"""Define a base class for all Drivers in OpenMDAO."""
from itertools import chain
import pprint
import sys
import time
import os
import weakref
import numpy as np
from openmdao.core.group import Group
from openmdao.core.total_jac import _TotalJacInfo
from openmdao.core.constants import INT_DTYPE, _SetupStatus
from openmdao.recorders.recording_manager import RecordingManager
from openmdao.recorders.recording_iteration_stack import Recording
from openmdao.utils.hooks import _setup_hooks
from openmdao.utils.record_util import create_local_meta, check_path, has_match
from openmdao.utils.general_utils import _src_name_iter
from openmdao.utils.mpi import MPI
from openmdao.utils.options_dictionary import OptionsDictionary
import openmdao.utils.coloring as coloring_mod
from openmdao.utils.array_utils import sizes2offsets
from openmdao.vectors.vector import _full_slice, _flat_full_indexer
from openmdao.utils.indexer import indexer
from openmdao.utils.om_warnings import issue_warning, DerivativesWarning, DriverWarning
[docs]class Driver(object):
"""
Top-level container for the systems and drivers.
Parameters
----------
**kwargs : dict of keyword arguments
Keyword arguments that will be mapped into the Driver options.
Attributes
----------
fail : bool
Reports whether the driver ran successfully.
iter_count : int
Keep track of iterations for case recording.
options : <OptionsDictionary>
Dictionary with general pyoptsparse options.
recording_options : <OptionsDictionary>
Dictionary with driver recording options.
cite : str
Listing of relevant citations that should be referenced when
publishing work that uses this class.
_problem : weakref to <Problem>
Pointer to the containing problem.
supports : <OptionsDictionary>
Provides a consistent way for drivers to declare what features they support.
_designvars : dict
Contains all design variable info.
_designvars_discrete : list
List of design variables that are discrete.
_dist_driver_vars : dict
Dict of constraints that are distributed outputs. Key is a 'user' variable name,
typically promoted name or an alias. Values are (local indices, local sizes).
_cons : dict
Contains all constraint info.
_objs : dict
Contains all objective info.
_responses : dict
Contains all response info.
_remote_dvs : dict
Dict of design variables that are remote on at least one proc. Values are
(owning rank, size).
_remote_cons : dict
Dict of constraints that are remote on at least one proc. Values are
(owning rank, size).
_remote_objs : dict
Dict of objectives that are remote on at least one proc. Values are
(owning rank, size).
_rec_mgr : <RecordingManager>
Object that manages all recorders added to this driver.
_coloring_info : dict
Metadata pertaining to total coloring.
_total_jac_format : str
Specifies the format of the total jacobian. Allowed values are 'flat_dict', 'dict', and
'array'.
_con_subjacs : dict
Dict of sparse subjacobians for use with certain optimizers, e.g. pyOptSparseDriver.
Keyed by sources and aliases.
_total_jac : _TotalJacInfo or None
Cached total jacobian handling object.
_total_jac_linear : _TotalJacInfo or None
Cached linear total jacobian handling object.
opt_result : dict
Dictionary containing information for use in the optimization report.
_has_scaling : bool
If True, scaling has been set for this driver.
"""
[docs] def __init__(self, **kwargs):
"""
Initialize the driver.
"""
self._rec_mgr = RecordingManager()
self._problem = None
self._designvars = None
self._designvars_discrete = []
self._cons = None
self._objs = None
self._responses = None
# Driver options
self.options = OptionsDictionary(parent_name=type(self).__name__)
self.options.declare('debug_print', types=list,
values=['desvars', 'nl_cons', 'ln_cons', 'objs', 'totals'],
desc="List of what type of Driver variables to print at each "
"iteration.",
default=[])
default_desvar_behavior = os.environ.get('OPENMDAO_INVALID_DESVAR_BEHAVIOR', 'warn').lower()
self.options.declare('invalid_desvar_behavior', values=('warn', 'raise', 'ignore'),
desc='Behavior of driver if the initial value of a design '
'variable exceeds its bounds. The default value may be'
'set using the `OPENMDAO_INVALID_DESVAR_BEHAVIOR` environment '
'variable to one of the valid options.',
default=default_desvar_behavior)
# Case recording options
self.recording_options = OptionsDictionary(parent_name=type(self).__name__)
self.recording_options.declare('record_desvars', types=bool, default=True,
desc='Set to True to record design variables at the '
'driver level')
self.recording_options.declare('record_responses', types=bool, default=False,
desc='Set True to record constraints and objectives at the '
'driver level')
self.recording_options.declare('record_objectives', types=bool, default=True,
desc='Set to True to record objectives at the driver level')
self.recording_options.declare('record_constraints', types=bool, default=True,
desc='Set to True to record constraints at the '
'driver level')
self.recording_options.declare('includes', types=list, default=[],
desc='Patterns for variables to include in recording. '
'Uses fnmatch wildcards')
self.recording_options.declare('excludes', types=list, default=[],
desc='Patterns for vars to exclude in recording '
'(processed post-includes). Uses fnmatch wildcards')
self.recording_options.declare('record_derivatives', types=bool, default=False,
desc='Set to True to record derivatives at the driver '
'level')
self.recording_options.declare('record_inputs', types=bool, default=True,
desc='Set to True to record inputs at the driver level')
self.recording_options.declare('record_outputs', types=bool, default=True,
desc='Set True to record outputs at the '
'driver level.')
self.recording_options.declare('record_residuals', types=bool, default=False,
desc='Set True to record residuals at the '
'driver level.')
