class Solver(object): """ Base solver class. This class is subclassed by NonlinearSolver and LinearSolver, which are in turn subclassed by actual solver implementations. Attributes ---------- _system : <System> Pointer to the owning system. _depth : int How many subsolvers deep this solver is (0 means not a subsolver). _vec_names : [str, ...] List of right-hand-side (RHS) vector names. _mode : str 'fwd' or 'rev', applicable to linear solvers only. _iter_count : int Number of iterations for the current invocation of the solver. _rec_mgr : <RecordingManager> object that manages all recorders added to this solver cite : str Listing of relevant citations that should be referenced when publishing work that uses this class. options : <OptionsDictionary> Options dictionary. recording_options : <OptionsDictionary> Recording options dictionary. supports : <OptionsDictionary> Options dictionary describing what features are supported by this solver. _filtered_vars_to_record: Dict Dict of list of var names to record _norm0: float Normalization factor _solver_info : SolverInfo A stack-like object shared by all Solvers in the model. """ # Object to store some formatting for iprint that is shared across all solvers. SOLVER = 'base_solver' def __init__(self, **kwargs): """ Initialize all attributes. Parameters ---------- **kwargs : dict of keyword arguments Keyword arguments that will be mapped into the Solver options. """ self._system = None self._depth = 0 self._vec_names = None self._mode = 'fwd' self._iter_count = 0 self._solver_info = None # Solver options self.options = OptionsDictionary(parent_name=self.msginfo) self.options.declare('maxiter', types=int, default=10, desc='maximum number of iterations') self.options.declare('atol', default=1e-10, desc='absolute error tolerance') self.options.declare('rtol', default=1e-10, desc='relative error tolerance') self.options.declare('iprint', types=int, default=1, desc='whether to print output') self.options.declare('err_on_maxiter', types=bool, default=None, allow_none=True, desc="Deprecated. Use 'err_on_non_converge'.") self.options.declare('err_on_non_converge', types=bool, default=False, desc="When True, AnalysisError will be raised if we don't converge.") # Case recording options self.recording_options = OptionsDictionary(parent_name=self.msginfo) self.recording_options.declare('record_abs_error', types=bool, default=True, desc='Set to True to record absolute error at the \ solver level') self.recording_options.declare('record_rel_error', types=bool, default=True, desc='Set to True to record relative error at the \ solver level') self.recording_options.declare('record_inputs', types=bool, default=True, desc='Set to True to record inputs at the solver level') self.recording_options.declare('record_outputs', types=bool, default=True, desc='Set to True to record outputs at the solver level') self.recording_options.declare('record_solver_residuals', types=bool, default=False, desc='Set to True to record residuals at the solver level') self.recording_options.declare('record_metadata', types=bool, desc='Record metadata', default=True) self.recording_options.declare('includes', types=list, default=['*'], desc='Patterns for variables to include in recording') self.recording_options.declare('excludes', types=list, default=[], desc='Patterns for vars to exclude in recording ' '(processed post-includes)') # Case recording related self._filtered_vars_to_record = {} self._norm0 = 0.0 # What the solver supports. self.supports = OptionsDictionary(parent_name=self.msginfo) self.supports.declare('gradients', types=bool, default=False) self.supports.declare('implicit_components', types=bool, default=False) self._declare_options() self.options.update(kwargs) self._rec_mgr = RecordingManager() self.cite = "" @property def msginfo(self): """ Return info to prepend to messages. Returns ------- str Info to prepend to messages. """ if self._system is None: return type(self).__name__ return '{} in {}'.format(type(self).__name__, self._system().msginfo) def _assembled_jac_solver_iter(self): """ Return an empty generator of lin solvers using assembled jacs. """ for i in (): yield def add_recorder(self, recorder): """ Add a recorder to the solver's RecordingManager. Parameters ---------- recorder : <CaseRecorder> A recorder instance to be added to RecManager. """ if MPI: raise RuntimeError( "Recording of Solvers when running parallel code is not supported yet") self._rec_mgr.append(recorder) def _declare_options(self): """ Declare options before kwargs are processed in the init method. This is optionally implemented by subclasses of Solver. """ pass def _setup_solvers(self, system, depth): """ Assign system instance, set depth, and optionally perform setup. Parameters ---------- system : <System> pointer to the owning system. depth : int depth of the current system (already incremented). """ self._system = weakref.ref(system) self._depth = depth self._solver_info = system._solver_info self._recording_iter = system._recording_iter if system.pathname: parent_name = self.msginfo self.options._parent_name = parent_name self.recording_options._parent_name = parent_name self.supports._parent_name = parent_name if isinstance(self, LinearSolver) and not system._use_derivatives: return self._rec_mgr.startup(self) self._rec_mgr.record_metadata(self) myoutputs = myresiduals = myinputs = set() incl = self.recording_options['includes'] excl = self.recording_options['excludes'] if self.recording_options['record_solver_residuals']: if isinstance(self, NonlinearSolver): residuals = system._residuals else: # it's a LinearSolver residuals = system._vectors['residual']['linear'] myresiduals = {n for n in residuals._names if check_path(n, incl, excl)} if self.recording_options['record_outputs']: if isinstance(self, NonlinearSolver): outputs = system._outputs else: # it's a LinearSolver outputs = system._vectors['output']['linear'] myoutputs = {n for n in outputs._names if check_path(n, incl, excl)} if self.recording_options['record_inputs']: if isinstance(self, NonlinearSolver): inputs = system._inputs else: inputs = system._vectors['input']['linear'] myinputs = {n for n in inputs._names if check_path(n, incl, excl)} self._filtered_vars_to_record = { 'in': myinputs, 'out': myoutputs, 'res': myresiduals } # Raise a deprecation warning for changed option. if 'err_on_maxiter' in self.options and self.options['err_on_maxiter'] is not None: self.options['err_on_non_converge'] = self.options['err_on_maxiter'] warn_deprecation("The 'err_on_maxiter' option provides backwards compatibility " "with earlier version of OpenMDAO; use options['err_on_non_converge'] " "instead.") def _set_solver_print(self, level=2, type_='all'): """ Control printing for solvers and subsolvers in the model. Parameters ---------- level : int iprint level. Set to 2 to print residuals each iteration; set to 1 to print just the iteration totals; set to 0 to disable all printing except for failures, and set to -1 to disable all printing including failures. type_ : str Type of solver to set: 'LN' for linear, 'NL' for nonlinear, or 'all' for all. """ self.options['iprint'] = level def _mpi_print(self, iteration, abs_res, rel_res): """ Print residuals from an iteration. Parameters ---------- iteration : int iteration counter, 0-based. abs_res : float current absolute residual norm. rel_res : float current relative residual norm. """ if (self.options['iprint'] == 2 and self._system().comm.rank == 0): prefix = self._solver_info.prefix solver_name = self.SOLVER if prefix.endswith('precon:'): solver_name = solver_name[3:] print_str = prefix + solver_name print_str += ' %d ; %.9g %.9g' % (iteration, abs_res, rel_res) print(print_str) def _mpi_print_header(self): """ Print header text before solving. """ pass def _solve(self): """ Run the iterative solver. """ maxiter = self.options['maxiter'] atol = self.options['atol'] rtol = self.options['rtol'] iprint = self.options['iprint'] self._mpi_print_header() self._iter_count = 0 norm0, norm = self._iter_initialize() self._norm0 = norm0 self._mpi_print(self._iter_count, norm, norm / norm0) while self._iter_count < maxiter and norm > atol and norm / norm0 > rtol: with Recording(type(self).__name__, self._iter_count, self) as rec: self._single_iteration() self._iter_count += 1 self._run_apply() norm = self._iter_get_norm() # With solvers, we want to record the norm AFTER the call, but the call needs to # be wrapped in the with for stack purposes, so we locally assign norm & norm0 # into the class. rec.abs = norm rec.rel = norm / norm0 if norm0 == 0: norm0 = 1 self._mpi_print(self._iter_count, norm, norm / norm0) system = self._system() if system.comm.rank == 0 or os.environ.get('USE_PROC_FILES'): prefix = self._solver_info.prefix + self.SOLVER # Solver terminated early because a Nan in the norm doesn't satisfy the while-loop # conditionals. if np.isinf(norm) or np.isnan(norm): msg = "Solver '{}' on system '{}': residuals contain 'inf' or 'NaN' after {} " + \ "iterations." if iprint > -1: print(prefix + msg.format(self.SOLVER, system.pathname, self._iter_count)) # Raise AnalysisError if requested. if self.options['err_on_non_converge']: raise AnalysisError(msg.format(self.SOLVER, system.pathname, self._iter_count)) # Solver hit maxiter without meeting desired tolerances. elif (norm > atol and norm / norm0 > rtol): msg = "Solver '{}' on system '{}' failed to converge in {} iterations." if iprint > -1: print(prefix + msg.format(self.SOLVER, system.pathname, self._iter_count)) # Raise AnalysisError if requested. if self.options['err_on_non_converge']: raise AnalysisError(msg.format(self.SOLVER, system.pathname, self._iter_count)) # Solver converged elif iprint == 1: print(prefix + ' Converged in {} iterations'.format(self._iter_count)) elif iprint == 2: print(prefix + ' Converged') def _iter_initialize(self): """ Perform any necessary pre-processing operations. Returns ------- float initial error. float error at the first iteration. """ pass def _run_apply(self): """ Run the appropriate apply method on the system. """ pass def _linearize(self): """ Perform any required linearization operations such as matrix factorization. """ pass def _linearize_children(self): """ Return a flag that is True when we need to call linearize on our subsystems' solvers. Returns ------- boolean Flag for indicating child linerization """ return True def __str__(self): """ Return a string representation of the solver. Returns ------- str String representation of the solver. """ return self.SOLVER def record_iteration(self, **kwargs): """ Record an iteration of the current Solver. Parameters ---------- **kwargs : dict Keyword arguments (used for abs and rel error). """ if not self._rec_mgr._recorders: return metadata = create_local_meta(self.SOLVER) # Get the data data = {} if self.recording_options['record_abs_error']: data['abs'] = kwargs.get('abs') else: data['abs'] = None if self.recording_options['record_rel_error']: data['rel'] = kwargs.get('rel') else: data['rel'] = None system = self._system() if isinstance(self, NonlinearSolver): outputs = system._outputs inputs = system._inputs residuals = system._residuals else: # it's a LinearSolver outputs = system._vectors['output']['linear'] inputs = system._vectors['input']['linear'] residuals = system._vectors['residual']['linear'] if self.recording_options['record_outputs']: data['o'] = {} if 'out' in self._filtered_vars_to_record: for out in self._filtered_vars_to_record['out']: if out in outputs._names: data['o'][out] = outputs._views[out] else: data['o'] = outputs else: data['o'] = None if self.recording_options['record_inputs']: data['i'] = {} if 'in' in self._filtered_vars_to_record: for inp in self._filtered_vars_to_record['in']: if inp in inputs._names: data['i'][inp] = inputs._views[inp] else: data['i'] = inputs else: data['i'] = None if self.recording_options['record_solver_residuals']: data['r'] = {} if 'res' in self._filtered_vars_to_record: for res in self._filtered_vars_to_record['res']: if res in residuals._names: data['r'][res] = residuals._views[res] else: data['r'] = residuals else: data['r'] = None self._rec_mgr.record_iteration(self, data, metadata) def cleanup(self): """ Clean up resources prior to exit. """ # shut down all recorders self._rec_mgr.shutdown() def _set_complex_step_mode(self, active): """ Turn on or off complex stepping mode. Recurses to turn on or off complex stepping mode in all subsystems and their vectors. Parameters ---------- active : bool Complex mode flag; set to True prior to commencing complex step. """ pass
