def calc_derivatives(self, first=False, second=False, savebase=True, required_inputs=None, required_outputs=None):
"""Calculate the derivatives for this non-differentiable block using
Finite Difference."""
# We don't do this in __init__ because some inputs and outputs
# are added after creation (for nested driver support).
if self.fd is None:
from openmdao.main.derivatives import FiniteDifference
self.fd = FiniteDifference(self)
if hasattr(self.wflow, "_severed_edges"):
self.wflow.sever_edges(self.wflow._severed_edges)
try:
# First, linearize about operating point.
# Note: Only needed for differentiable islands, which are handled
# with Fake Finite Difference.
# Don't do this for full-model finite difference.
if first and self.ffd_order > 0:
for name in self.comps:
comp = self.wflow.scope.get(name)
# Comp list contains full graph, so don't double up on
# the subdrivers.
if not has_interface(comp, IDriver) and name not in self.wflow._J_cache:
comp.calc_derivatives(first, second, True)
self.J = self.fd.calculate()
finally:
if hasattr(self.wflow, "_severed_edges"):
self.wflow.unsever_edges()
return self.J
class PseudoAssembly(object):
"""The PseudoAssembly is used to aggregate blocks of components that cannot
provide derivatives, and thus must be finite differenced. It is not a real
assembly, and should never be used in an OpenMDAO model."""
def __init__(self, name, comps, graph, wflow, fd=False, drv_name=None, boundary_params=()):
"""Initialized with list of components, and the parent workflow."""
inputs, outputs, renames = self._pre_init(name, comps, graph, fd, boundary_params)
self.comps = wflow.scope._depgraph.order_components(set(comps))
self.name = name
self.boundary_params = list(boundary_params)
self.wflow = wflow
self.inputs = list(inputs)
self.outputs = list(outputs)
self.renames = renames
self.mapped_inputs = []
self._removed_comps = []
self._provideJ_bounds = None
for varpath in self.inputs:
if isinstance(varpath, basestring):
val = to_PA_var(varpath, name).partition(".")[2]
else:
val = tuple([to_PA_var(vp, name).partition(".")[2] for vp in varpath])
self.mapped_inputs.append(val)
self.mapped_outputs = [to_PA_var(varpath, name).partition(".")[2] for varpath in self.outputs]
self.itername = ""
self.fd = None
self.J = None
if fd: # for full-model fd, turn off fake finite difference
self.ffd_order = 0
else: # use fake finite difference on comps having derivatives
self.ffd_order = 1
if fd:
self.itercomps = [c.name for c in wflow]
elif drv_name is not None:
self.itercomps = [drv_name]
else:
self.itercomps = list(self.comps)
def _pre_init(self, pa_name, group, dgraph, fd, boundary_params):
"""Return a tuple of the form (pa_inputs, pa_outputs, renames)
for the PseudoAssembly that would be created given the nodes in
group and the given graph.
"""
# First, find our group boundary
self._orig_group_nodes = list(group) + list(boundary_params)
allnodes = dgraph.find_prefixed_nodes(self._orig_group_nodes)
out_edges = nx.edge_boundary(dgraph, allnodes)
in_edges = nx.edge_boundary(dgraph, set(dgraph.nodes()).difference(allnodes))
solver_states = []
if fd is False:
for comp in group:
solver_states.extend(
[node for node in dgraph.predecessors(comp) if "solver_state" in dgraph.node[node]]
)
pa_inputs = edges_to_dict(in_edges).values()
pa_inputs.extend(solver_states)
pa_outputs = set([a for a, b in out_edges])
renames = {}
# Add pseudoassy inputs
for varpath in list(flatten_list_of_iters(pa_inputs)) + list(pa_outputs):
varname = to_PA_var(varpath, pa_name)
if varpath in dgraph:
renames[varpath] = varname
old = dgraph.base_var(varpath)
if old != varpath:
renames[old] = to_PA_var(old, pa_name)
# make boundary params outputs of the PA
pa_outputs.update(boundary_params)
return pa_inputs, pa_outputs, renames
def set_itername(self, name):
"""Comp API compatibility; allows iteration coord to be set in
components."""
self.itername = name
def run(self, ffd_order=0, case_id=""):
"""Run all components contained in this assy. Used by finite
difference."""
# Override fake finite difference if requested. This enables a pure
# finite-differences for check_derivatives.
if self.ffd_order == 0:
ffd_order = 0
for name in self.itercomps:
comp = self.wflow.scope.get(name)
comp.set_itername(self.itername + "-fd")
comp.run(ffd_order=ffd_order, case_id=case_id)
def calc_derivatives(self, first=False, second=False, savebase=True, required_inputs=None, required_outputs=None):
"""Calculate the derivatives for this non-differentiable block using
Finite Difference."""
