# Standard Python modules
from collections import OrderedDict
import copy
from typing import Dict, Iterable, List, Optional, Union
# External modules
import numpy as np
# Local modules
from .pyOpt_error import Error, pyOptSparseWarning
from .pyOpt_utils import INFINITY, _broadcast_to_array, convertToCOO
from .types import Dict1DType
[docs]
class Constraint:
def __init__(
self,
name: str,
nCon: int,
linear: bool,
wrt: Union[None, str, Iterable[str]],
jac: Dict1DType,
lower,
upper,
scale,
):
"""
This class holds the representation of a pyOptSparse constraint group
See Also
--------
pyoptsparse.pyOpt_optimization.Optimization.addConGroup : for the full documentation
"""
self.name = name
self.ncon = nCon
self.linear = linear
self.wrt = wrt
self.jac = jac
self.partialReturnOk: Optional[bool] = None
self.scale = scale
self.rs: Optional[int] = None
self.re: Optional[int] = None
# Before we can do the processing below we need to have lower
# and upper arguments expanded:
lower = _broadcast_to_array("lower", lower, nCon, allow_none=True)
upper = _broadcast_to_array("upper", upper, nCon, allow_none=True)
scale = _broadcast_to_array("scale", scale, nCon)
# Save lower and upper...they are only used for printing however
self.lower = lower
self.upper = upper
# The current value of the constraint (for printing purposes)
self.value = np.zeros(self.ncon)
# Now we determine what kind of constraint this is:
# 1. An equality constraint
# 2. A upper bound on a 1-sided constraint
# 3. A lower bound on a 1-sided constraint
# 4. Lower and Upper bounds on 2-sided constraint
# 5. No lower or upper bounds. Typically will only be used for
# dummy constraint on an unconstrained problem.
# The first 3, will give a single "constraint" in all
# optimizers. Some optimizers can only do 1-sided constraints
# so type 4 and 5 must be split into two separate constraints
# automatically.
# This keeps track of the equality constraints:
equalityConstraints: Dict[str, List] = {"value": [], "ind": [], "fact": []}
# All (inequality) constraints get added to
# "twoSidedConstraints". This will be used in optimizers that
# can do two-sided constraints properly
twoSidedConstraints: Dict[str, List] = {"lower": [], "upper": [], "ind": [], "fact": []}
# All (inequality) constraints are also added to
# "oneSidedConstraints". These are processed such that the
# lower bound is ALWAYS -INFINITY such that: con <= upper For
# optimizers that need things <= zero, this can be processed
# with a (-value) offset. One sided constraints need a fact
# defined which is precisely 1.0 or -1.0. The -1.0 appears
# when a greater-than-constraint is turned into a
# less-than-constraint.
oneSidedConstraints: Dict[str, List] = {"lower": [], "upper": [], "ind": [], "fact": []}
for icon in range(self.ncon):
# Check for equality constraint:
if lower[icon] == upper[icon] and lower[icon] is not None:
equalityConstraints["value"].append(lower[icon] * scale[icon])
equalityConstraints["ind"].append(icon)
equalityConstraints["fact"].append(1.0)
# Two sided constraint:
elif lower[icon] is not None and upper[icon] is not None:
twoSidedConstraints["lower"].append(lower[icon] * scale[icon])
twoSidedConstraints["upper"].append(upper[icon] * scale[icon])
twoSidedConstraints["ind"].append(icon)
twoSidedConstraints["fact"].append(1.0)
# TWO sets of 1 sided constraints:
oneSidedConstraints["lower"].append(-INFINITY)
oneSidedConstraints["upper"].append(upper[icon] * scale[icon])
oneSidedConstraints["ind"].append(icon)
oneSidedConstraints["fact"].append(1.0)
oneSidedConstraints["lower"].append(-INFINITY)
oneSidedConstraints["upper"].append(-lower[icon] * scale[icon])
oneSidedConstraints["ind"].append(icon)
oneSidedConstraints["fact"].append(-1.0)
# Upper bound only:
elif upper[icon] is not None:
twoSidedConstraints["lower"].append(-INFINITY)
twoSidedConstraints["upper"].append(upper[icon] * scale[icon])
twoSidedConstraints["ind"].append(icon)
twoSidedConstraints["fact"].append(1.0)
# Just one, 1-sided constraint
oneSidedConstraints["lower"].append(-INFINITY)
oneSidedConstraints["upper"].append(upper[icon] * scale[icon])
oneSidedConstraints["ind"].append(icon)
oneSidedConstraints["fact"].append(1.0)
# Lower bound only:
elif lower[icon] is not None:
twoSidedConstraints["lower"].append(lower[icon] * scale[icon])
twoSidedConstraints["upper"].append(INFINITY)
twoSidedConstraints["ind"].append(icon)
twoSidedConstraints["fact"].append(1.0)
# Just one, 1-sided constraint
oneSidedConstraints["lower"].append(-INFINITY)
oneSidedConstraints["upper"].append(-lower[icon] * scale[icon])
oneSidedConstraints["ind"].append(icon)
oneSidedConstraints["fact"].append(-1.0)