class Solver(object): """ Base solver class. This class is subclassed by NonlinearSolver and LinearSolver, which are in turn subclassed by actual solver implementations. Attributes ---------- _system : <System> Pointer to the owning system. _depth : int How many subsolvers deep this solver is (0 means not a subsolver). _vec_names : [str, ...] List of right-hand-side (RHS) vector names. _mode : str 'fwd' or 'rev', applicable to linear solvers only. _iter_count : int Number of iterations for the current invocation of the solver. _rec_mgr : <RecordingManager> object that manages all recorders added to this solver cite : str Listing of relevant citations that should be referenced when publishing work that uses this class. options : <OptionsDictionary> Options dictionary. recording_options : <OptionsDictionary> Recording options dictionary. supports : <OptionsDictionary> Options dictionary describing what features are supported by this solver. _filtered_vars_to_record : Dict Dict of list of var names to record _norm0 : float Normalization factor _problem_meta : dict Problem level metadata. """ # Object to store some formatting for iprint that is shared across all solvers. SOLVER = 'base_solver' def __init__(self, **kwargs): """ Initialize all attributes. Parameters ---------- **kwargs : dict of keyword arguments Keyword arguments that will be mapped into the Solver options. """ self._system = None self._depth = 0 self._vec_names = None self._mode = 'fwd' self._iter_count = 0 self._problem_meta = None # Solver options self.options = OptionsDictionary(parent_name=self.msginfo) self.options.declare('maxiter', types=int, default=10, desc='maximum number of iterations') self.options.declare('atol', default=1e-10, desc='absolute error tolerance') self.options.declare('rtol', default=1e-10, desc='relative error tolerance') self.options.declare('iprint', types=int, default=1, desc='whether to print output') self.options.declare( 'err_on_non_converge', types=bool, default=False, desc="When True, AnalysisError will be raised if we don't converge." ) # Case recording options self.recording_options = OptionsDictionary(parent_name=self.msginfo) self.recording_options.declare( 'record_abs_error', types=bool, default=True, desc='Set to True to record absolute error at the \ solver level') self.recording_options.declare( 'record_rel_error', types=bool, default=True, desc='Set to True to record relative error at the \ solver level') self.recording_options.declare( 'record_inputs', types=bool, default=True, desc='Set to True to record inputs at the solver level') self.recording_options.declare( 'record_outputs', types=bool, default=True, desc='Set to True to record outputs at the solver level') self.recording_options.declare( 'record_solver_residuals', types=bool, default=False, desc='Set to True to record residuals at the solver level') self.recording_options.declare( 'record_metadata', types=bool, desc='Deprecated. Recording ' 'of metadata will always be done', deprecation="The recording option, record_metadata, on " "Solver is " "deprecated. Recording of metadata will always be done", default=True) self.recording_options.declare( 'includes', types=list, default=['*'], desc="Patterns for variables to include in recording. \ Paths are relative to solver's Group. \ Uses fnmatch wildcards") self.recording_options.declare( 'excludes', types=list, default=[], desc="Patterns for vars to exclude in recording. \ (processed post-includes) \ Paths are relative to solver's Group. \ Uses fnmatch wildcards") # Case recording related self._filtered_vars_to_record = {} self._norm0 = 0.0 # What the solver supports. self.supports = OptionsDictionary(parent_name=self.msginfo) self.supports.declare('gradients', types=bool, default=False) self.supports.declare('implicit_components', types=bool, default=False) self._declare_options() self.options.update(kwargs) self._rec_mgr = RecordingManager() self.cite = "" @property def msginfo(self): """ Return info to prepend to messages. Returns ------- str Info to prepend to messages. """ if self._system is None: return type(self).__name__ return '{} in {}'.format(type(self).__name__, self._system().msginfo) @property def _recording_iter(self): if self._problem_meta is None: raise RuntimeError( f"{self.msginfo}: Can't access recording_iter because " "_setup_solvers has not been called.") return self._problem_meta['recording_iter'] @property def _solver_info(self): if self._problem_meta is None: raise RuntimeError( f"{self.msginfo}: Can't access solver_info because _setup_solvers " "has not been called.") return self._problem_meta['solver_info'] def _assembled_jac_solver_iter(self): """ Return an empty generator of lin solvers using assembled jacs. """ for i in (): yield def add_recorder(self, recorder): """ Add a recorder to the solver's RecordingManager. Parameters ---------- recorder : <CaseRecorder> A recorder instance to be added to RecManager. """ if MPI: raise RuntimeError( "Recording of Solvers when running parallel code is not supported yet" ) self._rec_mgr.append(recorder) def _declare_options(self): """ Declare options before kwargs are processed in the init method. This is optionally implemented by subclasses of Solver. """ pass def _setup_solvers(self, system, depth): """ Assign system instance, set depth, and optionally perform setup. Parameters ---------- system : <System> pointer to the owning system. depth : int depth of the current system (already incremented). """ self._system = weakref.ref(system) self._depth = depth self._problem_meta = system._problem_meta if system.pathname: parent_name = self.msginfo self.options._parent_name = parent_name self.recording_options._parent_name = parent_name self.supports._parent_name = parent_name if isinstance(self, LinearSolver) and not system._use_derivatives: return self._rec_mgr.startup(self) myoutputs = myresiduals = myinputs = [] incl = self.recording_options['includes'] excl = self.recording_options['excludes'] # doesn't matter if we're a linear or nonlinear solver. The names for # inputs, outputs, and residuals are the same for both the 'linear' and 'nonlinear' # vectors. if system.pathname: incl = ['.'.join((system.pathname, i)) for i in incl] excl = ['.'.join((system.pathname, i)) for i in excl] if self.recording_options['record_solver_residuals']: myresiduals = [ n for n in system._residuals._abs_iter() if check_path(n, incl, excl) ] if self.recording_options['record_outputs']: myoutputs = [ n for n in system._outputs._abs_iter() if check_path(n, incl, excl) ] if self.recording_options['record_inputs']: myinputs = [ n for n in system._inputs._abs_iter() if check_path(n, incl, excl) ] self._filtered_vars_to_record = { 'input': myinputs, 'output': myoutputs, 'residual': myresiduals } def _set_solver_print(self, level=2, type_='all'): """ Control printing for solvers and subsolvers in the model. Parameters ---------- level : int iprint level. Set to 2 to print residuals each iteration; set to 1 to print just the iteration totals; set to 0 to disable all printing except for failures, and set to -1 to disable all printing including failures. type_ : str Type of solver to set: 'LN' for linear, 'NL' for nonlinear, or 'all' for all. """ self.options['iprint'] = level def _mpi_print(self, iteration, abs_res, rel_res): """ Print residuals from an iteration. Parameters ---------- iteration : int iteration counter, 0-based. abs_res : float current absolute residual norm. rel_res : float current relative residual norm. """ if (self.options['iprint'] == 2 and (self._system().comm.rank == 0 or os.environ.get('USE_PROC_FILES'))): prefix = self._solver_info.prefix solver_name = self.SOLVER if prefix.endswith('precon:'): solver_name = solver_name[3:] print_str = prefix + solver_name print_str += ' %d ; %.9g %.9g' % (iteration, abs_res, rel_res) print(print_str) def _mpi_print_header(self): """ Print header text before solving. """ if (self.options['iprint'] > 0 and (self._system().comm.rank == 0 or os.environ.get('USE_PROC_FILES'))): pathname = self._system().pathname if pathname: nchar = len(pathname) prefix = self._solver_info.prefix header = prefix + "\n" header += prefix + nchar * "=" + "\n" header += prefix + pathname + "\n" header += prefix + nchar * "=" print(header) def _iter_initialize(self): """ Perform any necessary pre-processing operations. Returns ------- float initial error. float error at the first iteration. """ pass def _run_apply(self): """ Run the appropriate apply method on the system. """ pass def _linearize(self): """ Perform any required linearization operations such as matrix factorization. """ pass def _linearize_children(self): """ Return a flag that is True when we need to call linearize on our subsystems' solvers. Returns ------- boolean Flag for indicating child linerization """ return True def __str__(self): """ Return a string representation of the solver. Returns ------- str String representation of the solver. """ return self.SOLVER def record_iteration(self, **kwargs): """ Record an iteration of the current Solver. Parameters ---------- **kwargs : dict Keyword arguments (used for abs and rel error). """ if not self._rec_mgr._recorders: return metadata = create_local_meta(self.SOLVER) # Get the data data = { 'abs': kwargs.get('abs') if self.recording_options['record_abs_error'] else None, 'rel': kwargs.get('rel') if self.recording_options['record_rel_error'] else None, 'input': {}, 'output': {}, 'residual': {} } system = self._system() vec_name = 'nonlinear' if isinstance(self, NonlinearSolver) else 'linear' filt = self._filtered_vars_to_record parallel = self._rec_mgr._check_parallel( ) if system.comm.size > 1 else False if self.recording_options['record_outputs']: data['output'] = system._retrieve_data_of_kind( filt, 'output', vec_name, parallel) if self.recording_options['record_inputs']: data['input'] = system._retrieve_data_of_kind( filt, 'input', vec_name, parallel) if self.recording_options['record_solver_residuals']: data['residual'] = system._retrieve_data_of_kind( filt, 'residual', vec_name, parallel) self._rec_mgr.record_iteration(self, data, metadata) def cleanup(self): """ Clean up resources prior to exit. """ # shut down all recorders self._rec_mgr.shutdown() def _set_complex_step_mode(self, active): """ Turn on or off complex stepping mode. Recurses to turn on or off complex stepping mode in all subsystems and their vectors. Parameters ---------- active : bool Complex mode flag; set to True prior to commencing complex step. """ pass def _disallow_distrib_solve(self): """ Raise an exception if our system or any subsystems are distributed or non-local. """ s = self._system() if s.comm.size == 1: return from openmdao.core.group import Group if s._has_distrib_vars or (isinstance(s, Group) and s._contains_parallel_group): msg = "{} linear solver in {} cannot be used in or above a ParallelGroup or a " + \ "distributed component." raise RuntimeError(msg.format(type(self).__name__, s.msginfo))
class Driver(object): """ Top-level container for the systems and drivers. 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 rank, 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_sparsity : dict, str, or None Specifies sparsity of sub-jacobians of the total jacobian. Only used by pyOptSparseDriver. _res_jacs : dict Dict of sparse subjacobians for use with certain optimizers, e.g. pyOptSparseDriver. _total_jac : _TotalJacInfo or None Cached total jacobian handling object. """ def __init__(self, **kwargs): """ Initialize the driver. Parameters ---------- **kwargs : dict of keyword arguments Keyword arguments that will be mapped into the Driver options. """ 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, check_valid=_check_debug_print_opts_valid, desc="List of what type of Driver variables to print at each " "iteration. Valid items in list are 'desvars', 'ln_cons', " "'nl_cons', 'objs', 'totals'", default=[]) # Case recording options self.recording_options = OptionsDictionary( parent_name=type(self).__name__) self.recording_options.declare( 'record_model_metadata', types=bool, default=True, desc='Deprecated. Recording of model metadata will always ' 'be done', deprecation="The recording option, record_model_metadata, " "on Driver is " "deprecated. Recording of model metadata will always " "be done", ) 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('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.iter_count = 0 self.cite = "" self._coloring_info = coloring_mod._get_coloring_meta() self._total_jac_sparsity = None self._res_jacs = {} self._total_jac = None self.fail = False self._declare_options() self.options.update(kwargs) @property def msginfo(self): """ Return info to prepend to messages. Returns ------- str Info to prepend to messages. """ return type(self).__name__ def add_recorder(self, recorder): """ Add a recorder to the driver. Parameters ---------- recorder : CaseRecorder A recorder instance. """ self._rec_mgr.append(recorder) 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 _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. """ self._problem = weakref.ref(problem) model = problem.model self._total_jac = None self._has_scaling = (np.any( [r['total_scaler'] is not None for r in self._responses.values()]) or np.any([ dv['total_scaler'] is not None for dv in self._designvars.values() ])) # Determine if any design variables are discrete. self._designvars_discrete = [ name for name, meta in self._designvars.items() if meta['ivc_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) con_set = set() obj_set = set() dv_set = set() self._remote_dvs = remote_dv_dict = {} self._remote_cons = remote_con_dict = {} self._dist_driver_vars = dist_dict = {} self._remote_objs = remote_obj_dict = {} src_design_vars = prom2ivc_src_dict(self._designvars) src_cons = prom2ivc_src_dict(self._cons) src_objs = prom2ivc_src_dict(self._objs) responses = prom2ivc_src_dict(self._responses) # Now determine if later we'll need to allgather cons, objs, or desvars. if model.comm.size > 1 and model._subsystems_allprocs: local_out_vars = set(model._outputs._abs_iter()) remote_dvs = set(src_design_vars) - local_out_vars remote_cons = set(src_cons) - local_out_vars remote_objs = set(src_objs) - local_out_vars 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 sizes = model._var_sizes['nonlinear']['output'] abs2meta = model._var_allprocs_abs2meta rank = model.comm.rank nprocs = model.comm.size for i, vname in enumerate(model._var_allprocs_abs_names['output']): if vname in responses: indices = responses[vname].get('indices') elif vname in src_design_vars: indices = src_design_vars[vname].get('indices') else: continue if abs2meta[vname]['distributed']: idx = model._var_allprocs_abs2idx['nonlinear'][vname] dist_sizes = model._var_sizes['nonlinear']['output'][:, idx] total_dist_size = np.sum(dist_sizes) # Determine which indices are on our proc. offsets = sizes2offsets(dist_sizes) if indices is not None: indices = convert_neg(indices, total_dist_size) 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]))] if irank == rank: local_indices = dist_inds - offsets[rank] distrib_indices = dist_inds true_sizes[irank] = dist_inds.size dist_dict[vname] = (local_indices, true_sizes, distrib_indices) else: dist_dict[vname] = (_full_slice, dist_sizes, slice( offsets[rank], offsets[rank] + dist_sizes[rank])) else: owner = owning_ranks[vname] sz = sizes[owner, i] if vname in dv_set: remote_dv_dict[vname] = (owner, sz) if vname in con_set: remote_con_dict[vname] = (owner, sz) if vname 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) self._setup_simul_coloring() 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 _get_vars_to_record(self, recording_options): """ Get variables to record based on recording options. Parameters ---------- recording_options : <OptionsDictionary> Dictionary with recording options. Returns ------- dict Dictionary containing lists of variables to record. """ problem = self._problem() model = problem.model incl = recording_options['includes'] excl = recording_options['excludes'] # includes and excludes for outputs are specified using promoted names abs2prom = model._var_allprocs_abs2prom['output'] # 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']: myoutputs = sorted([ n for n, prom in abs2prom.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']: myresiduals = [ n for n in model._residuals._abs_iter() if check_path(abs2prom[n], incl, excl) ] myoutputs = set(myoutputs) if recording_options['record_desvars']: myoutputs.update(self._designvars) if recording_options['record_objectives'] or recording_options[ 'record_responses']: myoutputs.update(self._objs) if recording_options['record_constraints'] or recording_options[ 'record_responses']: myoutputs.update(self._cons) # inputs (if in options). inputs use _absolute_ names for includes/excludes if 'record_inputs' in recording_options: if recording_options['record_inputs']: # sort the results since _var_allprocs_abs2prom isn't ordered myinputs = sorted([ n for n in model._var_allprocs_abs2prom['input'] if check_path(n, incl, excl) ]) vars2record = { 'input': myinputs, 'output': list(myoutputs), 'residual': myresiduals } return vars2record def _setup_recording(self): """ Set up case recording. """ self._filtered_vars_to_record = self._get_vars_to_record( self.recording_options) self._rec_mgr.startup(self) def _get_voi_val(self, name, meta, remote_vois, driver_scaling=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. 