = {w: _lin_jacs[name][w] for w in wrt}
opt_prob.addConGroup(name, size, lower=lower, upper=upper,
linear=True, wrt=wrt, jac=jac)
else:
if name in self._res_jacs:
resjac = self._res_jacs[name]
jac = {n: resjac[input_meta[n]['ivc_source']] for n in wrt}
else:
jac = None
opt_prob.addConGroup(name, size, lower=lower, upper=upper, wrt=wrt, jac=jac)
self._quantities.append(name)
# Add all inequality constraints
for name, meta in con_meta.items():
if meta['equals'] is not None:
continue
size = meta['global_size'] if meta['distributed'] else meta['size']
# Bounds - double sided is supported
lower = meta['lower']
upper = meta['upper']
if fwd:
wrt = [v for v in indep_list if name in relevant[input_meta[v]['ivc_source']]]
else:
rels = relevant[name]
wrt = [v for v in indep_list if input_meta[v]['ivc_source'] in rels]
if meta['linear']:
jac = {w: _lin_jacs[name][w] for w in wrt}
opt_prob.addConGroup(name, size, upper=upper, lower=lower,
linear=True, wrt=wrt, jac=jac)
else:
if name in self._res_jacs:
resjac = self._res_jacs[name]
jac = {n: resjac[input_meta[n]['ivc_source']] for n in wrt}
else:
jac = None
opt_prob.addConGroup(name, size, upper=upper, lower=lower, wrt=wrt, jac=jac)
self._quantities.append(name)
# Instantiate the requested optimizer
try:
_tmp = __import__('pyoptsparse', globals(), locals(), [optimizer], 0)
opt = getattr(_tmp, optimizer)()
except Exception as err:
# Change whatever pyopt gives us to an ImportError, give it a readable message,
# but raise with the original traceback.
msg = "Optimizer %s is not available in this installation." % optimizer
raise ImportError(msg)
# Process any default optimizer-specific settings.
if optimizer in DEFAULT_OPT_SETTINGS:
for name, value in DEFAULT_OPT_SETTINGS[optimizer].items():
if name not in self.opt_settings:
self.opt_settings[name] = value
# Set optimization options
for option, value in self.opt_settings.items():
opt.setOption(option, value)
self._exc_info = None
try:
# Execute the optimization problem
if self.options['gradient method'] == 'pyopt_fd':
# Use pyOpt's internal finite difference
# TODO: Need to get this from OpenMDAO
# fd_step = problem.model.deriv_options['step_size']
fd_step = 1e-6
sol = opt(opt_prob, sens='FD', sensStep=fd_step, storeHistory=self.hist_file,
hotStart=self.hotstart_file)
elif self.options['gradient method'] == 'snopt_fd':
if self.options['optimizer'] == 'SNOPT':
# Use SNOPT's internal finite difference
# TODO: Need to get this from OpenMDAO
# fd_step = problem.model.deriv_options['step_size']
fd_step = 1e-6
sol = opt(opt_prob, sens=None, sensStep=fd_step, storeHistory=self.hist_file,
hotStart=self.hotstart_file)
else:
msg = "SNOPT's internal finite difference can only be used with SNOPT"
self._exc_info = Exception(msg)
else:
# Use OpenMDAO's differentiator for the gradient
sol = opt(opt_prob, sens=weak_method_wrapper(self, '_gradfunc'),
storeHistory=self.hist_file, hotStart=self.hotstart_file)
except Exception as _:
if not self._exc_info:
raise()
if self._exc_info:
raise self._exc_info
# Print results
if self.options['print_results']:
print(sol)
# Pull optimal parameters back into framework and re-run, so that
# framework is left in the right final state
dv_dict = sol.getDVs()
for name in indep_list:
self.set_design_var(name, dv_dict[name])
with RecordingDebugging(self._get_name(), self.iter_count, self) as rec:
try:
model.run_solve_nonlinear()
except AnalysisError:
model._clear_iprint()
