def test_psd_constraints(self):
""" Test positive semi-definite constraints
"""
C = Variable((3, 3))
obj = Maximize(C[0, 2])
constraints = [
diag(C) == 1, C[0, 1] == 0.6, C[1, 2] == -0.3, C == C.T, C >> 0
]
prob = Problem(obj, constraints)
self.assertTrue(FlipObjective().accepts(prob))
p_min = FlipObjective().apply(prob)
self.assertTrue(ConeMatrixStuffing().accepts(p_min[0]))
C = Variable((2, 2))
obj = Maximize(C[0, 1])
constraints = [C == 1, C >> [[2, 0], [0, 2]]]
prob = Problem(obj, constraints)
self.assertTrue(FlipObjective().accepts(prob))
p_min = FlipObjective().apply(prob)
self.assertTrue(ConeMatrixStuffing().accepts(p_min[0]))
C = Variable((2, 2), symmetric=True)
obj = Minimize(C[0, 0])
constraints = [C << [[2, 0], [0, 2]]]
prob, _ = CvxAttr2Constr().apply(Problem(obj, constraints))
self.assertTrue(ConeMatrixStuffing().accepts(prob))
def test_nonneg_constraints_backend(self) -> None:
x = Variable(shape=(2, ), name='x')
objective = Maximize(-4 * x[0] - 5 * x[1])
constr_expr = hstack(
[3 - (2 * x[0] + x[1]), 3 - (x[0] + 2 * x[1]), x[0], x[1]])
constraints = [NonNeg(constr_expr)]
prob = Problem(objective, constraints)
self.assertFalse(ConeMatrixStuffing().accepts(prob))
self.assertTrue(FlipObjective().accepts(prob))
p_min = FlipObjective().apply(prob)
self.assertTrue(ConeMatrixStuffing().accepts(p_min[0]))
def construct_intermediate_chain(problem, candidates, gp: bool = False):
"""
Builds a chain that rewrites a problem into an intermediate
representation suitable for numeric reductions.
Parameters
----------
problem : Problem
The problem for which to build a chain.
candidates : dict
Dictionary of candidate solvers divided in qp_solvers
and conic_solvers.
gp : bool
If True, the problem is parsed as a Disciplined Geometric Program
instead of as a Disciplined Convex Program.
Returns
-------
Chain
A Chain that can be used to convert the problem to an intermediate form.
Raises
------
DCPError
Raised if the problem is not DCP and `gp` is False.
DGPError
Raised if the problem is not DGP and `gp` is True.
"""
reductions = []
if len(problem.variables()) == 0:
return Chain(reductions=reductions)
# TODO Handle boolean constraints.
if complex2real.accepts(problem):
reductions += [complex2real.Complex2Real()]
if gp:
reductions += [Dgp2Dcp()]
if not gp and not problem.is_dcp():
append = build_non_disciplined_error_msg(problem, 'DCP')
if problem.is_dgp():
append += ("\nHowever, the problem does follow DGP rules. "
"Consider calling solve() with `gp=True`.")
elif problem.is_dqcp():
append += ("\nHowever, the problem does follow DQCP rules. "
"Consider calling solve() with `qcp=True`.")
raise DCPError("Problem does not follow DCP rules. Specifically:\n" +
append)
elif gp and not problem.is_dgp():
append = build_non_disciplined_error_msg(problem, 'DGP')
if problem.is_dcp():
append += ("\nHowever, the problem does follow DCP rules. "
"Consider calling solve() with `gp=False`.")
elif problem.is_dqcp():
append += ("\nHowever, the problem does follow DQCP rules. "
"Consider calling solve() with `qcp=True`.")
raise DGPError("Problem does not follow DGP rules." + append)
# Dcp2Cone and Qp2SymbolicQp require problems to minimize their objectives.
if type(problem.objective) == Maximize:
reductions += [FlipObjective()]
# First, attempt to canonicalize the problem to a linearly constrained QP.
if candidates['qp_solvers'] and qp2symbolic_qp.accepts(problem):
reductions += [CvxAttr2Constr(), Qp2SymbolicQp()]
return Chain(reductions=reductions)
# Canonicalize it to conic problem.
if not candidates['conic_solvers']:
raise SolverError("Problem could not be reduced to a QP, and no "
"conic solvers exist among candidate solvers "
"(%s)." % candidates)
reductions += [Dcp2Cone(), CvxAttr2Constr()]
return Chain(reductions=reductions)
def _solve(self,
solver=None,
warm_start=True,
verbose=False,
gp=False,
qcp=False,
requires_grad=False,
enforce_dpp=False,
**kwargs):
"""Solves a DCP compliant optimization problem.