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 distributed_vars = self._dist_driver_vars indices = meta['indices'] if meta.get('ivc_source') is not None: src_name = meta['ivc_source'] else: src_name = name if MPI: distributed = comm.size > 0 and src_name in distributed_vars else: distributed = False if src_name in remote_vois: owner, size = remote_vois[src_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=True, rank=rank, flat=True) if indices is not None: val = val[indices] else: if owner == comm.rank: if indices is None: val = get(name).copy() else: val = get(name)[indices] else: if indices is not None: size = len(indices) val = np.empty(size) comm.Bcast(val, root=owner) elif distributed: local_val = model.get_val(src_name, flat=True) local_indices, sizes, _ = distributed_vars[src_name] if local_indices is not None: local_val = local_val[local_indices] 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: if name in self._designvars_discrete: val = model._discrete_outputs[src_name] # At present, only integers are supported by OpenMDAO drivers. # We check the values here. msg = "Only integer scalars or ndarrays are supported as values for " + \ "discrete variables when used as a design variable. " if np.isscalar(val) and not isinstance(val, (int, np.integer)): msg += "A value of type '{}' was specified.".format( val.__class__.__name__) raise ValueError(msg) elif isinstance(val, np.ndarray) and not np.issubdtype( val[0], np.integer): msg += "An array of type '{}' was specified.".format( val[0].__class__.__name__) raise ValueError(msg) elif indices is None: val = get(src_name).copy() else: val = get(src_name)[indices] 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 def get_design_var_values(self): """ Return the design variable values. Returns ------- dict Dictionary containing values of each design variable. """ return { n: self._get_voi_val(n, dv, self._remote_dvs) for n, dv in self._designvars.items() } def set_design_var(self, name, value): """ Set the value of a design variable. Parameters ---------- name : str Global pathname of the design variable. value : float or ndarray Value for the design variable. """ problem = self._problem() meta = self._designvars[name] src_name = meta['ivc_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 indices = meta['indices'] if indices is None: indices = _full_slice 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) else: value = value.astype(np.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 src_name in self._dist_driver_vars: loc_idxs, _, dist_idxs = self._dist_driver_vars[src_name] else: loc_idxs = indices dist_idxs = _full_slice desvar[loc_idxs] = np.atleast_1d(value)[dist_idxs] # 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 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() } def get_constraint_values(self, ctype='all', lintype='all', driver_scaling=True): """ Return constraint values. Parameters ---------- ctype : string Default is 'all'. Optionally return just the inequality constraints with 'ineq' or the equality constraints with 'eq'. lintype : string 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 = OrderedDict() self._cons = cons = OrderedDict() model._setup_driver_units() 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 = sum(resps[n]['size'] for n in self._get_ordered_nl_responses()) # Gather up the information for design vars. self._designvars = designvars = model.get_design_vars( recurse=True, use_prom_ivc=True) desvar_size = sum(data['size'] for data in designvars.values()) return response_size, desvar_size def run(self): """ Execute this driver. The base `Driver` just runs the model. All other drivers overload this method. Returns ------- boolean 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', global_names=None, use_abs_names=True): """ Compute derivatives of desired quantities with respect to desired inputs. All derivatives are returned using driver scaling. Parameters ---------- of : list of variable name strings or None Variables whose derivatives will be computed. Default is None, which uses the driver's objectives and constraints. wrt : list of variable name strings or None Variables with respect to which the derivatives will be computed. Default is None, which uses the driver's desvars. return_format : string 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. global_names : bool Deprecated. Use 'use_abs_names' instead. use_abs_names : bool Set to True when passing in absolute names to skip some translation steps. Returns ------- derivs : object Derivatives in form requested by 'return_format'. """ problem = self._problem() total_jac = self._total_jac 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 global_names is not None: warn_deprecation( "'global_names' is deprecated in calls to _compute_totals. " "Use 'use_abs_names' instead.") use_abs_names = global_names if problem.model._owns_approx_jac: self._recording_iter.push(('_compute_totals_approx', 0)) try: if total_jac is None: total_jac = _TotalJacInfo(problem, of, wrt, use_abs_names, return_format, approx=True, debug_print=debug_print) # Don't cache linear constraint jacobian if not total_jac.has_lin_cons: self._total_jac = total_jac totals = total_jac.compute_totals_approx(initialize=True) else: totals = total_jac.compute_totals_approx() finally: self._recording_iter.pop() else: if total_jac is None: total_jac = _TotalJacInfo(problem, of, wrt, use_abs_names, return_format, debug_print=debug_print) # don't cache linear constraint jacobian if not total_jac.has_lin_cons: self._total_jac = total_jac self._recording_iter.push(('_compute_totals', 0)) try: totals = total_jac.compute_totals() finally: self._recording_iter.pop() if self._rec_mgr._recorders and self.recording_options[ 'record_derivatives']: metadata = create_local_meta(self._get_name()) total_jac.record_derivatives(self, metadata) return totals def record_iteration(self): """ Record an iteration of the current Driver. """ record_iteration(self, self._problem(), self._get_name()) 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" 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'] ): """ 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. """ 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['coloring'] = None self._coloring_info['show_summary'] = show_summary self._coloring_info['show_sparsity'] = show_sparsity def use_fixed_coloring(self, coloring=coloring_mod._STD_COLORING_FNAME): """ Tell the driver to use a precomputed coloring. Parameters ---------- coloring : str A coloring filename. 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 """ 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 not None: return coloring if 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) return coloring def _get_total_coloring_fname(self): return os.path.join(self._problem().options['coloring_dir'], 'total_coloring.pkl') def _setup_simul_coloring(self): """ Set up metadata for coloring of total derivative solution. If set_coloring was called with a filename, load the coloring file. """ # command line simul_coloring uses this env var to turn pre-existing coloring off if not coloring_mod._use_total_sparsity: return problem = self._problem() if not problem.model._use_derivatives: simple_warning( "Derivatives are turned off. Skipping simul deriv coloring.") return total_coloring = self._get_static_coloring() if total_coloring._rev and problem._orig_mode not in ('rev', 'auto'): revcol = total_coloring._rev[0][0] if revcol: raise RuntimeError( "Simultaneous coloring does reverse solves but mode has " "been set to '%s'" % problem._orig_mode) if total_coloring._fwd and problem._orig_mode not in ('fwd', 'auto'): fwdcol = total_coloring._fwd[0][0] if fwdcol: raise RuntimeError( "Simultaneous coloring does forward solves but mode has " "been set to '%s'" % problem._orig_mode) 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()
class Driver(object): """ Top-level container for the systems and drivers. 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 : <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. _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_sparsity : dict, str, or None Specifies sparsity of sub-jacobians of the total jacobian. Only used by pyOptSparseDriver. _res_jacs : dict Dict of sparse subjacobians for use with certain optimizers, e.g. pyOptSparseDriver. _total_jac : _TotalJacInfo or None Cached total jacobian handling object. """ def __init__(self, **kwargs): """ Initialize the driver. Parameters ---------- **kwargs : dict of keyword arguments Keyword arguments that will be mapped into the Driver options. """ 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, check_valid=_check_debug_print_opts_valid, desc="List of what type of Driver variables to print at each " "iteration. Valid items in list are 'desvars', 'ln_cons', " "'nl_cons', 'objs', 'totals'", default=[]) # Case recording options self.recording_options = OptionsDictionary( parent_name=type(self).__name__) self.recording_options.declare( 'record_model_metadata', types=bool, default=True, desc='Record metadata for all Systems in the model') 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 to True to record responses 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') self.recording_options.declare( 'excludes', types=list, default=[], desc='Patterns for vars to exclude in recording ' '(processed post-includes)') 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') # What the driver supports. self.supports = OptionsDictionary(parent_name=type(self).__name__) 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=False) 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.iter_count = 0 self.cite = "" self._coloring_info = coloring_mod._DEF_COMP_SPARSITY_ARGS.copy() self._coloring_info['coloring'] = None self._coloring_info['dynamic'] = False self._coloring_info['static'] = None self._total_jac_sparsity = None self._res_jacs = {} self._total_jac = None self.fail = False self._declare_options() self.options.update(kwargs) @property def msginfo(self): """ Return info to prepend to messages. Returns ------- str Info to prepend to messages. """ return type(self).__name__ def add_recorder(self, recorder): """ Add a recorder to the driver. Parameters ---------- recorder : CaseRecorder A recorder instance. """ self._rec_mgr.append(recorder) 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 _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. """ self._problem = problem self._recording_iter = problem._recording_iter model = problem.model self._total_jac = None self._has_scaling = (np.any([ r['scaler'] is not None for r in itervalues(self._responses) ]) or np.any( [dv['scaler'] is not None for dv in itervalues(self._designvars)])) # Determine if any design variables are discrete. self._designvars_discrete = [ dv for dv in self._designvars if dv 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) con_set = set() obj_set = set() dv_set = set() self._remote_dvs = dv_dict = {} self._remote_cons = con_dict = {} self._remote_objs = obj_dict = {} # Now determine if later we'll need to allgather cons, objs, or desvars. if model.comm.size > 1 and model._subsystems_allprocs: local_out_vars = set(model._outputs._views) remote_dvs = set(self._designvars) - local_out_vars remote_cons = set(self._cons) - local_out_vars remote_objs = set(self._objs) - local_out_vars 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 sizes = model._var_sizes['nonlinear']['output'] for i, vname in enumerate(model._var_allprocs_abs_names['output']): owner = owning_ranks[vname] if vname in dv_set: dv_dict[vname] = (owner, sizes[owner, i]) if vname in con_set: con_dict[vname] = (owner, sizes[owner, i]) if vname in obj_set: obj_dict[vname] = (owner, sizes[owner, i]) 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) self._setup_simul_coloring() 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 _get_vars_to_record(self, recording_options): """ Get variables to record based on recording options. Parameters ---------- recording_options : <OptionsDictionary> Dictionary with recording options. Returns ------- dict Dictionary containing lists of variables to record. """ problem = self._problem model = problem.model if MPI: # TODO: Eventually, we think we can get rid of this next check. # But to be safe, we are leaving it in there. if not model.is_active(): raise RuntimeError( "RecordingManager.startup should never be called when " "running in parallel on an inactive System") rrank = problem.comm.rank rowned = model._owning_rank incl = recording_options['includes'] excl = recording_options['excludes'] # includes and excludes for outputs are specified using promoted names # NOTE: only local var names are in abs2prom, all will be gathered later abs2prom = model._var_abs2prom['output'] all_desvars = { n for n in self._designvars if n in abs2prom and check_path(abs2prom[n], incl, excl, True) } all_objectives = { n for n in self._objs if n in abs2prom and check_path(abs2prom[n], incl, excl, True) } all_constraints = { n for n in self._cons if n in abs2prom and check_path(abs2prom[n], incl, excl, True) } # design variables, objectives and constraints are always in the options mydesvars = myobjectives = myconstraints = set() if recording_options['record_desvars']: if MPI: mydesvars = [n for n in all_desvars if rrank == rowned[n]] else: mydesvars = list(all_desvars) if recording_options['record_objectives']: if MPI: myobjectives = [ n for n in all_objectives if rrank == rowned[n] ] else: myobjectives = list(all_objectives) if recording_options['record_constraints']: if MPI: myconstraints = [ n for n in all_constraints if rrank == rowned[n] ] else: myconstraints = list(all_constraints) filtered_vars_to_record = { 'des': mydesvars, 'obj': myobjectives, 'con': myconstraints } # responses (if in options) if 'record_responses' in recording_options: myresponses = set() if recording_options['record_responses']: myresponses = { n for n in self._responses if n in abs2prom and check_path(abs2prom[n], incl, excl, True) } if MPI: myresponses = [ n for n in myresponses if rrank == rowned[n] ] filtered_vars_to_record['res'] = list(myresponses) # inputs (if in options) if 'record_inputs' in recording_options: myinputs = set() if recording_options['record_inputs']: myinputs = { n for n in model._inputs if check_path(n, incl, excl) } if MPI: # gather the variables from all ranks to rank 0 all_vars = model.comm.gather(myinputs, root=0) if MPI.COMM_WORLD.rank == 0: myinputs = all_vars[-1] for d in all_vars[:-1]: myinputs.update(d) myinputs = [n for n in myinputs if rrank == rowned[n]] filtered_vars_to_record['in'] = list(myinputs) # system outputs myoutputs = set() if incl: myoutputs = { n for n in model._outputs if n in abs2prom and check_path(abs2prom[n], incl, excl) } if MPI: # gather the variables from all ranks to rank 0 all_vars = model.comm.gather(myoutputs, root=0) if MPI.COMM_WORLD.rank == 0: myoutputs = all_vars[-1] for d in all_vars[:-1]: myoutputs.update(d) # de-duplicate myoutputs = myoutputs.difference(all_desvars, all_objectives, all_constraints) if MPI: myoutputs = [n for n in myoutputs if rrank == rowned[n]] filtered_vars_to_record['sys'] = list(myoutputs) return filtered_vars_to_record def _setup_recording(self): """ Set up case recording. """ self._filtered_vars_to_record = self._get_vars_to_record( self.recording_options) self._rec_mgr.startup(self) # record the system metadata to the recorders attached to this Driver if self.recording_options['record_model_metadata']: for sub in self._problem.model.system_iter(recurse=True, include_self=True): self._rec_mgr.record_metadata(sub) def _get_voi_val(self, name, meta, remote_vois, driver_scaling=True, ignore_indices=False): """ 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. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- float or ndarray The value of the named variable of interest. """ model = self._problem.model comm = model.comm vec = model._outputs._views_flat indices = meta['indices'] if name in remote_vois: owner, size = remote_vois[name] if owner == comm.rank: if indices is None or ignore_indices: val = vec[name].copy() else: val = vec[name][indices] else: if not (indices is None or ignore_indices): size = len(indices) val = np.empty(size) comm.Bcast(val, root=owner) else: if name in self._designvars_discrete: val = model._discrete_outputs[name] # At present, only integers are supported by OpenMDAO drivers. # We check the values here. valid = True msg = "Only integer scalars or ndarrays are supported as values for " + \ "discrete variables when used as a design variable. " if np.isscalar(val) and not isinstance(val, int): msg += "A value of type '{}' was specified.".format( val.__class__.__name__) valid = False elif isinstance(val, np.ndarray) and not np.issubdtype(val[0], int): msg += "An array of type '{}' was specified.".format( val[0].__class__.__name__) valid = False if valid is False: raise ValueError(msg) elif indices is None or ignore_indices: val = vec[name].copy() else: val = vec[name][indices] if self._has_scaling and driver_scaling: # Scale design variable values adder = meta['adder'] if adder is not None: val += adder scaler = meta['scaler'] if scaler is not None: val *= scaler return val def get_design_var_values(self, filter=None, driver_scaling=True, ignore_indices=False): """ Return the design variable values. This is called to gather the initial design variable state. Parameters ---------- filter : list List of desvar names used by recorders. 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. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- dict Dictionary containing values of each design variable. """ if filter: dvs = filter else: # use all the designvars dvs = self._designvars return { n: self._get_voi_val(n, self._designvars[n], self._remote_dvs, driver_scaling=driver_scaling, ignore_indices=ignore_indices) for n in dvs } def set_design_var(self, name, value): """ Set the value of a design variable. Parameters ---------- name : str Global pathname of the design variable. value : float or ndarray Value for the design variable. """ problem = self._problem if (name in self._remote_dvs and problem.model._owning_rank[name] != problem.comm.rank): return meta = self._designvars[name] indices = meta['indices'] if indices is None: indices = slice(None) if name in self._designvars_discrete: problem.model._discrete_outputs[name] = int(value) else: desvar = problem.model._outputs._views_flat[name] desvar[indices] = value # Undo driver scaling when setting design var values into model. if self._has_scaling: scaler = meta['scaler'] if scaler is not None: desvar[indices] *= 1.0 / scaler adder = meta['adder'] if adder is not None: desvar[indices] -= adder def get_response_values(self, filter=None): """ Return response values. Parameters ---------- filter : list List of response names used by recorders. Returns ------- dict Dictionary containing values of each response. """ if filter: resps = filter else: resps = self._responses return { n: self._get_voi_val(n, self._responses[n], self._remote_objs) for n in resps } def get_objective_values(self, driver_scaling=True, filter=None, ignore_indices=False): """ 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. filter : list List of objective names used by recorders. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- dict Dictionary containing values of each objective. """ if filter: objs = filter else: objs = self._objs return { n: self._get_voi_val(n, self._objs[n], self._remote_objs, driver_scaling=driver_scaling, ignore_indices=ignore_indices) for n in objs } def get_constraint_values(self, ctype='all', lintype='all', driver_scaling=True, filter=None, ignore_indices=False): """ Return constraint values. Parameters ---------- ctype : string Default is 'all'. Optionally return just the inequality constraints with 'ineq' or the equality constraints with 'eq'. lintype : string 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. filter : list List of constraint names used by recorders. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- dict Dictionary containing values of each constraint. """ if filter is not None: cons = filter else: cons = self._cons con_dict = {} for name in cons: meta = self._cons[name] 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, ignore_indices=ignore_indices) 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 iteritems(self._cons) 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 = OrderedDict() self._cons = cons = OrderedDict() self._responses = resps = model.get_responses(recurse=True) for name, data in iteritems(resps): if data['type'] == 'con': cons[name] = data else: objs[name] = data response_size = sum(resps[n]['size'] for n in self._get_ordered_nl_responses()) # Gather up the information for design vars. self._designvars = designvars = model.get_design_vars(recurse=True) desvar_size = sum(data['size'] for data in itervalues(designvars)) return response_size, desvar_size def run(self): """ Execute this driver. The base `Driver` just runs the model. All other drivers overload this method. Returns ------- boolean 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 def _compute_totals(self, of=None, wrt=None, return_format='flat_dict', global_names=True): """ Compute derivatives of desired quantities with respect to desired inputs. All derivatives are returned using driver scaling. Parameters ---------- of : list of variable name strings or None Variables whose derivatives will be computed. Default is None, which uses the driver's objectives and constraints. wrt : list of variable name strings or None Variables with respect to which the derivatives will be computed. Default is None, which uses the driver's desvars. return_format : string 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. global_names : bool Set to True when passing in global names to skip some translation steps. Returns ------- derivs : object Derivatives in form requested by 'return_format'. """ problem = self._problem total_jac = self._total_jac debug_print = 'totals' in self.options['debug_print'] and ( not MPI or MPI.COMM_WORLD.rank == 0) if debug_print: header = 'Driver total derivatives for iteration: ' + str( self.iter_count) print(header) print(len(header) * '-' + '\n') if problem.model._owns_approx_jac: self._recording_iter.stack.append(('_compute_totals_approx', 0)) try: if total_jac is None: total_jac = _TotalJacInfo(problem, of, wrt, global_names, return_format, approx=True, debug_print=debug_print) # Don't cache linear constraint jacobian if not total_jac.has_lin_cons: self._total_jac = total_jac totals = total_jac.compute_totals_approx(initialize=True) else: totals = total_jac.compute_totals_approx() finally: self._recording_iter.stack.pop() else: if total_jac is None: total_jac = _TotalJacInfo(problem, of, wrt, global_names, return_format, debug_print=debug_print) # don't cache linear constraint jacobian if not total_jac.has_lin_cons: self._total_jac = total_jac self._recording_iter.stack.append(('_compute_totals', 0)) try: totals = total_jac.compute_totals() finally: self._recording_iter.stack.pop() if self._rec_mgr._recorders and self.recording_options[ 'record_derivatives']: metadata = create_local_meta(self._get_name()) total_jac.record_derivatives(self, metadata) return totals def record_iteration(self): """ Record an iteration of the current Driver. """ if not self._rec_mgr._recorders: return # Get the data to record (collective calls that get across all ranks) opts = self.recording_options filt = self._filtered_vars_to_record if opts['record_desvars']: des_vars = self.get_design_var_values(driver_scaling=False, ignore_indices=True) else: des_vars = {} if opts['record_objectives']: obj_vars = self.get_objective_values(driver_scaling=False, ignore_indices=True) else: obj_vars = {} if opts['record_constraints']: con_vars = self.get_constraint_values(driver_scaling=False, ignore_indices=True) else: con_vars = {} if opts['record_responses']: # res_vars = self.get_response_values() # not really working yet res_vars = {} else: res_vars = {} des_vars = {name: des_vars[name] for name in filt['des']} obj_vars = {name: obj_vars[name] for name in filt['obj']} con_vars = {name: con_vars[name] for name in filt['con']} # res_vars = {name: res_vars[name] for name in filt['res']} model = self._problem.model names = model._outputs._names views = model._outputs._views sys_vars = {name: views[name] for name in names if name in filt['sys']} if self.recording_options['record_inputs']: names = model._inputs._names views = model._inputs._views in_vars = { name: views[name] for name in names if name in filt['in'] } else: in_vars = {} if MPI: des_vars = self._gather_vars(model, des_vars) res_vars = self._gather_vars(model, res_vars) obj_vars = self._gather_vars(model, obj_vars) con_vars = self._gather_vars(model, con_vars) sys_vars = self._gather_vars(model, sys_vars) in_vars = self._gather_vars(model, in_vars) outs = {} if not MPI or model.comm.rank == 0: outs.update(des_vars) outs.update(res_vars) outs.update(obj_vars) outs.update(con_vars) outs.update(sys_vars) data = {'out': outs, 'in': in_vars} metadata = create_local_meta(self._get_name()) self._rec_mgr.record_iteration(self, data, metadata) def _gather_vars(self, root, local_vars): """ Gather and return only variables listed in `local_vars` from the `root` System. Parameters ---------- root : <System> the root System for the Problem local_vars : dict local variable names and values Returns ------- dct : dict variable names and values. """ # if trace: # debug("gathering vars for recording in %s" % root.pathname) all_vars = root.comm.gather(local_vars, root=0) # if trace: # debug("DONE gathering rec vars for %s" % root.pathname) if root.comm.rank == 0: dct = all_vars[-1] for d in all_vars[:-1]: dct.update(d) return dct def _get_name(self): """ Get name of current Driver. Returns ------- str Name of current Driver. """ return "Driver" def set_total_jac_sparsity(self, sparsity): """ Set the sparsity of sub-jacobians of the total jacobian. Note: This currently will have no effect if you are not using the pyOptSparseDriver. Parameters ---------- sparsity : str or dict :: # Sparsity is a nested dictionary where the outer keys are response # names, the inner keys are design variable names, and the value is a tuple of # the form (row_list, col_list, shape). { resp1: { dv1: (rows, cols, shape), # for sub-jac d_resp1/d_dv1 dv2: (rows, cols, shape), ... }, resp2: { ... } ... } """ if self.supports['total_jac_sparsity']: self._total_jac_sparsity = sparsity else: raise RuntimeError( "Driver '%s' does not support setting of total jacobian sparsity." % self._get_name()) 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'] ): """ 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. """ 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['coloring'] = None self._coloring_info['show_summary'] = show_summary self._coloring_info['show_sparsity'] = show_sparsity def use_fixed_coloring(self, coloring=coloring_mod._STD_COLORING_FNAME): """ Tell the driver to use a precomputed coloring. Parameters ---------- coloring : str A coloring filename. 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 set_simul_deriv_color(self, coloring): """ See use_fixed_coloring. This method is deprecated. Parameters ---------- coloring : str or Coloring Information about simultaneous coloring for design vars and responses. If a string, then coloring is assumed to be the name of a file that contains the coloring information in pickle format. Otherwise it must be a Coloring object. See the docstring for Coloring for details. """ warn_deprecation( "set_simul_deriv_color is deprecated. Use use_fixed_coloring instead." ) self.use_fixed_coloring(coloring) def _setup_tot_jac_sparsity(self): """ Set up total jacobian subjac sparsity. Drivers that can use subjac sparsity should override this. """ 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 """ 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 not None: return coloring if static is coloring_mod._STD_COLORING_FNAME or isinstance( static, string_types): 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) return coloring def _get_total_coloring_fname(self): return os.path.join(self._problem.options['coloring_dir'], 'total_coloring.pkl') def _setup_simul_coloring(self): """ Set up metadata for coloring of total derivative solution. If set_coloring was called with a filename, load the coloring file. """ # command line simul_coloring uses this env var to turn pre-existing coloring off if not coloring_mod._use_total_sparsity: return problem = self._problem if not problem.model._use_derivatives: simple_warning( "Derivatives are turned off. Skipping simul deriv coloring.") return total_coloring = self._get_static_coloring() if total_coloring._rev and problem._orig_mode not in ('rev', 'auto'): revcol = total_coloring._rev[0][0] if revcol: raise RuntimeError( "Simultaneous coloring does reverse solves but mode has " "been set to '%s'" % problem._orig_mode) if total_coloring._fwd and problem._orig_mode not in ('fwd', 'auto'): fwdcol = total_coloring._fwd[0][0] if fwdcol: raise RuntimeError( "Simultaneous coloring does forward solves but mode has " "been set to '%s'" % problem._orig_mode) def _pre_run_model_debug_print(self): """ Optionally print some debugging information before the model runs. """ debug_opt = self.options['debug_print'] if not debug_opt or debug_opt == ['totals']: return if not MPI or MPI.COMM_WORLD.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: desvar_vals = self.get_design_var_values(driver_scaling=False, ignore_indices=True) if not MPI or MPI.COMM_WORLD.rank == 0: print("Design Vars") if desvar_vals: # Print desvars in non_flattened state. meta = self._problem.model._var_allprocs_abs2meta for name in desvar_vals: shape = meta[name]['shape'] desvar_vals[name] = desvar_vals[name].reshape(shape) 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. """ if 'nl_cons' in self.options['debug_print']: cons = self.get_constraint_values(lintype='nonlinear', driver_scaling=False) if not MPI or MPI.COMM_WORLD.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 MPI.COMM_WORLD.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 MPI.COMM_WORLD.rank == 0: print("Objectives") if objs: pprint.pprint(objs) else: print("None") print() sys.stdout.flush()