# What the driver supports.
self.supports = OptionsDictionary(parent_name=type(self).__name__)
self.supports.declare('optimization', types=bool, default=False)
self.supports.declare('inequality_constraints', types=bool, default=False)
self.supports.declare('equality_constraints', types=bool, default=False)
self.supports.declare('linear_constraints', types=bool, default=False)
self.supports.declare('two_sided_constraints', types=bool, default=False)
self.supports.declare('multiple_objectives', types=bool, default=False)
self.supports.declare('integer_design_vars', types=bool, default=True)
self.supports.declare('gradients', types=bool, default=False)
self.supports.declare('active_set', types=bool, default=False)
self.supports.declare('simultaneous_derivatives', types=bool, default=False)
self.supports.declare('total_jac_sparsity', types=bool, default=False)
self.supports.declare('distributed_design_vars', types=bool, default=True)
self.iter_count = 0
self.cite = ""
self._coloring_info = coloring_mod.ColoringMeta()
self._total_jac_format = 'flat_dict'
self._con_subjacs = {}
self._total_jac = None
self._total_jac_linear = None
self.fail = False
self._declare_options()
self.options.update(kwargs)
self.opt_result = {
'runtime': 0.0,
'iter_count': 0,
'obj_calls': 0,
'deriv_calls': 0,
'exit_status': 'NOT_RUN'
}
self._has_scaling = False
def _get_inst_id(self):
if self._problem is None:
return None
return f"{self._problem()._get_inst_id()}.driver"
@property
def msginfo(self):
"""
Return info to prepend to messages.
Returns
-------
str
Info to prepend to messages.
"""
return type(self).__name__
[docs] def add_recorder(self, recorder):
"""
Add a recorder to the driver.
Parameters
----------
recorder : CaseRecorder
A recorder instance.
"""
self._rec_mgr.append(recorder)
[docs] def cleanup(self):
"""
Clean up resources prior to exit.
"""
# shut down all recorders
self._rec_mgr.shutdown()
def _declare_options(self):
"""
Declare options before kwargs are processed in the init method.
This is optionally implemented by subclasses of Driver.
"""
pass
def _setup_comm(self, comm):
"""
Perform any driver-specific setup of communicators for the model.
Parameters
----------
comm : MPI.Comm or <FakeComm> or None
The communicator for the Problem.
Returns
-------
MPI.Comm or <FakeComm> or None
The communicator for the Problem model.
"""
return comm
def _set_problem(self, problem):
"""
Set a reference to the containing Problem.
Parameters
----------
problem : <Problem>
Reference to the containing problem.
"""
self._problem = weakref.ref(problem)
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.
"""
model = problem.model
self._total_jac = None
# Determine if any design variables are discrete.
self._designvars_discrete = [name for name, meta in self._designvars.items()
if meta['source'] in model._discrete_outputs]
if not self.supports['integer_design_vars'] and len(self._designvars_discrete) > 0:
msg = "Discrete design variables are not supported by this driver: "
msg += '.'.join(self._designvars_discrete)
raise RuntimeError(msg)
self._remote_dvs = remote_dv_dict = {}
self._remote_cons = remote_con_dict = {}
self._dist_driver_vars = dist_dict = {}
self._remote_objs = remote_obj_dict = {}
# Only allow distributed design variables on drivers that support it.
if self.supports['distributed_design_vars'] is False:
dist_vars = []
abs2meta_in = model._var_allprocs_abs2meta['input']
discrete_in = model._var_allprocs_discrete['input']
for dv, meta in self._designvars.items():
# For Auto-ivcs, we need to check the distributed metadata on the target instead.
if meta['source'].startswith('_auto_ivc.'):
for abs_name in model._var_allprocs_prom2abs_list['input'][dv]:
# we can use abs name to check for discrete vars here because
# relative names are absolute names at the model level.
if abs_name in discrete_in:
# Discrete vars aren't distributed.
break
if abs2meta_in[abs_name]['distributed']:
dist_vars.append(dv)
break
elif meta['distributed']:
dist_vars.append(dv)
if dist_vars:
dstr = ', '.join(dist_vars)
msg = "Distributed design variables are not supported by this driver, but the "
msg += f"following variables are distributed: [{dstr}]"
raise RuntimeError(msg)
# Now determine if later we'll need to allgather cons, objs, or desvars.
if model.comm.size > 1:
loc_vars = set(model._outputs._abs_iter())
# some of these lists could have duplicate src names if aliases are used. We'll
# fix that when we convert to sets after the allgather.
remote_dvs = [n for n in _src_name_iter(self._designvars) if n not in loc_vars]
remote_cons = [n for n in _src_name_iter(self._cons) if n not in loc_vars]
remote_objs = [n for n in _src_name_iter(self._objs) if n not in loc_vars]
con_set = set()
obj_set = set()
dv_set = set()
all_remote_vois = model.comm.allgather((remote_dvs, remote_cons, remote_objs))
for rem_dvs, rem_cons, rem_objs in all_remote_vois:
con_set.update(rem_cons)
obj_set.update(rem_objs)
dv_set.update(rem_dvs)
# If we have remote VOIs, pick an owning rank for each and use that
# to bcast to others later
owning_ranks = model._owning_rank
abs2idx = model._var_allprocs_abs2idx
abs2meta_out = model._var_allprocs_abs2meta['output']
sizes = model._var_sizes['output']
rank = model.comm.rank
nprocs = model.comm.size
# Loop over all VOIs.
for vname, voimeta in chain(self._responses.items(), self._designvars.items()):
# vname may be a promoted name or an alias
indices = voimeta['indices']
vsrc = voimeta['source']
meta = abs2meta_out[vsrc]
i = abs2idx[vsrc]
if meta['distributed']:
dist_sizes = sizes[:, i]
tot_size = np.sum(dist_sizes)
