# coding: utf-8
# ------------------------------------------------------------------------------
# Copyright (C) 2012-2017 Guillaume Sagnol
# Copyright (C) 2017-2019 Maximilian Stahlberg
#
# This file is part of PICOS.
#
# PICOS is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# PICOS is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program. If not, see <http://www.gnu.org/licenses/>.
# ------------------------------------------------------------------------------
"""Implementation of :class:`CVXOPTSolver`."""
import cvxopt
import numpy
from ..apidoc import api_end, api_start
from ..constraints import (AffineConstraint, DummyConstraint, LMIConstraint,
LogSumExpConstraint, RSOCConstraint, SOCConstraint)
from ..expressions import CONTINUOUS_VARTYPES, AffineExpression, LogSumExp
from ..modeling.footprint import Specification
from ..modeling.solution import (PS_FEASIBLE, PS_INFEASIBLE, PS_UNBOUNDED,
PS_UNKNOWN, SS_INFEASIBLE, SS_OPTIMAL,
SS_UNKNOWN)
from .solver import Solver, SolverError
_API_START = api_start(globals())
# -------------------------------
[docs]class CVXOPTSolver(Solver):
"""Interface to the CVXOPT solver.
Also used as an interface to the
:class:`SMCP solver <picos.solvers.solver_smcp.SMCPSolver>`.
"""
SUPPORTED = Specification(
objectives=[
AffineExpression,
LogSumExp],
variables=CONTINUOUS_VARTYPES,
constraints=[
DummyConstraint,
AffineConstraint,
SOCConstraint,
RSOCConstraint,
LMIConstraint,
LogSumExpConstraint])
[docs] @classmethod
def supports(cls, footprint):
"""Implement :meth:`~.solver.Solver.supports`."""
if not Solver.supports(footprint):
return False
return footprint in cls.SUPPORTED
[docs] @classmethod
def default_penalty(cls):
"""Implement :meth:`~.solver.Solver.default_penalty`."""
return 1.0 # Stable free/open source solver.
[docs] @classmethod
def test_availability(cls):
"""Implement :meth:`~.solver.Solver.test_availability`."""
# CVXOPT is a dependency of PICOS, so it is always available.
pass
[docs] @classmethod
def names(cls):
"""Implement :meth:`~.solver.Solver.names`."""
return "cvxopt", "CVXOPT", "Python Convex Optimization Solver"
[docs] @classmethod
def is_free(cls):
"""Implement :meth:`~.solver.Solver.is_free`."""
return True
@property
def is_smcp(self):
"""Whether to implement SMCP instead of CVXOPT."""
return False
[docs] def __init__(self, problem):
"""Initialize a CVXOPT solver interface.
:param ~picos.Problem problem: The problem to be solved.
"""
super(CVXOPTSolver, self).__init__(problem)
self._numVars = 0
"""Total number of scalar variables passed to CVXOPT."""
self._cvxoptVarOffset = {}
"""Maps a PICOS variable to its offset in the constraint matrix."""
# HACK: Setting this to false would result in variable bounds to be
# ignored, instead of added to the linear inequalities matrix.
# This is used by a Problem method that prints the problem to a
# file using a CVXOPTSolver instance.
self.import_variable_bounds = True
# SMCP currently offers no option reset, so keep a backup.
if self.is_smcp:
import smcp
self._smcp_default_options = smcp.solvers.options.copy()
[docs] def reset_problem(self):
"""Implement :meth:`~.solver.Solver.reset_problem`."""