class Solver(object): """ Base solver class. This class is subclassed by NonlinearSolver and LinearSolver, which are in turn subclassed by actual solver implementations. Attributes ---------- _system : <System> Pointer to the owning system. _depth : int How many subsolvers deep this solver is (0 means not a subsolver). _vec_names : [str, ...] List of right-hand-side (RHS) vector names. _mode : str 'fwd' or 'rev', applicable to linear solvers only. _iter_count : int Number of iterations for the current invocation of the solver. _rec_mgr : <RecordingManager> object that manages all recorders added to this solver cite : str Listing of relevant citations that should be referenced when publishing work that uses this class. options : <OptionsDictionary> Options dictionary. recording_options : <OptionsDictionary> Recording options dictionary. supports : <OptionsDictionary> Options dictionary describing what features are supported by this solver. _filtered_vars_to_record: Dict Dict of list of var names to record _norm0: float Normalization factor _solver_info : SolverInfo A stack-like object shared by all Solvers in the model. """ # Object to store some formatting for iprint that is shared across all solvers. SOLVER = 'base_solver' def __init__(self, **kwargs): """ Initialize all attributes. Parameters ---------- **kwargs : dict of keyword arguments Keyword arguments that will be mapped into the Solver options. """ self._system = None self._depth = 0 self._vec_names = None self._mode = 'fwd' self._iter_count = 0 self._solver_info = None # Solver options self.options = OptionsDictionary() self.options.declare('maxiter', types=int, default=10, desc='maximum number of iterations') self.options.declare('atol', default=1e-10, desc='absolute error tolerance') self.options.declare('rtol', default=1e-10, desc='relative error tolerance') self.options.declare('iprint', types=int, default=1, desc='whether to print output') self.options.declare('err_on_maxiter', types=bool, default=False, desc="When True, AnalysisError will be raised if we don't converge.") # Case recording options self.recording_options = OptionsDictionary() self.recording_options.declare('record_abs_error', types=bool, default=True, desc='Set to True to record absolute error at the \ solver level') self.recording_options.declare('record_rel_error', types=bool, default=True, desc='Set to True to record relative error at the \ solver level') self.recording_options.declare('record_inputs', types=bool, default=True, desc='Set to True to record inputs at the solver level') self.recording_options.declare('record_outputs', types=bool, default=True, desc='Set to True to record outputs at the solver level') self.recording_options.declare('record_solver_residuals', types=bool, default=False, desc='Set to True to record residuals at the solver level') self.recording_options.declare('record_metadata', types=bool, desc='Record metadata', default=True) self.recording_options.declare('includes', types=list, default=['*'], desc='Patterns for variables to include in recording') self.recording_options.declare('excludes', types=list, default=[], desc='Patterns for vars to exclude in recording ' '(processed post-includes)') # Case recording related self._filtered_vars_to_record = {} self._norm0 = 0.0 # What the solver supports. self.supports = OptionsDictionary() self.supports.declare('gradients', types=bool, default=False) self.supports.declare('implicit_components', types=bool, default=False) self._declare_options() self.options.update(kwargs) self._rec_mgr = RecordingManager() self.cite = "" def _assembled_jac_solver_iter(self): """ Return an empty generator of lin solvers using assembled jacs. """ for i in (): yield def add_recorder(self, recorder): """ Add a recorder to the solver's RecordingManager. Parameters ---------- recorder : <CaseRecorder> A recorder instance to be added to RecManager. """ if MPI: raise RuntimeError( "Recording of Solvers when running parallel code is not supported yet") self._rec_mgr.append(recorder) def _declare_options(self): """ Declare options before kwargs are processed in the init method. This is optionally implemented by subclasses of Solver. """ pass def _setup_solvers(self, system, depth): """ Assign system instance, set depth, and optionally perform setup. Parameters ---------- system : <System> pointer to the owning system. depth : int depth of the current system (already incremented). """ self._system = system self._depth = depth self._solver_info = system._solver_info self._recording_iter = system._recording_iter if isinstance(self, LinearSolver) and not system._use_derivatives: return self._rec_mgr.startup(self) self._rec_mgr.record_metadata(self) myoutputs = myresiduals = myinputs = set() incl = self.recording_options['includes'] excl = self.recording_options['excludes'] if self.recording_options['record_solver_residuals']: if isinstance(self, NonlinearSolver): residuals = system._residuals else: # it's a LinearSolver residuals = system._vectors['residual']['linear'] myresiduals = {n for n in residuals._names if check_path(n, incl, excl)} if self.recording_options['record_outputs']: if isinstance(self, NonlinearSolver): outputs = system._outputs else: # it's a LinearSolver outputs = system._vectors['output']['linear'] myoutputs = {n for n in outputs._names if check_path(n, incl, excl)} if self.recording_options['record_inputs']: if isinstance(self, NonlinearSolver): inputs = system._inputs else: inputs = system._vectors['input']['linear'] myinputs = {n for n in inputs._names if check_path(n, incl, excl)} self._filtered_vars_to_record = { 'in': myinputs, 'out': myoutputs, 'res': myresiduals } def _set_solver_print(self, level=2, type_='all'): """ Control printing for solvers and subsolvers in the model. Parameters ---------- level : int iprint level. Set to 2 to print residuals each iteration; set to 1 to print just the iteration totals; set to 0 to disable all printing except for failures, and set to -1 to disable all printing including failures. type_ : str Type of solver to set: 'LN' for linear, 'NL' for nonlinear, or 'all' for all. """ self.options['iprint'] = level def _mpi_print(self, iteration, abs_res, rel_res): """ Print residuals from an iteration. Parameters ---------- iteration : int iteration counter, 0-based. abs_res : float current absolute residual norm. rel_res : float current relative residual norm. """ if (self.options['iprint'] == 2 and self._system.comm.rank == 0): prefix = self._solver_info.prefix solver_name = self.SOLVER if prefix.endswith('precon:'): solver_name = solver_name[3:] print_str = prefix + solver_name print_str += ' %d ; %.9g %.9g' % (iteration, abs_res, rel_res) print(print_str) def _mpi_print_header(self): """ Print header text before solving. """ pass def _solve(self): """ Run the iterative solver. """ maxiter = self.options['maxiter'] atol = self.options['atol'] rtol = self.options['rtol'] iprint = self.options['iprint'] self._mpi_print_header() self._iter_count = 0 norm0, norm = self._iter_initialize() self._norm0 = norm0 self._mpi_print(self._iter_count, norm, norm / norm0) while self._iter_count < maxiter and norm > atol and norm / norm0 > rtol: with Recording(type(self).__name__, self._iter_count, self) as rec: self._single_iteration() self._iter_count += 1 self._run_apply() norm = self._iter_get_norm() # With solvers, we want to record the norm AFTER the call, but the call needs to # be wrapped in the with for stack purposes, so we locally assign norm & norm0 # into the class. rec.abs = norm rec.rel = norm / norm0 if norm0 == 0: norm0 = 1 self._mpi_print(self._iter_count, norm, norm / norm0) if self._system.comm.rank == 0 or os.environ.get('USE_PROC_FILES'): prefix = self._solver_info.prefix + self.SOLVER if np.isinf(norm) or np.isnan(norm) or (norm > atol and norm / norm0 > rtol): if iprint > -1: msg = ' Failed to Converge in {} iterations'.format(self._iter_count) print(prefix + msg) # Raise AnalysisError if requested. if self.options['err_on_maxiter']: msg = "Solver '{}' on system '{}' failed to converge." raise AnalysisError(msg.format(self.SOLVER, self._system.pathname)) elif iprint == 1: print(prefix + ' Converged in {} iterations'.format(self._iter_count)) elif iprint == 2: print(prefix + ' Converged') def _iter_initialize(self): """ Perform any necessary pre-processing operations. Returns ------- float initial error. float error at the first iteration. """ pass def _run_apply(self): """ Run the appropriate apply method on the system. """ pass def _linearize(self): """ Perform any required linearization operations such as matrix factorization. """ pass def _linearize_children(self): """ Return a flag that is True when we need to call linearize on our subsystems' solvers. Returns ------- boolean Flag for indicating child linerization """ return True def __str__(self): """ Return a string representation of the solver. Returns ------- str String representation of the solver. """ return self.SOLVER def record_iteration(self, **kwargs): """ Record an iteration of the current Solver. Parameters ---------- **kwargs : dict Keyword arguments (used for abs and rel error). """ if not self._rec_mgr._recorders: return metadata = create_local_meta(self.SOLVER) # Get the data data = {} if self.recording_options['record_abs_error']: data['abs'] = kwargs.get('abs') else: data['abs'] = None if self.recording_options['record_rel_error']: data['rel'] = kwargs.get('rel') else: data['rel'] = None system = self._system if isinstance(self, NonlinearSolver): outputs = system._outputs inputs = system._inputs residuals = system._residuals else: # it's a LinearSolver outputs = system._vectors['output']['linear'] inputs = system._vectors['input']['linear'] residuals = system._vectors['residual']['linear'] if self.recording_options['record_outputs']: data['o'] = {} if 'out' in self._filtered_vars_to_record: for out in self._filtered_vars_to_record['out']: if out in outputs._names: data['o'][out] = outputs._views[out] else: data['o'] = outputs else: data['o'] = None if self.recording_options['record_inputs']: data['i'] = {} if 'in' in self._filtered_vars_to_record: for inp in self._filtered_vars_to_record['in']: if inp in inputs._names: data['i'][inp] = inputs._views[inp] else: data['i'] = inputs else: data['i'] = None if self.recording_options['record_solver_residuals']: data['r'] = {} if 'res' in self._filtered_vars_to_record: for res in self._filtered_vars_to_record['res']: if res in residuals._names: data['r'][res] = residuals._views[res] else: data['r'] = residuals else: data['r'] = None self._rec_mgr.record_iteration(self, data, metadata) def cleanup(self): """ Clean up resources prior to exit. """ # shut down all recorders self._rec_mgr.shutdown() def _set_complex_step_mode(self, active): """ Turn on or off complex stepping mode. Recurses to turn on or off complex stepping mode in all subsystems and their vectors. Parameters ---------- active : bool Complex mode flag; set to True prior to commencing complex step. """ pass
class Driver(object): """ Top-level container for the systems and drivers. 