# Determine which indices are on our proc.
offsets = sizes2offsets(dist_sizes)
if indices is not None:
indices = indices.shaped_array()
true_sizes = np.zeros(nprocs, dtype=INT_DTYPE)
for irank in range(nprocs):
dist_inds = indices[np.logical_and(indices >= offsets[irank],
indices < (offsets[irank] +
dist_sizes[irank]))]
true_sizes[irank] = dist_inds.size
if irank == rank:
local_indices = dist_inds - offsets[rank]
distrib_indices = dist_inds
ind = indexer(local_indices, src_shape=(tot_size,), flat_src=True)
dist_dict[vname] = (ind, true_sizes, distrib_indices)
else:
dist_dict[vname] = (_flat_full_indexer, dist_sizes,
slice(offsets[rank], offsets[rank] + dist_sizes[rank]))
else:
owner = owning_ranks[vsrc]
sz = sizes[owner, i]
if vsrc in dv_set:
remote_dv_dict[vname] = (owner, sz)
if vsrc in con_set:
remote_con_dict[vname] = (owner, sz)
if vsrc in obj_set:
remote_obj_dict[vname] = (owner, sz)
self._remote_responses = self._remote_cons.copy()
self._remote_responses.update(self._remote_objs)
# set up simultaneous deriv coloring
if coloring_mod._use_total_sparsity:
# reset the coloring
if self._coloring_info.dynamic or self._coloring_info.static is not None:
self._coloring_info.coloring = None
coloring = self._get_static_coloring()
if coloring is not None and self.supports['simultaneous_derivatives']:
if model._owns_approx_jac:
coloring._check_config_partial(model)
else:
coloring._check_config_total(self, model)
if not problem.model._use_derivatives:
issue_warning("Derivatives are turned off. Skipping simul deriv coloring.",
category=DerivativesWarning)
def _check_for_missing_objective(self):
"""
Check for missing objective and raise error if no objectives found.
"""
if len(self._objs) == 0:
msg = "Driver requires objective to be declared"
raise RuntimeError(msg)
def _check_for_invalid_desvar_values(self):
"""
Check for design variable values that exceed their bounds.
This method's behavior is controlled by the OPENMDAO_INVALID_DESVAR environment variable,
which may take on values 'ignore', 'raise'', 'warn'.
- 'ignore' : Proceed without checking desvar bounds.
- 'warn' : Issue a warning if one or more desvar values exceed bounds.
- 'raise' : Raise an exception if one or more desvar values exceed bounds.
These options are case insensitive.
"""
if self.options['invalid_desvar_behavior'] != 'ignore':
invalid_desvar_data = []
for var, meta in self._designvars.items():
_val = self._problem().get_val(var, units=meta['units'], get_remote=True)
val = np.array([_val]) if np.ndim(_val) == 0 else _val # Handle discrete desvars
idxs = meta['indices']() if meta['indices'] else None
flat_idxs = meta['flat_indices']
scaler = meta['scaler'] if meta['scaler'] is not None else 1.
adder = meta['adder'] if meta['adder'] is not None else 0.
lower = meta['lower'] / scaler - adder
upper = meta['upper'] / scaler - adder
flat_val = val.ravel()[idxs] if flat_idxs else val[idxs].ravel()
if (flat_val < lower).any() or (flat_val > upper).any():
invalid_desvar_data.append((var, val, lower, upper))
if invalid_desvar_data:
s = 'The following design variable initial conditions are out of their ' \
'specified bounds:'
for var, val, lower, upper in invalid_desvar_data:
s += f'\n {var}\n val: {val.ravel()}' \
f'\n lower: {lower}\n upper: {upper}'
s += '\nSet the initial value of the design variable to a valid value or set ' \
'the driver option[\'invalid_desvar_behavior\'] to \'ignore\'.'
if self.options['invalid_desvar_behavior'] == 'raise':
raise ValueError(s)
else:
issue_warning(s, category=DriverWarning)
def _get_vars_to_record(self, obj=None):
"""
Get variables to record based on recording options.
Parameters
----------
obj : Problem or Driver
Parent object which has recording options.
Returns
-------
dict
Dictionary containing lists of variables to record.
"""
if obj is None:
obj = self
recording_options = obj.recording_options
problem = self._problem()
model = problem.model
incl = recording_options['includes']
excl = recording_options['excludes']
# includes and excludes for outputs are specified using promoted names
# includes and excludes for inputs are specified using _absolute_ names
abs2prom_output = model._var_allprocs_abs2prom['output']
abs2prom_inputs = model._var_allprocs_abs2prom['input']
# set of promoted output names and absolute input and residual names
# used for matching includes/excludes
match_names = set()
# 1. If record_outputs is True, get the set of outputs
# 2. Filter those using includes and excludes to get the baseline set of variables to record
# 3. Add or remove from that set any desvars, objs, and cons based on the recording
# options of those
# includes and excludes for outputs are specified using _promoted_ names
# vectors are keyed on absolute name, discretes on relative/promoted name
myinputs = myoutputs = myresiduals = []
if recording_options['record_outputs']:
match_names = match_names | set(abs2prom_output.values())
myoutputs = [n for n, prom in abs2prom_output.items() if check_path(prom, incl, excl)]
model_outs = model._outputs
if model._var_discrete['output']:
# if we have discrete outputs then residual name set doesn't match output one
if recording_options['record_residuals']:
myresiduals = [n for n in myoutputs if model_outs._contains_abs(n)]
elif recording_options['record_residuals']:
myresiduals = myoutputs
elif recording_options['record_residuals']:
match_names = match_names | set(model._residuals.keys())
myresiduals = [n for n in model._residuals._abs_iter()
if check_path(abs2prom_output[n], incl, excl)]
myoutputs = set(myoutputs)
if recording_options['record_desvars']:
myoutputs.update(_src_name_iter(self._designvars))
if recording_options['record_objectives'] or recording_options['record_responses']:
myoutputs.update(_src_name_iter(self._objs))
if recording_options['record_constraints'] or recording_options['record_responses']:
myoutputs.update(_src_name_iter(self._cons))
# inputs (if in options). inputs use _absolute_ names for includes/excludes
if 'record_inputs' in recording_options:
if recording_options['record_inputs']:
match_names = match_names | set(abs2prom_inputs.keys())
myinputs = [n for n in abs2prom_inputs if check_path(n, incl, excl)]
# check that all exclude/include globs have at least one matching output or input name
for pattern in excl:
if not has_match(pattern, match_names):
issue_warning(f"{obj.msginfo}: No matches for pattern '{pattern}' in "
"recording_options['excludes'].")
for pattern in incl:
if not has_match(pattern, match_names):
issue_warning(f"{obj.msginfo}: No matches for pattern '{pattern}' in "
"recording_options['includes'].")