rec.abs = 0.0
rec.rel = 0.0
self.iter_count += 1
# Save the most recent solution.
self.pyopt_solution = sol
try:
exit_status = sol.optInform['value']
self.fail = False
# These are various failed statuses.
if optimizer == 'IPOPT':
if exit_status not in {0, 1}:
self.fail = True
elif exit_status > 2:
self.fail = True
except KeyError:
# optimizers other than pySNOPT may not populate this dict
pass
# revert signal handler to cached version
sigusr = self.options['user_terminate_signal']
if sigusr is not None:
signal.signal(sigusr, self._signal_cache)
self._signal_cache = None # to prevent memory leak test from failing
return self.fail
def _objfunc(self, dv_dict):
"""
Compute the objective function and constraints.
This function is passed to pyOpt's Optimization object and is called
from its optimizers.
Parameters
----------
dv_dict : dict
Dictionary of design variable values.
Returns
-------
func_dict : dict
Dictionary of all functional variables evaluated at design point.
fail : int
0 for successful function evaluation
1 for unsuccessful function evaluation
"""
model = self._problem().model
fail = 0
# Note: we place our handler as late as possible so that codes that run in the
# workflow can place their own handlers.
sigusr = self.options['user_terminate_signal']
if sigusr is not None and self._signal_cache is None:
self._signal_cache = signal.getsignal(sigusr)
signal.signal(sigusr, self._signal_handler)
try:
for name in self._indep_list:
self.set_design_var(name, dv_dict[name])
# print("Setting DV")
# print(dv_dict)
# Check if we caught a termination signal while SNOPT was running.
if self._user_termination_flag:
func_dict = self.get_objective_values()
func_dict.update(self.get_constraint_values(lintype='nonlinear'))
return func_dict, 2
# Execute the model
with RecordingDebugging(self._get_name(), self.iter_count, self) as rec:
self.iter_count += 1
try:
self._in_user_function = True
model.run_solve_nonlinear()
# Let the optimizer try to handle the error
except AnalysisError:
model._clear_iprint()
fail = 1
# User requested termination
except UserRequestedException:
model._clear_iprint()
fail = 2
func_dict = self.get_objective_values()
func_dict.update(self.get_constraint_values(lintype='nonlinear'))
# Record after getting obj and constraint to assure they have
# been gathered in MPI.
rec.abs = 0.0
rec.rel = 0.0
except Exception as raised:
self._exc_info = raised
fail = 1
func_dict = {}
# print("Functions calculated")
# print(dv_dict)
# print(func_dict, flush=True)
self._in_user_function = False
return func_dict, fail
def _gradfunc(self, dv_dict, func_dict):
"""
Compute the gradient of the objective function and constraints.
This function is passed to pyOpt's Optimization object and is called
from its optimizers.
Parameters
----------
dv_dict : dict
Dictionary of design variable values.
func_dict : dict
Dictionary of all functional variables evaluated at design point.
Returns
-------
sens_dict : dict
Dictionary of dictionaries for gradient of each dv/func pair
fail : int
0 for successful function evaluation
1 for unsuccessful function evaluation
"""
prob = self._problem()
fail = 0
try:
# Check if we caught a termination signal while SNOPT was running.
if self._user_termination_flag:
return {}, 2
try:
self._in_user_function = True
sens_dict = self._compute_totals(of=self._quantities,
wrt=self._indep_list,
return_format='dict')
# First time through, check for zero row/col.
if self._check_jac:
raise_error = self.options['singular_jac_behavior'] == 'error'
self._total_jac.check_total_jac(raise_error=raise_error,
tol=self.options['singular_jac_tol'])
self._check_jac = False
# Let the optimizer try to handle the error
except AnalysisError:
prob.model._clear_iprint()
fail = 1
# User requested termination
except UserRequestedException:
prob.model._clear_iprint()
fail = 2
else:
# if we don't convert to 'coo' here, pyoptsparse will do a
# conversion of our dense array into a fully dense 'coo', which is bad.
# TODO: look into getting rid of all of these conversions!
new_sens = OrderedDict()
res_jacs = self._res_jacs
for okey in func_dict:
new_sens[okey] = newdv = OrderedDict()
okey_src = self._responses[okey]['ivc_source']
for ikey in dv_dict:
ikey_src = self._designvars[ikey]['ivc_source']
if okey_src in res_jacs and ikey_src in res_jacs[okey_src]:
arr = sens_dict[okey][ikey]
coo = res_jacs[okey_src][ikey_src]
row, col, data = coo['coo']
coo['coo'][2] = arr[row, col].flatten()
newdv[ikey] = coo
elif okey in sens_dict:
newdv[ikey] = sens_dict[okey][ikey]
sens_dict = new_sens
if fail > 0:
# We need to cobble together a sens_dict of the correct size.
# Best we can do is return zeros.
sens_dict = OrderedDict()
for okey, oval in func_dict.items():
sens_dict[okey] = OrderedDict()
osize = len(oval)
for ikey, ival in dv_dict.items():
isize = len(ival)
sens_dict[okey][ikey] = np.zeros((osize, isize))
except Exception as raised:
self._exc_info = raised
fail = 1
sens_dict = {}
# print("Derivatives calculated")
# print(dv_dict)
# print(sens_dict, flush=True)
self._in_user_function = False
return sens_dict, fail
def _get_name(self):
"""
Get name of current optimizer.
Returns
-------
str
The name of the current optimizer.
"""
return "pyOptSparse_" + self.options['optimizer']
def _get_ordered_nl_responses(self):
"""
Return the names of nonlinear responses in the order used by the driver.
Default order is objectives followed by nonlinear constraints. This is used for
simultaneous derivative coloring and sparsity determination.
Returns
-------
list of str
The nonlinear response names in order.
"""
nl_order = list(self._objs)
neq_order = []
for n, meta in self._cons.items():
if 'linear' not in meta or not meta['linear']:
if meta['equals'] is not None:
nl_order.append(n)
else:
neq_order.append(n)
nl_order.extend(neq_order)
return nl_order
def _setup_tot_jac_sparsity(self, coloring=None):
"""
Set up total jacobian subjac sparsity.
Parameters
----------
coloring : Coloring or None
Current coloring.
"""
total_sparsity = None
self._res_jacs = {}
coloring = coloring if coloring is not None else self._get_static_coloring()
if coloring is not None:
total_sparsity = coloring.get_subjac_sparsity()
if self._total_jac_sparsity is not None:
raise RuntimeError("Total jac sparsity was set in both _total_coloring"
" and _total_jac_sparsity.")
elif self._total_jac_sparsity is not None:
if isinstance(self._total_jac_sparsity, str):
with open(self._total_jac_sparsity, 'r') as f:
self._total_jac_sparsity = json.load(f)
total_sparsity = self._total_jac_sparsity
if total_sparsity is None:
return
for res, resdict in total_sparsity.items():
if res in self._objs: # skip objectives
continue
self._res_jacs[res] = {}
for dv, (rows, cols, shape) in resdict.items():
rows = np.array(rows, dtype=INT_DTYPE)
cols = np.array(cols, dtype=INT_DTYPE)
self._res_jacs[res][dv] = {
'coo': [rows, cols, np.zeros(rows.size)],
'shape': shape,
}
def _signal_handler(self, signum, frame):
# Subsystems (particularly external codes) may declare their own signal handling, so
# execute the cached handler first.
if self._signal_cache is not signal.Handlers.SIG_DFL:
self._signal_cache(signum, frame)
self._user_termination_flag = True
if self._in_user_function:
raise UserRequestedException('User requested termination.')