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. debug_print : <OptionsDictionary> Dictionary with debugging printing options. cite : str Listing of relevant citataions that should be referenced when publishing work that uses this class. _problem : <Problem> Pointer to the containing problem. supports : <OptionsDictionary> Provides a consistant way for drivers to declare what features they support. _designvars : dict Contains all design variable info. _cons : dict Contains all constraint info. _objs : dict Contains all objective info. _responses : dict Contains all response info. _rec_mgr : <RecordingManager> Object that manages all recorders added to this driver. _vars_to_record: dict Dict of lists of var names indicating what to record _model_viewer_data : dict Structure of model, used to make n2 diagram. _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). _remote_responses : dict A combined dict containing entries from _remote_cons and _remote_objs. _simul_coloring_info : tuple of dicts A data structure describing coloring for simultaneous derivs. _total_jac_sparsity : dict, str, or None Specifies sparsity of sub-jacobians of the total jacobian. Only used by pyOptSparseDriver. _res_jacs : dict Dict of sparse subjacobians for use with certain optimizers, e.g. pyOptSparseDriver. _total_jac : _TotalJacInfo or None Cached total jacobian handling object. """ def __init__(self, **kwargs): """ Initialize the driver. Parameters ---------- **kwargs : dict of keyword arguments Keyword arguments that will be mapped into the Driver options. """ self._rec_mgr = RecordingManager() self._vars_to_record = { 'desvarnames': set(), 'responsenames': set(), 'objectivenames': set(), 'constraintnames': set(), 'sysinclnames': set(), } self._problem = None self._designvars = None self._cons = None self._objs = None self._responses = None # Driver options self.options = OptionsDictionary() self.options.declare( 'debug_print', types=list, is_valid=_is_debug_print_opts_valid, desc="List of what type of Driver variables to print at each " "iteration. Valid items in list are 'desvars', 'ln_cons', " "'nl_cons', 'objs', 'totals'", default=[]) # Case recording options self.recording_options = OptionsDictionary() self.recording_options.declare('record_metadata', types=bool, default=True, desc='Record Driver metadata') self.recording_options.declare( 'record_model_metadata', types=bool, default=True, desc='Record metadata for all Systems in the model') 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 to True to record responses 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') self.recording_options.declare( 'excludes', types=list, default=[], desc='Patterns for vars to exclude in recording ' '(processed post-includes)') 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_n2_data', types=bool, default=True, desc='Set to True to record metadata required for ' 'N^2 viewing') # What the driver supports. self.supports = OptionsDictionary() 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=False) 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) # TODO, support these in OpenMDAO self.supports.declare('integer_design_vars', types=bool, default=False) self.iter_count = 0 self._model_viewer_data = None self.cite = "" self._simul_coloring_info = None self._total_jac_sparsity = None self._res_jacs = {} self._total_jac = None self.fail = False self._declare_options() self.options.update(kwargs) def add_recorder(self, recorder): """ Add a recorder to the driver. Parameters ---------- recorder : BaseRecorder A recorder instance. """ self._rec_mgr.append(recorder) 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 _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. """ self._problem = problem model = problem.model mode = problem._mode self._total_jac = None self._has_scaling = (np.any([ r['scaler'] is not None for r in itervalues(self._responses) ]) or np.any( [dv['scaler'] is not None for dv in itervalues(self._designvars)])) con_set = set() obj_set = set() dv_set = set() self._remote_dvs = dv_dict = {} self._remote_cons = con_dict = {} self._remote_objs = obj_dict = {} # Now determine if later we'll need to allgather cons, objs, or desvars. if model.comm.size > 1 and model._subsystems_allprocs: local_out_vars = set(model._outputs._views) remote_dvs = set(self._designvars) - local_out_vars remote_cons = set(self._cons) - local_out_vars remote_objs = set(self._objs) - local_out_vars 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 sizes = model._var_sizes['nonlinear']['output'] for i, vname in enumerate(model._var_allprocs_abs_names['output']): owner = owning_ranks[vname] if vname in dv_set: dv_dict[vname] = (owner, sizes[owner, i]) if vname in con_set: con_dict[vname] = (owner, sizes[owner, i]) if vname in obj_set: obj_dict[vname] = (owner, sizes[owner, i]) self._remote_responses = self._remote_cons.copy() self._remote_responses.update(self._remote_objs) # set up case recording self._setup_recording() # set up simultaneous deriv coloring if (coloring_mod._use_sparsity and self._simul_coloring_info and self.supports['simultaneous_derivatives']): self._setup_simul_coloring() def _setup_recording(self): """ Set up case recording. """ problem = self._problem model = problem.model mydesvars = myobjectives = myconstraints = myresponses = set() myinputs = set() mysystem_outputs = set() incl = self.recording_options['includes'] excl = self.recording_options['excludes'] rec_desvars = self.recording_options['record_desvars'] rec_objectives = self.recording_options['record_objectives'] rec_constraints = self.recording_options['record_constraints'] rec_responses = self.recording_options['record_responses'] rec_inputs = self.recording_options['record_inputs'] all_desvars = { n for n in self._designvars if check_path(n, incl, excl, True) } all_objectives = { n for n in self._objs if check_path(n, incl, excl, True) } all_constraints = { n for n in self._cons if check_path(n, incl, excl, True) } if rec_desvars: mydesvars = all_desvars if rec_objectives: myobjectives = all_objectives if rec_constraints: myconstraints = all_constraints if rec_responses: myresponses = { n for n in self._responses if check_path(n, incl, excl, True) } # get the includes that were requested for this Driver recording if incl: # The my* variables are sets # First gather all of the desired outputs # The following might only be the local vars if MPI mysystem_outputs = { n for n in model._outputs if check_path(n, incl, excl) } # If MPI, and on rank 0, need to gather up all the variables # even those not local to rank 0 if MPI: all_vars = model.comm.gather(mysystem_outputs, root=0) if MPI.COMM_WORLD.rank == 0: mysystem_outputs = all_vars[-1] for d in all_vars[:-1]: mysystem_outputs.update(d) # de-duplicate mysystem_outputs mysystem_outputs = mysystem_outputs.difference( all_desvars, all_objectives, all_constraints) if rec_inputs: prob = self._problem root = prob.model myinputs = {n for n in root._inputs if check_path(n, incl, excl)} if MPI: all_vars = root.comm.gather(myinputs, root=0) if MPI.COMM_WORLD.rank == 0: myinputs = all_vars[-1] for d in all_vars[:-1]: myinputs.update(d) if MPI: # filter based on who owns the variables # TODO Eventually, we think we can get rid of this next check. But to be safe, # we are leaving it in there. if not model.is_active(): raise RuntimeError( "RecordingManager.startup should never be called when " "running in parallel on an inactive System") rrank = problem.comm.rank rowned = model._owning_rank mydesvars = [n for n in mydesvars if rrank == rowned[n]] myresponses = [n for n in myresponses if rrank == rowned[n]] myobjectives = [n for n in myobjectives if rrank == rowned[n]] myconstraints = [n for n in myconstraints if rrank == rowned[n]] mysystem_outputs = [ n for n in mysystem_outputs if rrank == rowned[n] ] myinputs = [n for n in myinputs if rrank == rowned[n]] self._filtered_vars_to_record = { 'des': mydesvars, 'obj': myobjectives, 'con': myconstraints, 'res': myresponses, 'sys': mysystem_outputs, 'in': myinputs } self._rec_mgr.startup(self) if self.recording_options['record_metadata']: if self.recording_options['record_n2_data']: if self._rec_mgr._recorders: from openmdao.devtools.problem_viewer.problem_viewer import _get_viewer_data self._model_viewer_data = _get_viewer_data(problem) self._rec_mgr.record_metadata(self) # Also record the system metadata to the recorders attached to this Driver if self.recording_options['record_model_metadata']: for sub in model.system_iter(recurse=True, include_self=True): self._rec_mgr.record_metadata(sub) def _get_voi_val(self, name, meta, remote_vois, unscaled=False, ignore_indices=False): """ 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). unscaled : bool Set to True if unscaled (physical) design variables are desired. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- float or ndarray The value of the named variable of interest. """ model = self._problem.model comm = model.comm vec = model._outputs._views_flat indices = meta['indices'] if name in remote_vois: owner, size = remote_vois[name] if owner == comm.rank: if indices is None or ignore_indices: val = vec[name].copy() else: val = vec[name][indices] else: if indices is not None: size = len(indices) val = np.empty(size) comm.Bcast(val, root=owner) else: if indices is None or ignore_indices: val = vec[name].copy() else: val = vec[name][indices] if self._has_scaling and not unscaled: # Scale design variable values adder = meta['adder'] if adder is not None: val += adder scaler = meta['scaler'] if scaler is not None: val *= scaler return val def get_design_var_values(self, filter=None, unscaled=False, ignore_indices=False): """ Return the design variable values. This is called to gather the initial design variable state. Parameters ---------- filter : list List of desvar names used by recorders. unscaled : bool Set to True if unscaled (physical) design variables are desired. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- dict Dictionary containing values of each design variable. """ if filter: dvs = filter else: # use all the designvars dvs = self._designvars return { n: self._get_voi_val(n, self._designvars[n], self._remote_dvs, unscaled=unscaled, ignore_indices=ignore_indices) for n in dvs } def set_design_var(self, name, value): """ Set the value of a design variable. Parameters ---------- name : str Global pathname of the design variable. value : float or ndarray Value for the design variable. """ if (name in self._remote_dvs and self._problem.model._owning_rank[name] != self._problem.comm.rank): return meta = self._designvars[name] indices = meta['indices'] if indices is None: indices = slice(None) desvar = self._problem.model._outputs._views_flat[name] desvar[indices] = value if self._has_scaling: # Scale design variable values scaler = meta['scaler'] if scaler is not None: desvar[indices] *= 1.0 / scaler adder = meta['adder'] if adder is not None: desvar[indices] -= adder def get_response_values(self, filter=None): """ Return response values. Parameters ---------- filter : list List of response names used by recorders. Returns ------- dict Dictionary containing values of each response. """ if filter: resps = filter else: resps = self._responses return { n: self._get_voi_val(n, self._responses[n], self._remote_objs) for n in resps } def get_objective_values(self, unscaled=False, filter=None): """ Return objective values. Parameters ---------- unscaled : bool Set to True if unscaled (physical) design variables are desired. filter : list List of objective names used by recorders. Returns ------- dict Dictionary containing values of each objective. """ if filter: objs = filter else: objs = self._objs return { n: self._get_voi_val(n, self._objs[n], self._remote_objs, unscaled=unscaled) for n in objs } def get_constraint_values(self, ctype='all', lintype='all', unscaled=False, filter=None): """ Return constraint values. Parameters ---------- ctype : string Default is 'all'. Optionally return just the inequality constraints with 'ineq' or the equality constraints with 'eq'. lintype : string Default is 'all'. Optionally return just the linear constraints with 'linear' or the nonlinear constraints with 'nonlinear'. unscaled : bool Set to True if unscaled (physical) design variables are desired. filter : list List of constraint names used by recorders. Returns ------- dict Dictionary containing values of each constraint. """ if filter is not None: cons = filter else: cons = self._cons con_dict = {} for name in cons: meta = self._cons[name] 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, unscaled=unscaled) 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 iteritems(self._cons) 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 = OrderedDict() self._cons = cons = OrderedDict() self._responses = resps = model.get_responses(recurse=True) for name, data in iteritems(resps): if data['type'] == 'con': cons[name] = data else: objs[name] = data response_size = np.sum(resps[n]['size'] for n in self._get_ordered_nl_responses()) # Gather up the information for design vars. self._designvars = designvars = model.get_design_vars(recurse=True) desvar_size = np.sum(data['size'] for data in itervalues(designvars)) return response_size, desvar_size def run(self): """ Execute this driver. The base `Driver` just runs the model. All other drivers overload this method. Returns ------- boolean Failure flag; True if failed to converge, False is successful. """ with RecordingDebugging(self._get_name(), self.iter_count, self) as rec: failure_flag, _, _ = self._problem.model._solve_nonlinear() self.iter_count += 1 return failure_flag def _compute_totals(self, of=None, wrt=None, return_format='flat_dict', global_names=True): """ Compute derivatives of desired quantities with respect to desired inputs. All derivatives are returned using driver scaling. Parameters ---------- of : list of variable name strings or None Variables whose derivatives will be computed. Default is None, which uses the driver's objectives and constraints. wrt : list of variable name strings or None Variables with respect to which the derivatives will be computed. Default is None, which uses the driver's desvars. return_format : string 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. global_names : bool Set to True when passing in global names to skip some translation steps. Returns ------- derivs : object Derivatives in form requested by 'return_format'. """ total_jac = self._total_jac debug_print = 'totals' in self.options['debug_print'] and ( not MPI or MPI.COMM_WORLD.rank == 0) if debug_print: header = 'Driver total derivatives for iteration: ' + str( self.iter_count) print(header) print(len(header) * '-' + '\n') if self._problem.model._owns_approx_jac: recording_iteration.stack.append(('_compute_totals_approx', 0)) try: if total_jac is None: total_jac = _TotalJacInfo(self._problem, of, wrt, global_names, return_format, approx=True, debug_print=debug_print) self._total_jac = total_jac totals = total_jac.compute_totals_approx(initialize=True) else: totals = total_jac.compute_totals_approx() finally: recording_iteration.stack.pop() else: if total_jac is None: total_jac = _TotalJacInfo(self._problem, of, wrt, global_names, return_format, debug_print=debug_print) # don't cache linear constraint jacobian if not total_jac.has_lin_cons: self._total_jac = total_jac recording_iteration.stack.append(('_compute_totals', 0)) try: totals = total_jac.compute_totals() finally: recording_iteration.stack.pop() if self._rec_mgr._recorders and self.recording_options[ 'record_derivatives']: metadata = create_local_meta(self._get_name()) total_jac.record_derivatives(self, metadata) return totals def record_iteration(self): """ Record an iteration of the current Driver. """ if not self._rec_mgr._recorders: return # Get the data to record (collective calls that get across all ranks) opts = self.recording_options filt = self._filtered_vars_to_record if opts['record_desvars']: des_vars = self.get_design_var_values(unscaled=True) else: des_vars = {} if opts['record_objectives']: obj_vars = self.get_objective_values(unscaled=True) else: obj_vars = {} if opts['record_constraints']: con_vars = self.get_constraint_values(unscaled=True) else: con_vars = {} if opts['record_responses']: # res_vars = self.get_response_values() # not really working yet res_vars = {} else: res_vars = {} des_vars = {name: des_vars[name] for name in filt['des']} obj_vars = {name: obj_vars[name] for name in filt['obj']} con_vars = {name: con_vars[name] for name in filt['con']} # res_vars = {name: res_vars[name] for name in filt['res']} model = self._problem.model names = model._outputs._names views = model._outputs._views sys_vars = {name: views[name] for name in names if name in filt['sys']} if self.recording_options['record_inputs']: names = model._inputs._names views = model._inputs._views