# sort lists to ensure that vars are iterated over in the same order on all procs
vars2record = {
'input': sorted(myinputs),
'output': sorted(myoutputs),
'residual': sorted(myresiduals)
}
return vars2record
def _setup_recording(self):
"""
Set up case recording.
"""
self._filtered_vars_to_record = self._get_vars_to_record()
self._rec_mgr.startup(self, self._problem().comm)
def _run(self):
"""
Execute this driver.
This calls the run() method, which should be overriden by the subclass.
Returns
-------
bool
Failure flag; True if failed to converge, False is successful.
"""
problem = self._problem()
model = problem.model
if self.supports['optimization'] and problem.options['group_by_pre_opt_post']:
if model._pre_components:
with model._relevance.nonlinear_active('pre'):
model.run_solve_nonlinear()
with SaveOptResult(self):
with model._relevance.nonlinear_active('iter'):
result = self.run()
if model._post_components:
with model._relevance.nonlinear_active('post'):
model.run_solve_nonlinear()
return result
else:
with SaveOptResult(self):
return self.run()
def _get_voi_val(self, name, meta, remote_vois, driver_scaling=True,
get_remote=True, rank=None):
"""
Get the value of a variable of interest (objective, constraint, or design var).
This will retrieve the value if the VOI is remote.
Parameters
----------
name : str
Name of the variable of interest.
meta : dict
Metadata for the variable of interest.
remote_vois : dict
Dict containing (owning_rank, size) for all remote vois of a particular
type (design var, constraint, or objective).
driver_scaling : bool
When True, return values that are scaled according to either the adder and scaler or
the ref and ref0 values that were specified when add_design_var, add_objective, and
add_constraint were called on the model. Default is True.
get_remote : bool or None
If True, retrieve the value even if it is on a remote process. Note that if the
variable is remote on ANY process, this function must be called on EVERY process
in the Problem's MPI communicator.
If False, only retrieve the value if it is on the current process, or only the part
of the value that's on the current process for a distributed variable.
rank : int or None
If not None, gather value to this rank only.
Returns
-------
float or ndarray
The value of the named variable of interest.
"""
model = self._problem().model
comm = model.comm
get = model._outputs._abs_get_val
indices = meta['indices']
src_name = meta['source']
if MPI:
distributed = comm.size > 0 and name in self._dist_driver_vars
else:
distributed = False
if name in remote_vois:
owner, size = remote_vois[name]
# if var is distributed or only gathering to one rank
# TODO - support distributed var under a parallel group.
if owner is None or rank is not None:
val = model.get_val(src_name, get_remote=get_remote, rank=rank, flat=True)
if indices is not None:
val = val[indices.flat()]
else:
if owner == comm.rank:
if indices is None:
val = get(src_name, flat=True).copy()
else:
val = get(src_name, flat=True)[indices.as_array()]
else:
if indices is not None:
size = indices.indexed_src_size
val = np.empty(size)
if get_remote:
comm.Bcast(val, root=owner)
elif distributed:
local_val = model.get_val(src_name, get_remote=False, flat=True)
local_indices, sizes, _ = self._dist_driver_vars[name]
if local_indices is not _full_slice:
local_val = local_val[local_indices()]
if get_remote:
local_val = np.ascontiguousarray(local_val)
offsets = np.zeros(sizes.size, dtype=INT_DTYPE)
offsets[1:] = np.cumsum(sizes[:-1])
val = np.zeros(np.sum(sizes))
comm.Allgatherv(local_val, [val, sizes, offsets, MPI.DOUBLE])
else:
val = local_val
else:
if src_name in model._discrete_outputs:
val = model._discrete_outputs[src_name]
if name in self._designvars_discrete:
# At present, only integers are supported by OpenMDAO drivers.
# We check the values here.
if not ((np.isscalar(val) and isinstance(val, (int, np.integer))) or
(isinstance(val, np.ndarray) and np.issubdtype(val[0], np.integer))):
if np.isscalar(val):
suffix = f"A value of type '{type(val).__name__}' was specified."
elif isinstance(val, np.ndarray):
suffix = f"An array of type '{val.dtype.name}' was specified."
else:
suffix = ''
raise ValueError("Only integer scalars or ndarrays are supported as values "
"for discrete variables when used as a design variable. "
+ suffix)
elif indices is None:
val = get(src_name, flat=True).copy()
else:
val = get(src_name, flat=True)[indices.as_array()]
if self._has_scaling and driver_scaling:
# Scale design variable values
adder = meta['total_adder']
if adder is not None:
val += adder
scaler = meta['total_scaler']
if scaler is not None:
val *= scaler
return val
[docs] def get_driver_objective_calls(self):
"""
Return number of objective evaluations made during a driver run.
Returns
-------
int
Number of objective evaluations made during a driver run.
"""
return 0
[docs] def get_driver_derivative_calls(self):
"""
Return number of derivative evaluations made during a driver run.