# Fully unconstrained!
elif lower[icon] is None and upper[icon] is None:
twoSidedConstraints["lower"].append(-INFINITY)
twoSidedConstraints["upper"].append(INFINITY)
twoSidedConstraints["ind"].append(icon)
twoSidedConstraints["fact"].append(1.0)
# Since this is just a dummy constraint, we only need
# a single one....it can just be less than INFINITY
oneSidedConstraints["lower"].append(-INFINITY)
oneSidedConstraints["upper"].append(INFINITY)
oneSidedConstraints["ind"].append(icon)
oneSidedConstraints["fact"].append(1.0)
# end if (con type)
# end for (con loop)
# Convert the stuff to arrays:
oneSidedConstraints["ind"] = np.array(oneSidedConstraints["ind"], "intc")
twoSidedConstraints["ind"] = np.array(twoSidedConstraints["ind"], "intc")
equalityConstraints["ind"] = np.array(equalityConstraints["ind"], "intc")
oneSidedConstraints["fact"] = np.array(oneSidedConstraints["fact"])
twoSidedConstraints["fact"] = np.array(twoSidedConstraints["fact"])
equalityConstraints["fact"] = np.array(equalityConstraints["fact"])
equalityConstraints["value"] = np.array(equalityConstraints["value"])
# Now save this information:
self.equalityConstraints = equalityConstraints
self.oneSidedConstraints = oneSidedConstraints
self.twoSidedConstraints = twoSidedConstraints
[docs]
def finalize(self, variables: OrderedDict, dvOffset, index: int):
"""
After the design variables have been finalized and the order
is known we can check the constraint for consistency.
Parameters
----------
variables : OrderedDict
The pyOpt variable list after they have been finalized.
dvOffset : dict
Design variable offsets from pyOpt_optimization
index : int
The starting index of this constraint in natural order
Warnings
--------
This function should not need to be called by the user
"""
# Set the row start and end
self.rs = index
self.re = index + self.ncon
# First check if 'wrt' is supplied...if not we just take all
# the dvGroups
if self.wrt is None:
self.wrt = list(variables.keys())
else:
# Sanitize the wrt input:
if isinstance(self.wrt, str):
self.wrt = [self.wrt.lower()]
else:
try:
self.wrt = list(self.wrt)
except Exception:
raise Error(f"The 'wrt' argument to constraint '{self.name}' must be an iterable list")
# We allow 'None' to be in the list...they are null so
# just pop them out:
self.wrt = [dvGroup for dvGroup in self.wrt if dvGroup is not None]
# Now, make sure that each dvGroup the user supplied list
# *actually* are variables
for dvGroup in self.wrt:
if dvGroup not in variables:
raise Error(
f"The supplied dvGroup '{dvGroup}' in 'wrt' for the {self.name} constraint, does not exist. "
+ "It must be added with a call to addVar() or addVarGroup()."