self.int = None
self._numVars = 0
self._cvxoptVarOffset.clear()
def _is_geometric_program(self):
return isinstance(self.ext.no.function, LogSumExp) or \
any(isinstance(constraint, LogSumExpConstraint)
for constraint in self.ext.constraints.values())
def _affine_expression_to_G_and_h(self, expression):
assert isinstance(expression, AffineExpression)
length = len(expression)
# Construct G.
I, J, V = [], [], []
for variable, coefficients in expression._linear_coefs.items():
cvxoptVarOffset = self._cvxoptVarOffset[variable]
if not isinstance(coefficients, cvxopt.spmatrix):
coefficients = cvxopt.sparse(coefficients)
I.extend(coefficients.I)
J.extend([cvxoptVarOffset + j for j in coefficients.J])
V.extend(coefficients.V)
G = cvxopt.spmatrix(V, I, J, (length, self._numVars), tc="d")
# Construct h.
# TODO: Check if making a copy is necessary.
h = cvxopt.matrix(expression._constant_coef, tc="d")
return G, h
_Gh = _affine_expression_to_G_and_h
def _import_variables_without_bounds(self):
offset = 0
for variable in self.ext.variables.values():
dim = variable.dim
# Register the variable.
self._cvxoptVarOffset[variable] = offset
offset += dim
assert offset == self._numVars
def _import_variable_bounds(self):
for variable in self.ext.variables.values():
if self.import_variable_bounds:
bounds = variable.bound_constraint
if bounds:
self._import_affine_constraint(bounds)
def _import_objective(self):
direction, objective = self.ext.no
if self._is_geometric_program():
if isinstance(objective, LogSumExp):
(F, g) = self._Gh(objective.x)
else:
assert isinstance(objective, AffineExpression)
(F, g) = self._Gh(objective)
# NOTE: Needs to happen before LogSumExpConstraint are added.
self.int["K"] = [F.size[0]]
if direction == "max":
F, g = -F, -g
self.int["F"] = F
self.int["g"] = g
else:
(c, _) = self._Gh(objective)
if direction == "max":
c = -c
# Must be a dense column-vector.
c = cvxopt.matrix(c).T
self.int["c"] = c
def _import_affine_constraint(self, constraint):
assert isinstance(constraint, AffineConstraint)
(G_smaller, h_smaller) = self._Gh(constraint.smaller)
(G_greater, h_greater) = self._Gh(constraint.greater)
G = G_smaller - G_greater
h = h_greater - h_smaller
if constraint.is_equality():
self.int["A"] = cvxopt.sparse([self.int["A"], G])
self.int["b"] = cvxopt.matrix([self.int["b"], h])
else:
self.int["Gl"] = cvxopt.sparse([self.int["Gl"], G])
self.int["hl"] = cvxopt.matrix([self.int["hl"], h])
def _import_soc_constraint(self, constraint):
assert isinstance(constraint, SOCConstraint)
(A, b) = self._Gh(constraint.ne)
(c, d) = self._Gh(constraint.ub)
self.int["Gq"].append(cvxopt.sparse([-c, -A]))
self.int["hq"].append(cvxopt.matrix([d, b]))
def _import_rsoc_constraint(self, constraint):
assert isinstance(constraint, RSOCConstraint)
(A, b) = self._Gh(constraint.ne)
(c1, d1) = self._Gh(constraint.ub1)
(c2, d2) = self._Gh(constraint.ub2)
self.int["Gq"].append(cvxopt.sparse([-c1 - c2, -2 * A, c2 - c1]))
self.int["hq"].append(cvxopt.matrix([d1 + d2, 2 * b, d1 - d2]))
def _import_lmi_constraint(self, constraint):
assert isinstance(constraint, LMIConstraint)
(G_smaller, h_smaller) = self._Gh(constraint.smaller)
(G_greater, h_greater) = self._Gh(constraint.greater)
self.int["Gs"].append(G_smaller - G_greater)
self.int["hs"].append(h_greater - h_smaller)
def _import_lse_constraint(self, constraint):
assert isinstance(constraint, LogSumExpConstraint)
(F, g) = self._Gh(constraint.le0.x)
self.int["F"] = cvxopt.sparse([self.int["F"], F])
self.int["g"] = cvxopt.matrix([self.int["g"], g])
self.int["K"].append(F.size[0])
def _import_constraint(self, constraint):
if isinstance(constraint, AffineConstraint):
self._import_affine_constraint(constraint)
elif isinstance(constraint, SOCConstraint):
self._import_soc_constraint(constraint)
elif isinstance(constraint, RSOCConstraint):
self._import_rsoc_constraint(constraint)
elif isinstance(constraint, LMIConstraint):
self._import_lmi_constraint(constraint)
elif isinstance(constraint, LogSumExpConstraint):
self._import_lse_constraint(constraint)
else:
assert isinstance(constraint, DummyConstraint), \
"Unexpected constraint type: {}".format(
constraint.__class__.__name__)
def _import_problem(self):
self._numVars = sum(var.dim for var in self.ext.variables.values())