Returns
-------
int
Number of derivative evaluations made during a driver run.
"""
return 0
[docs] def get_design_var_values(self, get_remote=True, driver_scaling=True):
"""
Return the design variable values.
Parameters
----------
get_remote : bool or None
If True, retrieve the value even if it is on a remote process. Note that if the
variable is remote on ANY process, this function must be called on EVERY process
in the Problem's MPI communicator.
If False, only retrieve the value if it is on the current process, or only the part
of the value that's on the current process for a distributed variable.
driver_scaling : bool
When True, return values that are scaled according to either the adder and scaler or
the ref and ref0 values that were specified when add_design_var, add_objective, and
add_constraint were called on the model. Default is True.
Returns
-------
dict
Dictionary containing values of each design variable.
"""
return {n: self._get_voi_val(n, dvmeta, self._remote_dvs, get_remote=get_remote,
driver_scaling=driver_scaling)
for n, dvmeta in self._designvars.items()}
[docs] def set_design_var(self, name, value, set_remote=True):
"""
Set the value of a design variable.
'name' can be a promoted output name or an alias.
Parameters
----------
name : str
Global pathname of the design variable.
value : float or ndarray
Value for the design variable.
set_remote : bool
If True, set the global value of the variable (value must be of the global size).
If False, set the local value of the variable (value must be of the local size).
"""
problem = self._problem()
meta = self._designvars[name]
src_name = meta['source']
# if the value is not local, don't set the value
if (src_name in self._remote_dvs and
problem.model._owning_rank[src_name] != problem.comm.rank):
return
if name in self._designvars_discrete:
# Note, drivers set values here and generally should know it is setting an integer.
# However, the DOEdriver may pull a non-integer value from its generator, so we
# convert it.
if isinstance(value, float):
value = int(value)
elif isinstance(value, np.ndarray):
if isinstance(problem.model._discrete_outputs[src_name], int):
# Setting an integer value with a 1D array - don't want to convert to array.
value = int(value.item())
else:
value = value.astype(int)
problem.model._discrete_outputs[src_name] = value
elif problem.model._outputs._contains_abs(src_name):
desvar = problem.model._outputs._abs_get_val(src_name)
if name in self._dist_driver_vars:
loc_idxs, _, dist_idxs = self._dist_driver_vars[name]
loc_idxs = loc_idxs() # don't use indexer here
else:
loc_idxs = meta['indices']
if loc_idxs is None:
loc_idxs = _full_slice
else:
loc_idxs = loc_idxs()
dist_idxs = _full_slice
if set_remote:
# provided value is the global value, use indices for this proc
desvar[loc_idxs] = np.atleast_1d(value)[dist_idxs]
else:
# provided value is the local value
desvar[loc_idxs] = np.atleast_1d(value)
# Undo driver scaling when setting design var values into model.
if self._has_scaling:
scaler = meta['total_scaler']
if scaler is not None:
desvar[loc_idxs] *= 1.0 / scaler
adder = meta['total_adder']
if adder is not None:
desvar[loc_idxs] -= adder
[docs] def get_objective_values(self, driver_scaling=True):
"""
Return objective values.
Parameters
----------
driver_scaling : bool
When True, return values that are scaled according to either the adder and scaler or
the ref and ref0 values that were specified when add_design_var, add_objective, and
add_constraint were called on the model. Default is True.
Returns
-------
dict
Dictionary containing values of each objective.
"""
return {n: self._get_voi_val(n, obj, self._remote_objs,
driver_scaling=driver_scaling)
for n, obj in self._objs.items()}
[docs] def get_constraint_values(self, ctype='all', lintype='all', driver_scaling=True):
"""
Return constraint values.
Parameters
----------
ctype : str
Default is 'all'. Optionally return just the inequality constraints
with 'ineq' or the equality constraints with 'eq'.
lintype : str
Default is 'all'. Optionally return just the linear constraints
with 'linear' or the nonlinear constraints with 'nonlinear'.
driver_scaling : bool
When True, return values that are scaled according to either the adder and scaler or
the ref and ref0 values that were specified when add_design_var, add_objective, and
add_constraint were called on the model. Default is True.
Returns
-------
dict
Dictionary containing values of each constraint.
"""
con_dict = {}
for name, meta in self._cons.items():
if lintype == 'linear' and not meta['linear']:
continue
if lintype == 'nonlinear' and meta['linear']:
continue
if ctype == 'eq' and meta['equals'] is None:
continue
if ctype == 'ineq' and meta['equals'] is not None:
continue
con_dict[name] = self._get_voi_val(name, meta, self._remote_cons,
driver_scaling=driver_scaling)
return con_dict
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.
"""
order = list(self._objs)
order.extend(n for n, meta in self._cons.items()
if not ('linear' in meta and meta['linear']))
return order
def _update_voi_meta(self, model):
"""
Collect response and design var metadata from the model and size desvars and responses.
Parameters
----------
model : System
The System that represents the entire model.
Returns
-------
int
Total size of responses, with linear constraints excluded.
int
Total size of design vars.
"""
self._objs = objs = {}
self._cons = cons = {}
# driver _responses are keyed by either the alias or the promoted name
response_size = 0
self._responses = resps = model.get_responses(recurse=True, use_prom_ivc=True)
for name, data in resps.items():
if data['type'] == 'con':
cons[name] = data
else:
objs[name] = data
response_size += data['global_size']
# Gather up the information for design vars. _designvars are keyed by the promoted name
self._designvars = designvars = model.get_design_vars(recurse=True, use_prom_ivc=True)
desvar_size = sum(data['global_size'] for data in designvars.values())
self._has_scaling = model._setup_driver_units()
return response_size, desvar_size
[docs] def get_exit_status(self):
"""
Return exit status of driver run.
Returns
-------
str
String indicating result of driver run.