# We don't do this in __init__ because some inputs and outputs
# are added after creation (for nested driver support).
if self.fd is None:
from openmdao.main.derivatives import FiniteDifference
self.fd = FiniteDifference(self)
if hasattr(self.wflow, "_severed_edges"):
self.wflow.sever_edges(self.wflow._severed_edges)
try:
# First, linearize about operating point.
# Note: Only needed for differentiable islands, which are handled
# with Fake Finite Difference.
# Don't do this for full-model finite difference.
if first and self.ffd_order > 0:
for name in self.comps:
comp = self.wflow.scope.get(name)
# Comp list contains full graph, so don't double up on
# the subdrivers.
if not has_interface(comp, IDriver) and name not in self.wflow._J_cache:
comp.calc_derivatives(first, second, True)
self.J = self.fd.calculate()
finally:
if hasattr(self.wflow, "_severed_edges"):
self.wflow.unsever_edges()
return self.J
def provideJ(self):
"""Jacobian for this block"""
return self.J
def list_deriv_vars(self):
"""Derivative inputs and outputs for this block"""
return self.mapped_inputs, self.mapped_outputs
def get(self, varname):
""" Return the value of a variable in the Pseudoassembly. Used
when sizing variables in the Jacobian."""
return self.wflow.scope.get(from_PA_var(self.name + "." + varname))
def add_to_graph(self, startgraph, dgraph, excludes=()):
"""Add this PseudoAssembly to the given graph, absorbing
nodes that represent components contained in this PA.
"""
from openmdao.main.depgraph import is_subvar_node, is_input_base_node, is_output_base_node
self._removed_comps = [c for c in self.comps if c not in excludes]
# Add pseudoassys to graph
dgraph.add_node(self.name, pa_object=self, comp=True, pseudo="assembly", valid=True)
empty = {}
for oldname, newname in self.renames.items():
attrs = startgraph.node.get(oldname, empty).copy()
if oldname in self.boundary_params:
attrs["iotype"] = "out"
dgraph.add_node(newname, attr_dict=attrs)
for u, v, data in dgraph.edges(oldname, data=True):
dgraph.add_edge(newname, v, attr_dict=data.copy())
for u, v, data in dgraph.in_edges(oldname, data=True):
dgraph.add_edge(u, newname, attr_dict=data.copy())
if is_subvar_node(dgraph, newname):
dgraph.node[newname]["basevar"] = newname.split("[", 1)[0]
if is_input_base_node(dgraph, newname):
dgraph.add_edge(newname, self.name)
elif is_output_base_node(dgraph, newname):
dgraph.add_edge(self.name, newname)
dgraph.config_changed()
def clean_graph(self, startgraph, dgraph, excludes=()):
""" Clean up the old nodes in the graph, leaving out ones that
are excluded because they're used by ancestor drivers.
"""
dgraph.remove_nodes_from(
dgraph.find_prefixed_nodes([n for n in self._orig_group_nodes if n not in excludes and n in dgraph])
)
# if this PA represents a driver, remove the corresponding driver
# node
if self.name[1:] in dgraph:
dgraph.remove_nodes_from(dgraph.find_prefixed_nodes([self.name[1:]]))
# update mapped inputs/outputs in the graph
map_inputs = dgraph.graph["mapped_inputs"]
map_outputs = dgraph.graph["mapped_outputs"]
for i, varpath in enumerate(map_inputs):
if isinstance(varpath, basestring):
varpath = [varpath]
mapped = []
for path in varpath:
compname, _, varname = path.partition(".")
if varname and (compname in self.comps):
mapped.append(to_PA_var(path, self.name))
else:
mapped.append(path) # keep old value in that spot
map_inputs[i] = tuple(mapped)
# Add requested outputs
for i, varpath in enumerate(map_outputs):
compname, _, varname = varpath.partition(".")
if varname and (compname in self.comps):
map_outputs[i] = to_PA_var(varpath, self.name)
dgraph.config_changed()
def calc_derivatives(self,
first=False,
second=False,
savebase=True,
required_inputs=None,
required_outputs=None):
"""Calculate the derivatives for this non-differentiable block using
Finite Difference."""