# CVXOPT's internal problem representation is stateless; a number of
# matrices are supplied to the appropriate solver function each time a
# search is started. These matrices are thus stored in self.int.
self.int = {
# Objective function coefficients.
"c": None,
# Linear equality left hand side.
"A": cvxopt.spmatrix([], [], [], (0, self._numVars), tc="d"),
# Linear equality right hand side.
"b": cvxopt.matrix([], (0, 1), tc="d"),
# Linear inequality left hand side.
"Gl": cvxopt.spmatrix([], [], [], (0, self._numVars), tc="d"),
# Linear inequality right hand side.
"hl": cvxopt.matrix([], (0, 1), tc="d"),
# Second order cone inequalities left hand sides.
"Gq": [],
# Second order cone inequalities right hand sides.
"hq": [],
# Semidefinite cone inequalities left hand sides.
"Gs": [],
# Semidefinite cone inequalities right hand sides.
"hs": [],
# Geometric program data.
"F": None,
"g": None,
"K": None
}
# Import variables without their bounds.
self._import_variables_without_bounds()
# Set objective.
# NOTE: This needs to happen before constraints are added as
# self.int["K"][0] refers to an LSE objective while
# self.int["K"][i] with i > 0 refers to LSE constraints.
self._import_objective()
# Import constraints.
for constraint in self.ext.constraints.values():
self._import_constraint(constraint)
# Import variable bounds as additional affine constraints.
# NOTE: This needs to happen after constraints are added due to how the
# dual values for constraints are extracted.
self._import_variable_bounds()
def _update_problem(self):
raise NotImplementedError
def _solve(self):
if self.is_smcp:
import smcp
p = self.int
isGP = self._is_geometric_program()
# Clear all options set previously. This is necessary because CVXOPT
# options are global, and might be changed even by another problem.
cvxopt.solvers.options.clear()
# verbosity
cvxopt.solvers.options["show_progress"] = (self.verbosity() >= 1)
# rel_prim_fsb_tol, rel_dual_fsb_tol
feasibilityTols = [tol for tol in (self.ext.options.rel_prim_fsb_tol,
self.ext.options.rel_dual_fsb_tol) if tol is not None]
if feasibilityTols:
cvxopt.solvers.options["feastol"] = min(feasibilityTols)
# abs_ipm_opt_tol
if self.ext.options.abs_ipm_opt_tol is not None:
cvxopt.solvers.options["abstol"] = self.ext.options.abs_ipm_opt_tol
# rel_ipm_opt_tol
if self.ext.options.rel_ipm_opt_tol is not None:
cvxopt.solvers.options["reltol"] = self.ext.options.rel_ipm_opt_tol
# max_iterations
if self.ext.options.max_iterations is not None:
cvxopt.solvers.options["maxiters"] = self.ext.options.max_iterations
else:
cvxopt.solvers.options["maxiters"] = int(1e6)
# Allow solving GPs that are not strictly feasible and work around
# constraint matrix rank requirements.
if self.ext.options.cvxopt_kktreg is not None:
cvxopt.solvers.options["kktreg"] = self.ext.options.cvxopt_kktreg
# Handle unsupported options.
self._handle_unsupported_options(
"lp_root_method", "lp_node_method", "timelimit", "treememory",
"max_fsb_nodes", "hotstart")
# TODO: Add CVXOPT-sepcific options. Candidates are:
# - refinement
# Set options for SMCP.
if self.is_smcp:
# Restore default options.
smcp.solvers.options = self._smcp_default_options.copy()
# Copy options also used by CVXOPT.
smcp.solvers.options.update(cvxopt.solvers.options)
# Further handle "verbose" option.
smcp.solvers.options["debug"] = (self.verbosity() >= 2)
# TODO: Add SMCP-sepcific options.
if self._debug():
from pprint import pformat
self._debug("Setting options:\n{}\n".format(pformat(
smcp.solvers.options if self.is_smcp
else cvxopt.solvers.options)))
# Print a header.
if self.is_smcp:
subsolverText = None
else:
if isGP:
subsolverText = "internal GP solver"
else:
subsolverText = "internal CONELP solver"