"""
return 'FAIL' if self.fail else 'SUCCESS'
[docs] def check_relevance(self):
"""
Check if there are constraints that don't depend on any design vars.
This usually indicates something is wrong with the problem formulation.
"""
# relevance not relevant if not using derivatives
if not self.supports['gradients']:
return
if 'singular_jac_behavior' in self.options:
singular_behavior = self.options['singular_jac_behavior']
if singular_behavior == 'ignore':
return
else:
singular_behavior = 'warn'
problem = self._problem()
# Do not perform this check if any subgroup uses approximated partials.
# This causes the relevance graph to be invalid.
for system in problem.model.system_iter(include_self=True, recurse=True, typ=Group):
if system._has_approx:
return
bad = {n for n in self._problem().model._relevance._no_dv_responses
if n not in self._designvars}
if bad:
bad_conns = [m['name'] for m in self._cons.values() if m['source'] in bad]
bad_objs = [m['name'] for m in self._objs.values() if m['source'] in bad]
badmsg = []
if bad_conns:
badmsg.append(f"constraint(s) {bad_conns}")
if bad_objs:
badmsg.append(f"objective(s) {bad_objs}")
bad = ' and '.join(badmsg)
# Note: There is a hack in ScipyOptimizeDriver for older versions of COBYLA that
# implements bounds on design variables by adding them as constraints.
# These design variables as constraints will not appear in the wrt list.
msg = f"{self.msginfo}: {bad} do not depend on any " \
"design variables. Please check your problem formulation."
if singular_behavior == 'error':
raise RuntimeError(msg)
else:
issue_warning(msg, category=DriverWarning)
[docs] def run(self):
"""
Execute this driver.
The base `Driver` just runs the model. All other drivers overload
this method.
Returns
-------
bool
Failure flag; True if failed to converge, False is successful.
"""
with RecordingDebugging(self._get_name(), self.iter_count, self):
self._problem().model.run_solve_nonlinear()
self.iter_count += 1
return False
@property
def _recording_iter(self):
return self._problem()._metadata['recording_iter']
def _compute_totals(self, of=None, wrt=None, return_format='flat_dict', driver_scaling=True):
"""
Compute derivatives of desired quantities with respect to desired inputs.
All derivatives are returned using driver scaling.
Parameters
----------
of : list of variable name str or None
Variables whose derivatives will be computed. Default is None, which
uses the driver's objectives and constraints.
wrt : list of variable name str or None
Variables with respect to which the derivatives will be computed.
Default is None, which uses the driver's desvars.
return_format : str
Format to return the derivatives. Default is a 'flat_dict', which
returns them in a dictionary whose keys are tuples of form (of, wrt). For
the scipy optimizer, 'array' is also supported.
driver_scaling : bool
If True (default), scale derivative values by the quantities specified when the desvars
and responses were added. If False, leave them unscaled.
Returns
-------
derivs : object
Derivatives in form requested by 'return_format'.
"""
problem = self._problem()
debug_print = 'totals' in self.options['debug_print'] and (not MPI or
problem.comm.rank == 0)
if debug_print:
header = 'Driver total derivatives for iteration: ' + str(self.iter_count)
print(header)
print(len(header) * '-' + '\n')
if self._total_jac is None:
total_jac = _TotalJacInfo(problem, of, wrt, return_format,
approx=problem.model._owns_approx_jac,
debug_print=debug_print,
driver_scaling=driver_scaling)
if total_jac.has_lin_cons:
# if we're doing a scaling report, cache the linear total jacobian so we
# don't have to recreate it
if problem._has_active_report('scaling'):
self._total_jac_linear = total_jac
else:
self._total_jac = total_jac
else:
total_jac = self._total_jac
totals = total_jac.compute_totals()
if self.recording_options['record_derivatives']:
self.record_derivatives()
return totals
[docs] def record_derivatives(self):
"""
Record the current total jacobian.
"""
if self._total_jac is not None and self._rec_mgr._recorders:
metadata = create_local_meta(self._get_name())
self._total_jac.record_derivatives(self, metadata)
[docs] def record_iteration(self):
"""
Record an iteration of the current Driver.
"""
status = -1 if self._problem is None else self._problem()._metadata['setup_status']
if status >= _SetupStatus.POST_FINAL_SETUP:
record_iteration(self, self._problem(), self._get_name())
else:
raise RuntimeError(f'{self.msginfo} attempted to record iteration but '
'driver has not been initialized; `run_model()`, '
'`run_driver()`, or `final_setup()` must be called '
'before recording.')
def _get_recorder_metadata(self, case_name):
"""
Return metadata from the latest iteration for use in the recorder.
Parameters
----------
case_name : str
Name of current case.
Returns
-------
dict
Metadata dictionary for the recorder.
"""
return create_local_meta(case_name)
def _get_name(self):
"""
Get name of current Driver.
Returns
-------
str
Name of current Driver.
"""
return "Driver"
[docs] def declare_coloring(self, num_full_jacs=coloring_mod._DEF_COMP_SPARSITY_ARGS['num_full_jacs'],
tol=coloring_mod._DEF_COMP_SPARSITY_ARGS['tol'],
orders=coloring_mod._DEF_COMP_SPARSITY_ARGS['orders'],
perturb_size=coloring_mod._DEF_COMP_SPARSITY_ARGS['perturb_size'],
min_improve_pct=coloring_mod._DEF_COMP_SPARSITY_ARGS['min_improve_pct'],
show_summary=coloring_mod._DEF_COMP_SPARSITY_ARGS['show_summary'],
show_sparsity=coloring_mod._DEF_COMP_SPARSITY_ARGS['show_sparsity'],
use_scaling=coloring_mod._DEF_COMP_SPARSITY_ARGS['use_scaling']):
"""
Set options for total deriv coloring.