# We don't do this in __init__ because some inputs and outputs
# are added after creation (for nested driver support).
if self.fd is None:
from openmdao.main.derivatives import FiniteDifference
self.fd = FiniteDifference(self)
if hasattr(self.wflow, '_severed_edges'):
self.wflow.sever_edges(self.wflow._severed_edges)
try:
# First, linearize about operating point.
# Note: Only needed for differentiable islands, which are handled
# with Fake Finite Difference.
# Don't do this for full-model finite difference.
if first and self.ffd_order > 0:
for name in self.comps:
# TODO: I think the cache is blown away each time before
# this is called
if name in self.wflow._J_cache:
continue
comp = self.wflow.scope.get(name)
# Assemblies need some required inputs and outputs
# to calculate the Jacobians
if has_interface(comp, IAssembly):
# Need to know which assy bdry variables are
# required, and pass them in. Cache this once.
if name not in self.ffd_cache:
dgraph = self.wflow.scope._depgraph
inputs = [
dgraph.base_var(inp)
for inp in flatten_list_of_iters(self.inputs)
]
outputs = [
dgraph.base_var(outp) for outp in self.outputs
]
from openmdao.main.depgraph import _get_inner_edges
edges = _get_inner_edges(dgraph, inputs, outputs)
req_inputs = []
req_outputs = []
for inp in inputs:
comp_str, _, var_str = inp.partition('.')
if comp_str == name:
req_inputs.append(var_str)
for inp in outputs:
comp_str, _, var_str = inp.partition('.')
if comp_str == name:
req_outputs.append(var_str)
for edge in edges:
src, target = edge
comp_str, _, var_str = src.partition('.')
if comp_str == name:
req_outputs.append(var_str)
comp_str, _, var_str = target.partition('.')
if comp_str == name:
req_inputs.append(var_str)
self.ffd_cache[name] = (req_inputs, req_outputs)
req_inputs, req_outputs = self.ffd_cache[name]
comp.calc_derivatives(first,
second,
savebase=True,
required_inputs=req_inputs,
required_outputs=req_outputs)
# Comp list contains full graph, so don't double up on
# the subdrivers.
elif not has_interface(comp, IDriver):
comp.calc_derivatives(first, second, True)
self.J = self.fd.calculate()
finally:
if hasattr(self.wflow, '_severed_edges'):
self.wflow.unsever_edges()
return self.J
class PseudoAssembly(object):
"""The PseudoAssembly is used to aggregate blocks of components that cannot
provide derivatives, and thus must be finite differenced. It is not a real
assembly, and should never be used in an OpenMDAO model."""
def __init__(self,
name,
comps,
graph,
wflow,
fd=False,
drv_name=None,
boundary_params=()):
"""Initialized with list of components, and the parent workflow."""
inputs, outputs, renames = self._pre_init(name, comps, graph, fd,
boundary_params)
self.comps = wflow.scope._depgraph.order_components(set(comps))
self.name = name
self.boundary_params = list(boundary_params)
self.wflow = wflow
self.inputs = list(inputs)
self.outputs = list(outputs)
self.renames = renames
self.mapped_inputs = []
self._removed_comps = []
self._provideJ_bounds = None
for varpath in self.inputs:
if isinstance(varpath, basestring):
val = to_PA_var(varpath, name).partition('.')[2]
else:
val = tuple(
[to_PA_var(vp, name).partition('.')[2] for vp in varpath])
self.mapped_inputs.append(val)
self.mapped_outputs = [
to_PA_var(varpath, name).partition('.')[2]
for varpath in self.outputs
]
self.itername = ''
self.fd = None
self.J = None
self.ffd_cache = {}
if fd: # for full-model fd, turn off fake finite difference
self.ffd_order = 0
else: # use fake finite difference on comps having derivatives
# TODO: Fake Finite Difference has been disabled until we upgrade
# it to support arrays and vartrees.
self.ffd_order = 0
if fd:
self.itercomps = [c.name for c in wflow]
elif drv_name is not None:
self.itercomps = [drv_name]
else:
self.itercomps = list(self.comps)
#print "Created PseudoAssembly:", self.comps
def _pre_init(self, pa_name, group, dgraph, fd, boundary_params):
"""Return a tuple of the form (pa_inputs, pa_outputs, renames)
for the PseudoAssembly that would be created given the nodes in
group and the given graph.