Saves the values of primal and dual variables in the variable
and constraint objects, respectively.
Arguments
---------
solver : str, optional
The solver to use. Defaults to ECOS.
warm_start : bool, optional
Should the previous solver result be used to warm start?
verbose : bool, optional
Overrides the default of hiding solver output.
gp : bool, optional
If True, parses the problem as a disciplined geometric program.
qcp : bool, optional
If True, parses the problem as a disciplined quasiconvex program.
requires_grad : bool, optional
Makes it possible to compute gradients with respect to
parameters by calling `backward()` after solving, or to compute
perturbations to the variables by calling `derivative()`. When
True, the solver must be SCS, and dqcp must be False.
A DPPError is thrown when problem is not DPP.
enforce_dpp : bool, optional
When True, a DPPError will be thrown when trying to solve a non-DPP
problem (instead of just a warning). Defaults to False.
kwargs : dict, optional
A dict of options that will be passed to the specific solver.
In general, these options will override any default settings
imposed by cvxpy.
Returns
-------
float
The optimal value for the problem, or a string indicating
why the problem could not be solved.
"""
for parameter in self.parameters():
if parameter.value is None:
raise error.ParameterError(
"A Parameter (whose name is '%s') does not have a value "
"associated with it; all Parameter objects must have "
"values before solving a problem." % parameter.name())
if requires_grad:
dpp_context = 'dgp' if gp else 'dcp'
if qcp:
raise ValueError("Cannot compute gradients of DQCP problems.")
elif not self.is_dpp(dpp_context):
raise error.DPPError("Problem is not DPP (when requires_grad "
"is True, problem must be DPP).")
elif solver is not None and solver not in [s.SCS, s.DIFFCP]:
raise ValueError("When requires_grad is True, the only "
"supported solver is SCS "
"(received %s)." % solver)
elif s.DIFFCP not in slv_def.INSTALLED_SOLVERS:
raise ImportError(
"The Python package diffcp must be installed to "
"differentiate through problems. Please follow the "
"installation instructions at "
"https://github.com/cvxgrp/diffcp")
else:
solver = s.DIFFCP
else:
if gp and qcp:
raise ValueError("At most one of `gp` and `qcp` can be True.")
if qcp and not self.is_dcp():
if not self.is_dqcp():
raise error.DQCPError("The problem is not DQCP.")
reductions = [dqcp2dcp.Dqcp2Dcp()]
if type(self.objective) == Maximize:
reductions = [FlipObjective()] + reductions
chain = Chain(problem=self, reductions=reductions)
soln = bisection.bisect(chain.reduce(),
solver=solver,
verbose=verbose,
**kwargs)
self.unpack(chain.retrieve(soln))
return self.value
data, solving_chain, inverse_data = self.get_problem_data(
solver, gp, enforce_dpp)
solution = solving_chain.solve_via_data(self, data, warm_start,
verbose, kwargs)
self.unpack_results(solution, solving_chain, inverse_data)
return self.value
def test_scalar_lp(self):
"""Test scalar LP problems.
"""
for solver in self.solvers:
p = Problem(Minimize(3 * self.a), [self.a >= 2])
self.assertTrue(ConeMatrixStuffing().accepts(p))
result = p.solve(solver.name())
p_new = ConeMatrixStuffing().apply(p)
result_new = p_new[0].solve(solver.name())
self.assertAlmostEqual(result, result_new)
sltn = solver.solve(p_new[0], False, False, {})
self.assertAlmostEqual(sltn.opt_val, result)
inv_sltn = ConeMatrixStuffing().invert(sltn, p_new[1])
self.assertAlmostEqual(inv_sltn.opt_val, result)
self.assertAlmostEqual(inv_sltn.primal_vars[self.a.id],
self.a.value)
# TODO: Maximize
p = Problem(Minimize(-3 * self.a + self.b),
[self.a <= 2, self.b == self.a, self.b <= 5])
result = p.solve(solver.name())
self.assertTrue(ConeMatrixStuffing().accepts(p))
p_new = ConeMatrixStuffing().apply(p)
sltn = solver.solve(p_new[0], False, False, {})
self.assertAlmostEqual(sltn.opt_val, result)
inv_sltn = ConeMatrixStuffing().invert(sltn, p_new[1])
self.assertAlmostEqual(inv_sltn.opt_val, result)
self.assertAlmostEqual(inv_sltn.primal_vars[self.a.id],
self.a.value)
self.assertAlmostEqual(inv_sltn.primal_vars[self.b.id],
self.b.value)