Parameters
----------
num_full_jacs : int
Number of times to repeat partial jacobian computation when computing sparsity.
tol : float
Tolerance used to determine if an array entry is nonzero during sparsity determination.
orders : int
Number of orders above and below the tolerance to check during the tolerance sweep.
perturb_size : float
Size of input/output perturbation during generation of sparsity.
min_improve_pct : float
If coloring does not improve (decrease) the number of solves more than the given
percentage, coloring will not be used.
show_summary : bool
If True, display summary information after generating coloring.
show_sparsity : bool
If True, display sparsity with coloring info after generating coloring.
use_scaling : bool
If True, use driver scaling when generating the sparsity.
"""
self._coloring_info.coloring = None
self._coloring_info.num_full_jacs = num_full_jacs
self._coloring_info.tol = tol
self._coloring_info.orders = orders
self._coloring_info.perturb_size = perturb_size
self._coloring_info.min_improve_pct = min_improve_pct
if self._coloring_info.static is None:
self._coloring_info.dynamic = True
else:
self._coloring_info.dynamic = False
self._coloring_info.show_summary = show_summary
self._coloring_info.show_sparsity = show_sparsity
self._coloring_info.use_scaling = use_scaling
[docs] def use_fixed_coloring(self, coloring=coloring_mod._STD_COLORING_FNAME):
"""
Tell the driver to use a precomputed coloring.
Parameters
----------
coloring : str or Coloring
A coloring filename or a Coloring object. If no arg is passed, filename will be
determined automatically.
"""
if self.supports['simultaneous_derivatives']:
if coloring_mod._force_dyn_coloring and coloring is coloring_mod._STD_COLORING_FNAME:
# force the generation of a dynamic coloring this time
self._coloring_info.dynamic = True
self._coloring_info.static = None
else:
self._coloring_info.static = coloring
self._coloring_info.dynamic = False
self._coloring_info.coloring = None
else:
raise RuntimeError("Driver '%s' does not support simultaneous derivatives." %
self._get_name())
def _setup_tot_jac_sparsity(self, coloring=None):
"""
Set up total jacobian subjac sparsity.
Drivers that can use subjac sparsity should override this.
Parameters
----------
coloring : Coloring or None
Current coloring.
"""
pass
def _get_static_coloring(self):
"""
Get the Coloring for this driver.
If necessary, load the Coloring from a file.
Returns
-------
Coloring or None
The pre-existing or loaded Coloring, or None
"""
coloring = None
info = self._coloring_info
static = info.static
if isinstance(static, coloring_mod.Coloring):
coloring = static
info.coloring = coloring
else:
coloring = info.coloring
if coloring is None and (static is coloring_mod._STD_COLORING_FNAME or
isinstance(static, str)):
if static is coloring_mod._STD_COLORING_FNAME:
fname = self._get_total_coloring_fname()
else:
fname = static
print("loading total coloring from file %s" % fname)
coloring = info.coloring = coloring_mod.Coloring.load(fname)
info.update(coloring._meta)
if coloring is not None and info.static is not None:
problem = self._problem()
if coloring._rev and problem._orig_mode not in ('rev', 'auto'):
revcol = coloring._rev[0][0]
if revcol:
raise RuntimeError("Simultaneous coloring does reverse solves but mode has "
f"been set to '{problem._orig_mode}'")
if coloring._fwd and problem._orig_mode not in ('fwd', 'auto'):
fwdcol = coloring._fwd[0][0]
if fwdcol:
raise RuntimeError("Simultaneous coloring does forward solves but mode has "
f"been set to '{problem._orig_mode}'")
return coloring
def _get_total_coloring_fname(self):
return os.path.join(self._problem().options['coloring_dir'], 'total_coloring.pkl')
[docs] def scaling_report(self, outfile='driver_scaling_report.html', title=None, show_browser=True,
jac=True):
"""
Generate a self-contained html file containing a detailed connection viewer.
Optionally pops up a web browser to view the file.
Parameters
----------
outfile : str, optional
The name of the output html file. Defaults to 'driver_scaling_report.html'.
title : str, optional
Sets the title of the web page.
show_browser : bool, optional
If True, pop up a browser to view the generated html file. Defaults to True.
jac : bool
If True, show jacobian information.
Returns
-------
dict
Data used to create html file.
"""
from openmdao.visualization.scaling_viewer.scaling_report import view_driver_scaling
# Run the model if it hasn't been run yet.
status = -1 if self._problem is None else self._problem()._metadata['setup_status']
if status < _SetupStatus.POST_FINAL_SETUP:
raise RuntimeError("Either 'run_model' or 'final_setup' must be called before the "
"scaling report can be generated.")
prob = self._problem()
if prob._run_counter < 0:
prob.run_model()
return view_driver_scaling(self, outfile=outfile, show_browser=show_browser, jac=jac,
title=title)
def _pre_run_model_debug_print(self):
"""
Optionally print some debugging information before the model runs.
"""
debug_opt = self.options['debug_print']
rank = self._problem().comm.rank
if not debug_opt or debug_opt == ['totals']:
return
if not MPI or rank == 0:
header = 'Driver debug print for iter coord: {}'.format(
self._recording_iter.get_formatted_iteration_coordinate())
print(header)
print(len(header) * '-')
if 'desvars' in debug_opt:
model = self._problem().model
desvar_vals = {n: model.get_val(n, get_remote=True, rank=0) for n in self._designvars}
if not MPI or rank == 0:
print("Design Vars")
if desvar_vals:
pprint.pprint(desvar_vals)
else:
print("None")
print()
sys.stdout.flush()
def _post_run_model_debug_print(self):
"""
Optionally print some debugging information after the model runs.