# Further prepare the problem for the CVXOPT/SMCP CONELP solvers.
# TODO: This should be done during import.
if not isGP:
# Retrieve the structure of the cone, which is a cartesian product
# of the non-negative orthant of dimension l, a number of second
# order cones with dimensions in q and a number of positive
# semidefinite cones with dimensions in s.
dims = {
"l": p["Gl"].size[0],
"q": [Gqi.size[0] for Gqi in p["Gq"]],
"s": [int(numpy.sqrt(Gsi.size[0])) for Gsi in p["Gs"]]
}
# Construct G and h to contain all conic inequalities, starting with
# those with respect to the non-negative orthant.
G = p["Gl"]
h = p["hl"]
# SMCP's ConeLP solver does not handle (linear) equalities, so cast
# them as inequalities.
if self.is_smcp:
smcp_eps = min(feasibilityTols)
if p["A"].size[0] > 0:
G = cvxopt.sparse([G, p["A"]])
G = cvxopt.sparse([G, -p["A"]])
h = cvxopt.matrix([h, p["b"]+smcp_eps])
h = cvxopt.matrix([h, smcp_eps-p["b"]])
dims["l"] += (2 * p["A"].size[0])
# Remove the lines in G and h corresponding to 0==0 or 0<=0
JP = list(set(G.I))
IP = range(len(JP))
VP = [1] * len(JP)
if len(JP) != dims["l"]:
# is there a constraint of the form 0<=a, (a<0) ?
if any([b < -smcp_eps for (i, b) in enumerate(h)
if i not in JP]):
raise Exception(
'infeasible constraint of the form '
'0 <= a, with a<0')
# left-multiply with PPP-matrix to remove 0-constraints
PPP = cvxopt.spmatrix(VP, IP, JP, (len(IP), G.size[0]))
dims["l"] = len(JP)
G = PPP * G
h = PPP * h
# Add second-order cone inequalities.
for i in range(len(dims["q"])):
G = cvxopt.sparse([G, p["Gq"][i]])
h = cvxopt.matrix([h, p["hq"][i]])
# Add semidefinite cone inequalities.
for i in range(len(dims["s"])):
G = cvxopt.sparse([G, p["Gs"][i]])
h = cvxopt.matrix([h, p["hs"][i]])
# Remove zero lines from linear equality constraint matrix, as
# CVXOPT expects this matrix to have full row rank.
JP = list(set(p["A"].I))
IP = range(len(JP))
VP = [1] * len(JP)
if any([b for (i, b) in enumerate(p["b"]) if i not in JP]):
raise SolverError(
"Infeasible constraint of the form 0 = a with a != 0.")
P = cvxopt.spmatrix(VP, IP, JP, (len(IP), p["A"].size[0]))
A = P * p["A"]
b = P * p["b"]
# Attempt to solve the problem.
with self._header(subsolverText), self._stopwatch():
if self.is_smcp:
if self._debug():
self._debug("Calling smcp.solvers.conelp(c, G, h, dims) "
"with\nc:\n{}\nG:\n{}\nh:\n{}\ndims:\n{}\n"
.format(p["c"], G, h, dims))
try:
result = smcp.solvers.conelp(p["c"], G, h, dims)
except TypeError:
# HACK: Work around "'NoneType' object is not subscriptable"
# exception with infeasible/unbounded problems.
result = None
else:
kwargs = {}
if self.ext.options.cvxopt_kktsolver is not None:
kwargs["kktsolver"] = self.ext.options.cvxopt_kktsolver
if isGP:
result = cvxopt.solvers.gp(p["K"], p["F"], p["g"], p["Gl"],
p["hl"], p["A"], p["b"], **kwargs)
else:
result = cvxopt.solvers.conelp(
p["c"], G, h, dims, A, b, **kwargs)