in_vars = { name: views[name] for name in names if name in filt['in'] } else: in_vars = {} if MPI: des_vars = self._gather_vars(model, des_vars) res_vars = self._gather_vars(model, res_vars) obj_vars = self._gather_vars(model, obj_vars) con_vars = self._gather_vars(model, con_vars) sys_vars = self._gather_vars(model, sys_vars) in_vars = self._gather_vars(model, in_vars) outs = {} if not MPI or model.comm.rank == 0: outs.update(des_vars) outs.update(res_vars) outs.update(obj_vars) outs.update(con_vars) outs.update(sys_vars) data = {'out': outs, 'in': in_vars} metadata = create_local_meta(self._get_name()) self._rec_mgr.record_iteration(self, data, metadata) def _gather_vars(self, root, local_vars): """ Gather and return only variables listed in `local_vars` from the `root` System. Parameters ---------- root : <System> the root System for the Problem local_vars : dict local variable names and values Returns ------- dct : dict variable names and values. """ # if trace: # debug("gathering vars for recording in %s" % root.pathname) all_vars = root.comm.gather(local_vars, root=0) # if trace: # debug("DONE gathering rec vars for %s" % root.pathname) if root.comm.rank == 0: dct = all_vars[-1] for d in all_vars[:-1]: dct.update(d) return dct def _get_name(self): """ Get name of current Driver. Returns ------- str Name of current Driver. """ return "Driver" def set_simul_deriv_color(self, simul_info): """ Set the coloring (and possibly the sub-jac sparsity) for simultaneous total derivatives. Parameters ---------- simul_info : str or dict :: # Information about simultaneous coloring for design vars and responses. If a # string, then simul_info is assumed to be the name of a file that contains the # coloring information in JSON format. If a dict, the structure looks like this: { "fwd": [ # First, a list of column index lists, each index list representing columns # having the same color, except for the very first index list, which contains # indices of all columns that are not colored. [ [i1, i2, i3, ...] # list of non-colored columns [ia, ib, ...] # list of columns in first color [ic, id, ...] # list of columns in second color ... # remaining color lists, one list of columns per color ], # Next is a list of lists, one for each column, containing the nonzero rows for # that column. If a column is not colored, then it will have a None entry # instead of a list. [ [r1, rn, ...] # list of nonzero rows for column 0 None, # column 1 is not colored [ra, rb, ...] # list of nonzero rows for column 2 ... ], ], # This example is not a bidirectional coloring, so the opposite direction, "rev" # in this case, has an empty row index list. It could also be removed entirely. "rev": [[[]], []], "sparsity": # The sparsity entry can be absent, indicating that no sparsity structure is # specified, or it can be a nested dictionary where the outer keys are response # names, the inner keys are design variable names, and the value is a tuple of # the form (row_list, col_list, shape). { resp1_name: { dv1_name: (rows, cols, shape), # for sub-jac d_resp1/d_dv1 dv2_name: (rows, cols, shape), ... }, resp2_name: { ... } ... } } """ if self.supports['simultaneous_derivatives']: self._simul_coloring_info = simul_info else: raise RuntimeError( "Driver '%s' does not support simultaneous derivatives." % self._get_name()) def set_total_jac_sparsity(self, sparsity): """ Set the sparsity of sub-jacobians of the total jacobian. Note: This currently will have no effect if you are not using the pyOptSparseDriver. Parameters ---------- sparsity : str or dict :: # Sparsity is a nested dictionary where the outer keys are response # names, the inner keys are design variable names, and the value is a tuple of # the form (row_list, col_list, shape). { resp1: { dv1: (rows, cols, shape), # for sub-jac d_resp1/d_dv1 dv2: (rows, cols, shape), ... }, resp2: { ... } ... } """ if self.supports['total_jac_sparsity']: self._total_jac_sparsity = sparsity else: raise RuntimeError( "Driver '%s' does not support setting of total jacobian sparsity." % self._get_name()) def _setup_simul_coloring(self): """ Set up metadata for simultaneous derivative solution. """ # command line simul_coloring uses this env var to turn pre-existing coloring off if not coloring_mod._use_sparsity: return if isinstance(self._simul_coloring_info, string_types): with open(self._simul_coloring_info, 'r') as f: self._simul_coloring_info = json.load(f) if 'rev' in self._simul_coloring_info and self._problem._orig_mode not in ( 'rev', 'auto'): raise RuntimeError( "Simultaneous coloring does reverse solves but mode has " "been set to '%s'" % self._problem._orig_mode) # simul_coloring_info can contain data for either fwd, rev, or both, along with optional # sparsity patterns if 'sparsity' in self._simul_coloring_info: sparsity = self._simul_coloring_info['sparsity'] del self._simul_coloring_info['sparsity'] else: sparsity = None if sparsity is not None and self._total_jac_sparsity is not None: raise RuntimeError( "Total jac sparsity was set in both _simul_coloring_info" " and _total_jac_sparsity.") self._total_jac_sparsity = sparsity def _pre_run_model_debug_print(self): """ Optionally print some debugging information before the model runs. """ debug_opt = self.options['debug_print'] if not debug_opt or debug_opt == ['totals']: return if not MPI or MPI.COMM_WORLD.rank == 0: header = 'Driver debug print for iter coord: {}'.format( recording_iteration.get_formatted_iteration_coordinate()) print(header) print(len(header) * '-') if 'desvars' in debug_opt: desvar_vals = self.get_design_var_values(unscaled=True, ignore_indices=True) if not MPI or MPI.COMM_WORLD.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. """ if 'nl_cons' in self.options['debug_print']: cons = self.get_constraint_values(lintype='nonlinear', unscaled=True) if not MPI or MPI.COMM_WORLD.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', unscaled=True) if not MPI or MPI.COMM_WORLD.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(unscaled=True) if not MPI or MPI.COMM_WORLD.rank == 0: print("Objectives") if objs: pprint.pprint(objs) else: print("None") print() sys.stdout.flush()
class Driver(object): """ Top-level container for the systems and drivers. 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 : <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. _cons : dict Contains all constraint info. _objs : dict Contains all objective info. _responses : dict Contains all response info. _rec_mgr : <RecordingManager> Object that manages all recorders added to this driver. _vars_to_record: dict Dict of lists of var names indicating what to record _model_viewer_data : dict Structure of model, used to make n2 diagram. _simul_coloring_info : tuple of dicts A data structure describing coloring for simultaneous derivs. _total_jac_sparsity : dict, str, or None Specifies sparsity of sub-jacobians of the total jacobian. Only used by pyOptSparseDriver. _res_jacs : dict Dict of sparse subjacobians for use with certain optimizers, e.g. pyOptSparseDriver. _total_jac : _TotalJacInfo or None Cached total jacobian handling object. """ def __init__(self, **kwargs): """ Initialize the driver. Parameters ---------- **kwargs : dict of keyword arguments Keyword arguments that will be mapped into the Driver options. """ self._rec_mgr = RecordingManager() self._vars_to_record = { 'desvarnames': set(), 'responsenames': set(), 'objectivenames': set(), 'constraintnames': set(), 'sysinclnames': set(), } self._problem = None self._designvars = None self._cons = None self._objs = None self._responses = None # Driver options self.options = OptionsDictionary() self.options.declare('debug_print', types=list, check_valid=_check_debug_print_opts_valid, desc="List of what type of Driver variables to print at each " "iteration. Valid items in list are 'desvars', 'ln_cons', " "'nl_cons', 'objs', 'totals'", default=[]) # Case recording options self.recording_options = OptionsDictionary() self.recording_options.declare('record_model_metadata', types=bool, default=True, desc='Record metadata for all Systems in the model') 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 to True to record responses 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') self.recording_options.declare('excludes', types=list, default=[], desc='Patterns for vars to exclude in recording ' '(processed post-includes)') 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') # What the driver supports. self.supports = OptionsDictionary() 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=False) 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.iter_count = 0 self._model_viewer_data = None self.cite = "" self._simul_coloring_info = None self._total_jac_sparsity = None self._res_jacs = {} self._total_jac = None self.fail = False self._declare_options() self.options.update(kwargs) def add_recorder(self, recorder): """ Add a recorder to the driver. Parameters ---------- recorder : CaseRecorder A recorder instance. """ self._rec_mgr.append(recorder) 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 _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. """ self._problem = problem self._recording_iter = problem._recording_iter model = problem.model self._total_jac = None self._has_scaling = ( np.any([r['scaler'] is not None for r in itervalues(self._responses)]) or np.any([dv['scaler'] is not None for dv in itervalues(self._designvars)]) ) con_set = set() obj_set = set() dv_set = set() self._remote_dvs = dv_dict = {} self._remote_cons = con_dict = {} self._remote_objs = obj_dict = {} # Now determine if later we'll need to allgather cons, objs, or desvars. if model.comm.size > 1 and model._subsystems_allprocs: local_out_vars = set(model._outputs._views) remote_dvs = set(self._designvars) - local_out_vars remote_cons = set(self._cons) - local_out_vars remote_objs = set(self._objs) - local_out_vars 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 sizes = model._var_sizes['nonlinear']['output'] for i, vname in enumerate(model._var_allprocs_abs_names['output']): owner = owning_ranks[vname] if vname in dv_set: dv_dict[vname] = (owner, sizes[owner, i]) if vname in con_set: con_dict[vname] = (owner, sizes[owner, i]) if vname in obj_set: obj_dict[vname] = (owner, sizes[owner, i]) self._remote_responses = self._remote_cons.copy() self._remote_responses.update(self._remote_objs) # set up simultaneous deriv coloring if (coloring_mod._use_sparsity and self._simul_coloring_info and self.supports['simultaneous_derivatives']): self._setup_simul_coloring() def _get_vars_to_record(self, recording_options): """ Get variables to record based on recording options. Parameters ---------- recording_options : <OptionsDictionary> Dictionary with recording options. Returns ------- dict Dictionary containing lists of variables to record. """ problem = self._problem model = problem.model if MPI: # TODO: Eventually, we think we can get rid of this next check. # But to be safe, we are leaving it in there. if not model.is_active(): raise RuntimeError("RecordingManager.startup should never be called when " "running in parallel on an inactive System") rrank = problem.comm.rank rowned = model._owning_rank incl = recording_options['includes'] excl = recording_options['excludes'] # includes and excludes for outputs are specified using promoted names # NOTE: only local var names are in abs2prom, all will be gathered later abs2prom = model._var_abs2prom['output'] all_desvars = {n for n in self._designvars if n in abs2prom and check_path(abs2prom[n], incl, excl, True)} all_objectives = {n for n in self._objs if n in abs2prom and check_path(abs2prom[n], incl, excl, True)} all_constraints = {n for n in self._cons if n in abs2prom and check_path(abs2prom[n], incl, excl, True)} # design variables, objectives and constraints are always in the options mydesvars = myobjectives = myconstraints = set() if recording_options['record_desvars']: if MPI: mydesvars = [n for n in all_desvars if rrank == rowned[n]] else: mydesvars = list(all_desvars) if recording_options['record_objectives']: if MPI: myobjectives = [n for n in all_objectives if rrank == rowned[n]] else: myobjectives = list(all_objectives) if recording_options['record_constraints']: if MPI: myconstraints = [n for n in all_constraints if rrank == rowned[n]] else: myconstraints = list(all_constraints) filtered_vars_to_record = { 'des': mydesvars, 'obj': myobjectives, 'con': myconstraints } # responses (if in options) if 'record_responses' in recording_options: myresponses = set() if recording_options['record_responses']: myresponses = {n for n in self._responses if n in abs2prom and check_path(abs2prom[n], incl, excl, True)} if MPI: myresponses = [n for n in myresponses if rrank == rowned[n]] filtered_vars_to_record['res'] = list(myresponses) # inputs (if in options) if 'record_inputs' in recording_options: myinputs = set() if recording_options['record_inputs']: myinputs = {n for n in model._inputs if check_path(n, incl, excl)} if MPI: # gather the variables from all ranks to rank 0 all_vars = model.comm.gather(myinputs, root=0) if MPI.COMM_WORLD.rank == 0: myinputs = all_vars[-1] for d in all_vars[:-1]: myinputs.update(d) myinputs = [n for n in myinputs if rrank == rowned[n]] filtered_vars_to_record['in'] = list(myinputs) # system outputs (if the options being processed are for the driver itself) if recording_options is self.recording_options: myoutputs = set() if incl: myoutputs = {n for n in model._outputs if n in abs2prom and check_path(abs2prom[n], incl, excl)} if MPI: # gather the variables from all ranks to rank 0 all_vars = model.comm.gather(myoutputs, root=0) if MPI.COMM_WORLD.rank == 0: myoutputs = all_vars[-1] for d in all_vars[:-1]: myoutputs.update(d) # de-duplicate myoutputs = myoutputs.difference(all_desvars, all_objectives, all_constraints) if MPI: myoutputs = [n for n in myoutputs if rrank == rowned[n]] filtered_vars_to_record['sys'] = list(myoutputs) return filtered_vars_to_record def _setup_recording(self): """ Set up case recording. """ self._filtered_vars_to_record = self._get_vars_to_record(self.recording_options) self._rec_mgr.startup(self) # record the system metadata to the recorders attached to this Driver if self.recording_options['record_model_metadata']: for sub in self._problem.model.system_iter(recurse=True, include_self=True): self._rec_mgr.record_metadata(sub) def _get_voi_val(self, name, meta, remote_vois, unscaled=False, ignore_indices=False): """ 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). unscaled : bool Set to True if unscaled (physical) design variables are desired. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- float or ndarray The value of the named variable of interest. """ model = self._problem.model comm = model.comm vec = model._outputs._views_flat indices = meta['indices'] if name in remote_vois: owner, size = remote_vois[name] if owner == comm.rank: if indices is None or ignore_indices: val = vec[name].copy() else: val = vec[name][indices] else: if not (indices is None or ignore_indices): size = len(indices) val = np.empty(size) comm.Bcast(val, root=owner) else: if indices is None or ignore_indices: val = vec[name].copy() else: val = vec[name][indices] if self._has_scaling and not unscaled: # Scale design variable values adder = meta['adder'] if adder is not None: val += adder scaler = meta['scaler'] if scaler is not None: val *= scaler return val def get_design_var_values(self, filter=None, unscaled=False, ignore_indices=False): """ Return the design variable values. This is called to gather the initial design variable state. Parameters ---------- filter : list List of desvar names used by recorders. unscaled : bool Set to True if unscaled (physical) design variables are desired. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- dict Dictionary containing values of each design variable. """ if filter: dvs = filter else: # use all the designvars dvs = self._designvars return {n: self._get_voi_val(n, self._designvars[n], self._remote_dvs, unscaled=unscaled, ignore_indices=ignore_indices) for n in dvs} def set_design_var(self, name, value): """ Set the value of a design variable. Parameters ---------- name : str Global pathname of the design variable. value : float or ndarray Value for the design variable. """ problem = self._problem if (name in self._remote_dvs and problem.model._owning_rank[name] != problem.comm.rank): return meta = self._designvars[name] indices = meta['indices'] if indices is None: indices = slice(None) desvar = problem.model._outputs._views_flat[name] desvar[indices] = value if self._has_scaling: # Scale design variable values scaler = meta['scaler'] if scaler is not None: desvar[indices] *= 1.0 / scaler adder = meta['adder'] if adder is not None: desvar[indices] -= adder def get_response_values(self, filter=None): """ Return response values. Parameters ---------- filter : list List of response names used by recorders. Returns ------- dict Dictionary containing values of each response. """ if filter: resps = filter else: resps = self._responses return {n: self._get_voi_val(n, self._responses[n], self._remote_objs) for n in resps} def get_objective_values(self, unscaled=False, filter=None, ignore_indices=False): """ Return objective values. Parameters ---------- unscaled : bool Set to True if unscaled (physical) design variables are desired. filter : list List of objective names used by recorders. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- dict Dictionary containing values of each objective. """ if filter: objs = filter else: objs = self._objs return {n: self._get_voi_val(n, self._objs[n], self._remote_objs, unscaled=unscaled, ignore_indices=ignore_indices) for n in objs} def get_constraint_values(self, ctype='all', lintype='all', unscaled=False, filter=None, ignore_indices=False): """ Return constraint values. Parameters ---------- ctype : string Default is 'all'. Optionally return just the inequality constraints with 'ineq' or the equality constraints with 'eq'. lintype : string Default is 'all'. Optionally return just the linear constraints with 'linear' or the nonlinear constraints with 'nonlinear'. unscaled : bool Set to True if unscaled (physical) design variables are desired. filter : list List of constraint names used by recorders. ignore_indices : bool Set to True if the full array is desired, not just those indicated by indices. Returns ------- dict Dictionary containing values of each constraint. """ if filter is not None: cons = filter else: cons = self._cons con_dict = {} for name in cons: meta = self._cons[name] 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, unscaled=unscaled, ignore_indices=ignore_indices) 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 iteritems(self._cons) 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 = OrderedDict() self._cons = cons = OrderedDict() self._responses = resps = model.get_responses(recurse=True) for name, data in iteritems(resps): if data['type'] == 'con': cons[name] = data else: objs[name] = data response_size = sum(resps[n]['size'] for n in self._get_ordered_nl_responses()) # Gather up the information for design vars. self._designvars = designvars = model.get_design_vars(recurse=True) desvar_size = sum(data['size'] for data in itervalues(designvars)) return response_size, desvar_size def run(self): """ Execute this driver. The base `Driver` just runs the model. All other drivers overload this method. Returns ------- boolean Failure flag; True if failed to converge, False is successful. """ with RecordingDebugging(self._get_name(), self.iter_count, self): self._problem.model._solve_nonlinear() self.iter_count += 1 return False def _compute_totals(self, of=None, wrt=None, return_format='flat_dict', global_names=True): """ Compute derivatives of desired quantities with respect to desired inputs. All derivatives are returned using driver scaling. Parameters ---------- of : list of variable name strings or None Variables whose derivatives will be computed. Default is None, which uses the driver's objectives and constraints. wrt : list of variable name strings or None Variables with respect to which the derivatives will be computed. Default is None, which uses the driver's desvars. return_format : string 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. global_names : bool Set to True when passing in global names to skip some translation steps. Returns ------- derivs : object Derivatives in form requested by 'return_format'. """ problem = self._problem total_jac = self._total_jac debug_print = 'totals' in self.options['debug_print'] and (not MPI or MPI.COMM_WORLD.rank == 0) if debug_print: header = 'Driver total derivatives for iteration: ' + str(self.iter_count) print(header) print(len(header) * '-' + '\n') if problem.model._owns_approx_jac: self._recording_iter.stack.append(('_compute_totals_approx', 0)) try: if total_jac is None: total_jac = _TotalJacInfo(problem, of, wrt, global_names, return_format, approx=True, debug_print=debug_print) self._total_jac = total_jac totals = total_jac.compute_totals_approx(initialize=True) else: totals = total_jac.compute_totals_approx() finally: self._recording_iter.stack.pop() else: if total_jac is None: total_jac = _TotalJacInfo(problem, of, wrt, global_names, return_format, debug_print=debug_print) # don't cache linear constraint jacobian if not total_jac.has_lin_cons: self._total_jac = total_jac self._recording_iter.stack.append(('_compute_totals', 0)) try: totals = total_jac.compute_totals() finally: self._recording_iter.stack.pop() if self._rec_mgr._recorders and self.recording_options['record_derivatives']: metadata = create_local_meta(self._get_name()) total_jac.record_derivatives(self, metadata) return totals def record_iteration(self): """ Record an iteration of the current Driver. """ if not self._rec_mgr._recorders: return # Get the data to record (collective calls that get across all ranks) opts = self.recording_options filt = self._filtered_vars_to_record if opts['record_desvars']: des_vars = self.get_design_var_values(unscaled=True, ignore_indices=True) else: des_vars = {} if opts['record_objectives']: obj_vars = self.get_objective_values(unscaled=True, ignore_indices=True) else: obj_vars = {} if opts['record_constraints']: con_vars = self.get_constraint_values(unscaled=True, ignore_indices=True) else: con_vars = {} if opts['record_responses']: # res_vars = self.get_response_values() # not really working yet res_vars = {} else: res_vars = {} des_vars = {name: des_vars[name] for name in filt['des']} obj_vars = {name: obj_vars[name] for name in filt['obj']} con_vars = {name: con_vars[name] for name in filt['con']} # res_vars = {name: res_vars[name] for name in filt['res']} model = self._problem.model names = model._outputs._names views = model._outputs._views sys_vars = {name: views[name] for name in names if name in filt['sys']} if self.recording_options['record_inputs']: names = model._inputs._names views = model._inputs._views in_vars = {name: views[name] for name in names if name in filt['in']} else: in_vars = {} if MPI: des_vars = self._gather_vars(model, des_vars) res_vars = self._gather_vars(model, res_vars) obj_vars = self._gather_vars(model, obj_vars) con_vars = self._gather_vars(model, con_vars) sys_vars = self._gather_vars(model, sys_vars) in_vars = self._gather_vars(model, in_vars) outs = {} if not MPI or model.comm.rank == 0: outs.update(des_vars) outs.update(res_vars) outs.update(obj_vars) outs.update(con_vars) outs.update(sys_vars) data = { 'out': outs, 'in': in_vars } metadata = create_local_meta(self._get_name()) self._rec_mgr.record_iteration(self, data, metadata) def _gather_vars(self, root, local_vars): """ Gather and return only variables listed in `local_vars` from the `root` System. Parameters ---------- root : <System> the root System for the Problem local_vars : dict local variable names and values Returns ------- dct : dict variable names and values. """ # if trace: # debug("gathering vars for recording in %s" % root.pathname) all_vars = root.comm.gather(local_vars, root=0) # if trace: # debug("DONE gathering rec vars for %s" % root.pathname) if root.comm.rank == 0: dct = all_vars[-1] for d in all_vars[:-1]: dct.update(d) return dct def _get_name(self): """ Get name of current Driver. Returns ------- str Name of current Driver. """ return "Driver" def set_simul_deriv_color(self, simul_info): """ Set the coloring (and possibly the sub-jac sparsity) for simultaneous total derivatives. Parameters ---------- simul_info : str or dict :: # Information about simultaneous coloring for design vars and responses. If a # string, then simul_info is assumed to be the name of a file that contains the # coloring information in JSON format. If a dict, the structure looks like this: { "fwd": [ # First, a list of column index lists, each index list representing columns # having the same color, except for the very first index list, which contains # indices of all columns that are not colored. [ [i1, i2, i3, ...] # list of non-colored columns [ia, ib, ...] # list of columns in first color [ic, id, ...] # list of columns in second color ... # remaining color lists, one list of columns per color ], # Next is a list of lists, one for each column, containing the nonzero rows for # that column. If a column is not colored, then it will have a None entry # instead of a list. [ [r1, rn, ...] # list of nonzero rows for column 0 None, # column 1 is not colored [ra, rb, ...] # list of nonzero rows for column 2 ... ], ], # This example is not a bidirectional coloring, so the opposite direction, "rev" # in this case, has an empty row index list. It could also be removed entirely. "rev": [[[]], []], "sparsity": # The sparsity entry can be absent, indicating that no sparsity structure is # specified, or it can be a nested dictionary where the outer keys are response # names, the inner keys are design variable names, and the value is a tuple of # the form (row_list, col_list, shape). { resp1_name: { dv1_name: (rows, cols, shape), # for sub-jac d_resp1/d_dv1 dv2_name: (rows, cols, shape), ... }, resp2_name: { ... } ... } } """ if self.supports['simultaneous_derivatives']: self._simul_coloring_info = simul_info else: raise RuntimeError("Driver '%s' does not support simultaneous derivatives." % self._get_name()) def set_total_jac_sparsity(self, sparsity): """ Set the sparsity of sub-jacobians of the total jacobian. Note: This currently will have no effect if you are not using the pyOptSparseDriver. Parameters ---------- sparsity : str or dict :: # Sparsity is a nested dictionary where the outer keys are response # names, the inner keys are design variable names, and the value is a tuple of # the form (row_list, col_list, shape). { resp1: { dv1: (rows, cols, shape), # for sub-jac d_resp1/d_dv1 dv2: (rows, cols, shape), ... }, resp2: { ... } ... } """ if self.supports['total_jac_sparsity']: self._total_jac_sparsity = sparsity else: raise RuntimeError("Driver '%s' does not support setting of total jacobian sparsity." % self._get_name()) def _setup_simul_coloring(self): """ Set up metadata for simultaneous derivative solution. """ # command line simul_coloring uses this env var to turn pre-existing coloring off if not coloring_mod._use_sparsity: return problem = self._problem if not problem.model._use_derivatives: simple_warning("Derivatives are turned off. Skipping simul deriv coloring.") return if isinstance(self._simul_coloring_info, string_types): with open(self._simul_coloring_info, 'r') as f: self._simul_coloring_info = coloring_mod._json2coloring(json.load(f)) if 'rev' in self._simul_coloring_info and problem._orig_mode not in ('rev', 'auto'): revcol = self._simul_coloring_info['rev'][0][0] if revcol: raise RuntimeError("Simultaneous coloring does reverse solves but mode has " "been set to '%s'" % problem._orig_mode) if 'fwd' in self._simul_coloring_info and problem._orig_mode not in ('fwd', 'auto'): fwdcol = self._simul_coloring_info['fwd'][0][0] if fwdcol: raise RuntimeError("Simultaneous coloring does forward solves but mode has " "been set to '%s'" % problem._orig_mode) # simul_coloring_info can contain data for either fwd, rev, or both, along with optional # sparsity patterns if 'sparsity' in self._simul_coloring_info: sparsity = self._simul_coloring_info['sparsity'] del self._simul_coloring_info['sparsity'] else: sparsity = None if sparsity is not None and self._total_jac_sparsity is not None: raise RuntimeError("Total jac sparsity was set in both _simul_coloring_info" " and _total_jac_sparsity.") self._total_jac_sparsity = sparsity def _pre_run_model_debug_print(self): """ Optionally print some debugging information before the model runs. """ debug_opt = self.options['debug_print'] if not debug_opt or debug_opt == ['totals']: return if not MPI or MPI.COMM_WORLD.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: desvar_vals = self.get_design_var_values(unscaled=True, ignore_indices=True) if not MPI or MPI.COMM_WORLD.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. """ if 'nl_cons' in self.options['debug_print']: cons = self.get_constraint_values(lintype='nonlinear', unscaled=True) if not MPI or MPI.COMM_WORLD.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', unscaled=True) if not MPI or MPI.COMM_WORLD.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(unscaled=True) if not MPI or MPI.COMM_WORLD.rank == 0: print("Objectives") if objs: pprint.pprint(objs) else: print("None") print() sys.stdout.flush()