)
# Check for duplicates in wrt
wrt_uniq = list(set(self.wrt))
if len(wrt_uniq) < len(self.wrt):
duplicate_vars = list({x for x in self.wrt if self.wrt.count(x) > 1})
pyOptSparseWarning(
f"The constraint {self.name} was created with duplicate variables in 'wrt'. "
+ "The following duplicates were automatically removed: "
)
for var in duplicate_vars:
print(f"\t\t{var}")
self.wrt = wrt_uniq
# Last thing for wrt is to reorder them such that dvGroups are
# in order. This way when the Jacobian is assembled in
# processDerivatives() the coorindate matrix will in the right
# order.
dvStart = []
for dvGroup in self.wrt:
dvStart.append(dvOffset[dvGroup][0])
# This sort wrt using the keys in dvOffset
self.wrt = [x for (y, x) in sorted(zip(dvStart, self.wrt))]
# Now we know which dvGroups this constraint will have a
# derivative with respect to (i.e. what is in the wrt list)
# Now, it is possible that Jacobians were given for none, some
# or all the dvGroups defined in wrt.
if self.jac is None:
# If the constraint is linear we have to *Force* the user to
# supply a constraint Jacobian for *each* of the values in
# wrt. Otherwise, a matrix of zeros isn't meaningful for the
# sparse constraints.
if self.linear:
raise Error(
"The 'jac' keyword to argument to addConGroup() must be supplied for a linear constraint. "
+ f"The constraint in error is {self.name}."
)
# without any additional information about the Jacobian
# structure, we must assume they are all dense.
self.jac = {}
for dvGroup in self.wrt:
ss = dvOffset[dvGroup]
ndvs = ss[1] - ss[0]
self.jac[dvGroup] = convertToCOO(np.zeros((self.ncon, ndvs)))
# Set a flag for the constraint object, that not returning
# them all is ok.
self.partialReturnOk = True
else:
# First sanitize input:
if not isinstance(self.jac, dict):
raise Error(
"The 'jac' keyword argument to addConGroup() must be a dictionary. "
+ f"The constraint in error is {self.name}."
)
# Now loop over the set we *know* we need and see if any
# are in jac. We will actually pop them out, and that way
# if there is anything left at the end, we can tell the
# user supplied information was unused.
tmp = copy.deepcopy(self.jac)
self.jac = {}
for dvGroup in self.wrt:
ss = dvOffset[dvGroup]
ndvs = ss[1] - ss[0]
try:
self.jac[dvGroup] = tmp.pop(dvGroup)
except KeyError:
# No big deal, just make a dense component...and
# set to zero
self.jac[dvGroup] = convertToCOO(np.zeros((self.ncon, ndvs)))
# Convert Now check that the supplied Jacobian to COO:
self.jac[dvGroup] = convertToCOO(self.jac[dvGroup])
# Generically check the shape:
if self.jac[dvGroup]["shape"][0] != self.ncon or self.jac[dvGroup]["shape"][1] != ndvs:
raise Error(
f"The supplied Jacobian for dvGroup {dvGroup}' in constraint {self.name}, was the incorrect size. "
+ f"Expecting a Jacobian of size ({self.ncon}, {ndvs}) but received a Jacobian of size "
+ f"({self.jac[dvGroup]['shape'][0]}, {self.jac[dvGroup]['shape'][1]})."
)
# end for (dvGroup)
# If there is anything left in jac print a warning:
for dvGroup in tmp:
pyOptSparseWarning(
f"A Jacobian with dvGroup key of '{dvGroup}' was unused in constraint {self.name}. "
+ "This will be ignored."
)
# Since this function *may* be called multiple times, only
# set paritalReturnOk if it was the first pass:
if self.partialReturnOk is None:
# Finally partial returns NOT ok, since the user has
# supplied information about the sparsity:
self.partialReturnOk = False
# end if (if Jac)
def __str__(self) -> str:
"""
Print Constraint
"""
res = ""
for i in range(self.ncon):
lower = self.lower[i]
upper = self.upper[i]
value = self.value[i]
if lower is None:
lower = -INFINITY
if upper is None:
upper = INFINITY
res += (
" "
+ str(self.name).center(9)
+ f" i {np.real(lower):15.2e} <= {np.real(value):8f} <= {np.real(upper):8.2e}\n"
)
return res