"""
# First, find our group boundary
self._orig_group_nodes = list(group) + list(boundary_params)
allnodes = dgraph.find_prefixed_nodes(self._orig_group_nodes)
out_edges = nx.edge_boundary(dgraph, allnodes)
in_edges = nx.edge_boundary(dgraph,
set(dgraph.nodes()).difference(allnodes))
solver_states = []
if fd is False:
for comp in group:
solver_states.extend([
node for node in dgraph.predecessors(comp)
if 'solver_state' in dgraph.node[node]
])
pa_inputs = edges_to_dict(in_edges).values()
pa_inputs.extend(solver_states)
pa_outputs = set([a for a, b in out_edges])
renames = {}
# Add pseudoassy inputs
for varpath in list(flatten_list_of_iters(pa_inputs)) + \
list(pa_outputs):
varname = to_PA_var(varpath, pa_name)
if varpath in dgraph:
renames[varpath] = varname
old = dgraph.base_var(varpath)
if old != varpath:
renames[old] = to_PA_var(old, pa_name)
# make boundary params outputs of the PA
pa_outputs.update(boundary_params)
return pa_inputs, pa_outputs, renames
def set_itername(self, name):
"""Comp API compatibility; allows iteration coord to be set in
components."""
self.itername = name
def run(self, ffd_order=0, case_id=''):
"""Run all components contained in this assy. Used by finite
difference."""
# Override fake finite difference if requested. This enables a pure
# finite-differences for check_derivatives.
if self.ffd_order == 0:
ffd_order = 0
for name in self.itercomps:
comp = self.wflow.scope.get(name)
comp.set_itername(self.itername + '-fd')
comp.run(ffd_order=ffd_order, case_id=case_id)
def calc_derivatives(self,
first=False,
second=False,
savebase=True,
required_inputs=None,
required_outputs=None):
"""Calculate the derivatives for this non-differentiable block using
Finite Difference."""
# We don't do this in __init__ because some inputs and outputs
# are added after creation (for nested driver support).
if self.fd is None:
from openmdao.main.derivatives import FiniteDifference
self.fd = FiniteDifference(self)
if hasattr(self.wflow, '_severed_edges'):
self.wflow.sever_edges(self.wflow._severed_edges)
try:
# First, linearize about operating point.
# Note: Only needed for differentiable islands, which are handled
# with Fake Finite Difference.
# Don't do this for full-model finite difference.
if first and self.ffd_order > 0:
for name in self.comps:
# TODO: I think the cache is blown away each time before
# this is called
if name in self.wflow._J_cache:
continue
comp = self.wflow.scope.get(name)
# Assemblies need some required inputs and outputs
# to calculate the Jacobians
if has_interface(comp, IAssembly):
# Need to know which assy bdry variables are
# required, and pass them in. Cache this once.
if name not in self.ffd_cache:
dgraph = self.wflow.scope._depgraph
inputs = [
dgraph.base_var(inp)
for inp in flatten_list_of_iters(self.inputs)
]
outputs = [
dgraph.base_var(outp) for outp in self.outputs
]
from openmdao.main.depgraph import _get_inner_edges
edges = _get_inner_edges(dgraph, inputs, outputs)
req_inputs = []
req_outputs = []
for inp in inputs:
comp_str, _, var_str = inp.partition('.')
if comp_str == name:
req_inputs.append(var_str)
for inp in outputs:
comp_str, _, var_str = inp.partition('.')
if comp_str == name:
req_outputs.append(var_str)
for edge in edges:
src, target = edge
comp_str, _, var_str = src.partition('.')
if comp_str == name:
req_outputs.append(var_str)
comp_str, _, var_str = target.partition('.')
if comp_str == name:
req_inputs.append(var_str)
self.ffd_cache[name] = (req_inputs, req_outputs)
req_inputs, req_outputs = self.ffd_cache[name]
comp.calc_derivatives(first,
second,
savebase=True,
required_inputs=req_inputs,
required_outputs=req_outputs)
# Comp list contains full graph, so don't double up on
# the subdrivers.
elif not has_interface(comp, IDriver):
comp.calc_derivatives(first, second, True)
self.J = self.fd.calculate()
finally:
if hasattr(self.wflow, '_severed_edges'):
self.wflow.unsever_edges()
return self.J
def provideJ(self):
"""Jacobian for this block"""
return self.J
def list_deriv_vars(self):
"""Derivative inputs and outputs for this block"""
return self.mapped_inputs, self.mapped_outputs
def get(self, varname):
""" Return the value of a variable in the Pseudoassembly. Used
when sizing variables in the Jacobian."""
return self.wflow.scope.get(from_PA_var(self.name + '.' + varname))
def add_to_graph(self, startgraph, dgraph, excludes=()):
"""Add this PseudoAssembly to the given graph, absorbing
nodes that represent components contained in this PA.