# Retrieve primals.
primals = {}
if self.ext.options.primals is not False and result is not None \
and result["x"] is not None:
for variable in self.ext.variables.values():
offset = self._cvxoptVarOffset[variable]
value = list(result["x"][offset:offset + variable.dim])
primals[variable] = value
# Retrieve duals.
duals = {}
if self.ext.options.duals is not False and result is not None:
(indy, indzl, indzq, indznl, indzs) = (0, 0, 0, 0, 0)
if isGP:
zkey = "zl"
zqkey = "zq"
zskey = "zs"
else:
zkey = "z"
zqkey = "z"
zskey = "z"
indzq = dims["l"]
indzs = dims["l"] + sum(dims["q"])
if self.is_smcp:
# Equality constraints were cast as two inequalities.
ieq = p["Gl"].size[0]
neq = (dims["l"] - ieq) // 2
soleq = result["z"][ieq:ieq + neq]
soleq -= result["z"][ieq + neq:ieq + 2 * neq]
else:
soleq = result["y"]
for constraint in self.ext.constraints.values():
if isinstance(constraint, DummyConstraint):
duals[constraint] = cvxopt.spmatrix(
[], [], [], constraint.size)
continue
dual = None
consSz = len(constraint)
if isinstance(constraint, AffineConstraint):
if constraint.is_equality():
if soleq is not None:
dual = -(P.T * soleq)[indy:indy + consSz]
indy += consSz
else:
if result[zkey] is not None:
dual = result[zkey][indzl:indzl + consSz]
indzl += consSz
elif isinstance(constraint, SOCConstraint) \
or isinstance(constraint, RSOCConstraint):
if result[zqkey] is not None:
if isGP:
dual = result[zqkey][indzq]
dual[1:] = -dual[1:]
indzq += 1
else:
dual = result[zqkey][indzq:indzq + consSz]
if isinstance(constraint, RSOCConstraint):
# RScone were cast as a SOcone on import, so
# transform the dual to a proper RScone dual.
alpha = dual[0] + dual[-1]
beta = dual[0] - dual[-1]
z = 2.0 * dual[1:-1]
dual = cvxopt.matrix([alpha, beta, z])
indzq += consSz
elif isinstance(constraint, LMIConstraint):
if result[zskey] is not None:
matSz = constraint.size[0]
if isGP:
dual = cvxopt.matrix(
result[zskey][indzs], (matSz, matSz))
indzs += 1
else:
dual = cvxopt.matrix(
result[zskey][indzs:indzs + consSz],
(matSz, matSz))
indzs += consSz
elif isinstance(constraint, LogSumExpConstraint):
# TODO: Retrieve LSE duals.
indznl += 1
duals[constraint] = dual
# Retrieve objective value.
if result is None:
value = None
elif isGP:
value = None
else:
p = result['primal objective']
d = result['dual objective']
if p is not None and d is not None:
value = 0.5 * (p + d)
elif p is not None:
value = p
elif d is not None:
value = d
else:
value = None
if value is not None and self.ext.no.direction == "max":
value = -value
if self.is_smcp:
value = -value
# Retrieve solution status.
status = result["status"] if result else "unknown"
if status == "optimal":
primalStatus = SS_OPTIMAL
dualStatus = SS_OPTIMAL
problemStatus = PS_FEASIBLE
elif status == "primal infeasible":
primalStatus = SS_INFEASIBLE
dualStatus = SS_UNKNOWN
problemStatus = PS_INFEASIBLE
elif status == "dual infeasible":
primalStatus = SS_UNKNOWN
dualStatus = SS_INFEASIBLE
problemStatus = PS_UNBOUNDED
else:
primalStatus = SS_UNKNOWN
dualStatus = SS_UNKNOWN
problemStatus = PS_UNKNOWN
return self._make_solution(value, primals, duals, primalStatus,
dualStatus, problemStatus, {"cvxopt_sol": result})
# --------------------------------------
__all__ = api_end(_API_START, globals())