"""
rank = self._problem().comm.rank
if 'nl_cons' in self.options['debug_print']:
cons = self.get_constraint_values(lintype='nonlinear', driver_scaling=False)
if not MPI or rank == 0:
print("Nonlinear constraints")
if cons:
pprint.pprint(cons)
else:
print("None")
print()
if 'ln_cons' in self.options['debug_print']:
cons = self.get_constraint_values(lintype='linear', driver_scaling=False)
if not MPI or rank == 0:
print("Linear constraints")
if cons:
pprint.pprint(cons)
else:
print("None")
print()
if 'objs' in self.options['debug_print']:
objs = self.get_objective_values(driver_scaling=False)
if not MPI or rank == 0:
print("Objectives")
if objs:
pprint.pprint(objs)
else:
print("None")
print()
sys.stdout.flush()
[docs] def get_reports_dir(self):
"""
Get the path to the directory where the report files should go.
If it doesn't exist, it will be created.
Returns
-------
str
The path to the directory where reports should be written.
"""
return self._problem().get_reports_dir()
def _get_coloring(self, run_model=None):
"""
Get the total coloring for this driver.
If necessary, dynamically generate it.
Parameters
----------
run_model : bool or None
If False, don't run model, else use problem _run_counter to decide.
This is ignored if the coloring has already been computed.
Returns
-------
Coloring or None
Coloring object, possible loaded from a file or dynamically generated, or None
"""
if coloring_mod._use_total_sparsity:
if run_model and self._coloring_info.coloring is not None:
issue_warning("The 'run_model' argument is ignored because the coloring has "
"already been computed.")
if self._coloring_info.dynamic:
if self._coloring_info.do_compute_coloring():
self._coloring_info.coloring = \
coloring_mod.dynamic_total_coloring(self, run_model=run_model,
fname=self._get_total_coloring_fname())
return self._coloring_info.coloring
[docs]class SaveOptResult(object):
"""
A context manager that saves details about a driver run.
Parameters
----------
driver : Driver
The driver.
Attributes
----------
_driver : Driver
The driver for which we are saving results.
_start_time : float
The start time used to compute the run time.
"""
[docs] def __init__(self, driver):
"""
Initialize attributes.
"""
self._driver = driver
def __enter__(self):
"""
Set start time for the driver run.
This uses 'perf_counter()' which gives "the value (in fractional seconds)
of a performance counter, i.e. a clock with the highest available resolution
to measure a short duration. It does include time elapsed during sleep and
is system-wide."
Returns
-------
self : object
self
"""
self._start_time = time.perf_counter()
return self
def __exit__(self, *args):
"""
Save driver run information in the 'opt_result' attribute.
Parameters
----------
*args : array
Solver recording requires extra args.
"""
driver = self._driver
driver.opt_result = {
'runtime': time.perf_counter() - self._start_time,
'iter_count': driver.iter_count,
'obj_calls': driver.get_driver_objective_calls(),
'deriv_calls': driver.get_driver_derivative_calls(),
'exit_status': driver.get_exit_status()
}
[docs]class RecordingDebugging(Recording):
"""
A class that acts as a context manager.
Handles doing the case recording and also the Driver
debugging printing.
Parameters
----------
name : str
Name of object getting recorded.
iter_count : int
Current counter of iterations completed.
recording_requester : object
Object to which this recorder is attached.
"""
def __enter__(self):
"""
Do things before the code inside the 'with RecordingDebugging' block.
Returns
-------
self : object
self
"""
super().__enter__()
self.recording_requester()._pre_run_model_debug_print()
return self
def __exit__(self, *args):
"""
Do things after the code inside the 'with RecordingDebugging' block.
Parameters
----------
*args : array
Solver recording requires extra args.
"""
self.recording_requester()._post_run_model_debug_print()
super().__exit__()
[docs]def record_iteration(requester, prob, case_name):
"""
Record an iteration of the current Problem or Driver.
Parameters
----------
requester : Problem or Driver
The recording requester.
prob : Problem
The Problem.
case_name : str
The name of this case.
"""
rec_mgr = requester._rec_mgr
if not rec_mgr._recorders:
return
# Get the data to record (collective calls that get across all ranks)
model = prob.model
parallel = rec_mgr._check_parallel() if model.comm.size > 1 else False
inputs, outputs, residuals = model.get_nonlinear_vectors()
discrete_inputs = model._discrete_inputs
discrete_outputs = model._discrete_outputs
opts = requester.recording_options
data = {'input': {}, 'output': {}, 'residual': {}}
filt = requester._filtered_vars_to_record
if opts['record_inputs'] and (inputs._names or len(discrete_inputs) > 0):
data['input'] = model._retrieve_data_of_kind(filt, 'input', 'nonlinear', parallel)
if opts['record_outputs'] and (outputs._names or len(discrete_outputs) > 0):
data['output'] = model._retrieve_data_of_kind(filt, 'output', 'nonlinear', parallel)
if opts['record_residuals'] and residuals._names:
data['residual'] = model._retrieve_data_of_kind(filt, 'residual', 'nonlinear', parallel)
from openmdao.core.problem import Problem
if isinstance(requester, Problem):
# Record total derivatives
if opts['record_derivatives'] and prob.driver._designvars and prob.driver._responses:
data['totals'] = requester.compute_totals(return_format='flat_dict_structured_key')
# Record solver info
if opts['record_abs_error'] or opts['record_rel_error']:
norm = residuals.get_norm()
if opts['record_abs_error']:
data['abs'] = norm
if opts['record_rel_error']:
solver = model.nonlinear_solver
norm0 = solver._norm0 if solver._norm0 != 0.0 else 1.0 # runonce never sets _norm0
data['rel'] = norm / norm0
rec_mgr.record_iteration(requester, data, requester._get_recorder_metadata(case_name))