"""
from openmdao.main.depgraph import is_subvar_node, \
is_input_base_node, is_output_base_node
self._removed_comps = [c for c in self.comps if c not in excludes]
# Add pseudoassys to graph
dgraph.add_node(self.name,
pa_object=self,
comp=True,
pseudo='assembly',
valid=True)
empty = {}
for oldname, newname in self.renames.items():
attrs = startgraph.node.get(oldname, empty).copy()
if oldname in self.boundary_params:
attrs['iotype'] = 'out'
dgraph.add_node(newname, attr_dict=attrs)
for u, v, data in dgraph.edges(oldname, data=True):
dgraph.add_edge(newname, v, attr_dict=data.copy())
for u, v, data in dgraph.in_edges(oldname, data=True):
dgraph.add_edge(u, newname, attr_dict=data.copy())
if is_subvar_node(dgraph, newname):
dgraph.node[newname]['basevar'] = \
newname.split('[', 1)[0]
if is_input_base_node(dgraph, newname):
dgraph.add_edge(newname, self.name)
elif is_output_base_node(dgraph, newname):
dgraph.add_edge(self.name, newname)
def clean_graph(self, startgraph, dgraph, excludes=()):
""" Clean up the old nodes in the graph, leaving out ones that
are excluded because they're used by ancestor drivers.
"""
dgraph.remove_nodes_from(
dgraph.find_prefixed_nodes([
n for n in self._orig_group_nodes
if n not in excludes and n in dgraph
]))
# if this PA represents a driver, remove the corresponding driver
# node
if self.name[1:] in dgraph:
dgraph.remove_nodes_from(
dgraph.find_prefixed_nodes([self.name[1:]]))
# update mapped inputs/outputs in the graph
map_inputs = dgraph.graph['mapped_inputs']
map_outputs = dgraph.graph['mapped_outputs']
for i, varpath in enumerate(map_inputs):
if isinstance(varpath, basestring):
varpath = [varpath]
mapped = []
for path in varpath:
compname, _, varname = path.partition('.')
if varname and (compname in self.comps):
mapped.append(to_PA_var(path, self.name))
else:
mapped.append(path) # keep old value in that spot
map_inputs[i] = tuple(mapped)
# Add requested outputs
for i, varpath in enumerate(map_outputs):
compname, _, varname = varpath.partition('.')
if varname and (compname in self.comps):
map_outputs[i] = to_PA_var(varpath, self.name)
def calc_derivatives(self, first=False, second=False, savebase=True,
required_inputs=None, required_outputs=None):
"""Calculate the derivatives for this non-differentiable block using
Finite Difference."""
# We don't do this in __init__ because some inputs and outputs
# are added after creation (for nested driver support).
if self.fd is None:
from openmdao.main.derivatives import FiniteDifference
self.fd = FiniteDifference(self)
if hasattr(self.wflow, '_severed_edges'):
self.wflow.sever_edges(self.wflow._severed_edges)
try:
# First, linearize about operating point.
# Note: Only needed for differentiable islands, which are handled
# with Fake Finite Difference.
# Don't do this for full-model finite difference.
if first and self.ffd_order > 0:
for name in self.comps:
# TODO: I think the cache is blown away each time before
# this is called
if name in self.wflow._J_cache:
continue
comp = self.wflow.scope.get(name)
# Assemblies need some required inputs and outputs
# to calculate the Jacobians
if has_interface(comp, IAssembly):
# Need to know which assy bdry variables are
# required, and pass them in. Cache this once.
if name not in self.ffd_cache:
dgraph = self.wflow.scope._depgraph
inputs = [dgraph.base_var(inp)
for inp in flatten_list_of_iters(self.inputs)]
outputs = [dgraph.base_var(outp)
for outp in self.outputs]
from openmdao.main.depgraph import _get_inner_edges
edges = _get_inner_edges(dgraph, inputs, outputs)
req_inputs = []
req_outputs = []
for inp in inputs:
comp_str, _, var_str = inp.partition('.')
if comp_str == name:
req_inputs.append(var_str)
for inp in outputs:
comp_str, _, var_str = inp.partition('.')
if comp_str == name:
req_outputs.append(var_str)
for edge in edges:
src, target = edge
comp_str, _, var_str = src.partition('.')
if comp_str == name:
req_outputs.append(var_str)
comp_str, _, var_str = target.partition('.')
if comp_str == name:
req_inputs.append(var_str)
self.ffd_cache[name] = (req_inputs, req_outputs)
req_inputs, req_outputs = self.ffd_cache[name]
comp.calc_derivatives(first, second, savebase=True,
required_inputs=req_inputs,
required_outputs=req_outputs)
# Comp list contains full graph, so don't double up on
# the subdrivers.
elif not has_interface(comp, IDriver):
comp.calc_derivatives(first, second, True)
self.J = self.fd.calculate()
finally:
if hasattr(self.wflow, '_severed_edges'):
self.wflow.unsever_edges()
return self.J