# With a constant in the objective.
p = Problem(Minimize(3 * self.a - self.b + 100), [
self.a >= 2, self.b + 5 * self.c - 2 == self.a,
self.b <= 5 + self.c
])
self.assertTrue(ConeMatrixStuffing().accepts(p))
result = p.solve(solver.name())
p_new = ConeMatrixStuffing().apply(p)
sltn = solver.solve(p_new[0], False, False, {})
self.assertAlmostEqual(sltn.opt_val, result - 100)
inv_sltn = ConeMatrixStuffing().invert(sltn, p_new[1])
self.assertAlmostEqual(inv_sltn.opt_val, result)
self.assertAlmostEqual(inv_sltn.primal_vars[self.a.id],
self.a.value)
self.assertAlmostEqual(inv_sltn.primal_vars[self.b.id],
self.b.value)
# Unbounded problems.
# TODO: Maximize
p = Problem(Minimize(-self.a), [self.a >= 2])
self.assertTrue(ConeMatrixStuffing().accepts(p))
try:
result = p.solve(solver.name())
except SolverError: # Gurobi fails on this one
return
p_new = ConeMatrixStuffing().apply(p)
self.assertTrue(solver.accepts(p_new[0]))
sltn = solver.solve(p_new[0], False, False, {})
self.assertAlmostEqual(sltn.opt_val, result)
inv_sltn = ConeMatrixStuffing().invert(sltn, p_new[1])
self.assertAlmostEqual(inv_sltn.opt_val, result)
# Infeasible problems.
p = Problem(Maximize(self.a), [self.a >= 2, self.a <= 1])
result = p.solve(solver.name())
self.assertTrue(FlipObjective().accepts(p))
p_min = FlipObjective().apply(p)
self.assertTrue(ConeMatrixStuffing().accepts(p_min[0]))
p_new = ConeMatrixStuffing().apply(p_min[0])
result_new = p_new[0].solve(solver.name())
self.assertAlmostEqual(result, -result_new)
self.assertTrue(solver.accepts(p_new[0]))
sltn = solver.solve(p_new[0], False, False, {})
self.assertAlmostEqual(sltn.opt_val, -result)
inv_sltn = ConeMatrixStuffing().invert(sltn, p_new[1])
self.assertAlmostEqual(inv_sltn.opt_val, -result)
inv_flipped_sltn = FlipObjective().invert(inv_sltn, p_min[1])
self.assertAlmostEqual(inv_flipped_sltn.opt_val, result)
def _reductions_for_problem_class(problem,
candidates,
gp: bool = False) -> List[Any]:
"""
Builds a chain that rewrites a problem into an intermediate
representation suitable for numeric reductions.
Parameters
----------
problem : Problem
The problem for which to build a chain.
candidates : dict
Dictionary of candidate solvers divided in qp_solvers
and conic_solvers.
gp : bool
If True, the problem is parsed as a Disciplined Geometric Program
instead of as a Disciplined Convex Program.
Returns
-------
list of Reduction objects
A list of reductions that can be used to convert the problem to an
intermediate form.
Raises
------
DCPError
Raised if the problem is not DCP and `gp` is False.
DGPError
Raised if the problem is not DGP and `gp` is True.
"""
reductions = []
# TODO Handle boolean constraints.
if complex2real.accepts(problem):
reductions += [complex2real.Complex2Real()]
if gp:
reductions += [Dgp2Dcp()]
if not gp and not problem.is_dcp():
append = build_non_disciplined_error_msg(problem, 'DCP')
if problem.is_dgp():
append += ("\nHowever, the problem does follow DGP rules. "
"Consider calling solve() with `gp=True`.")
elif problem.is_dqcp():
append += ("\nHowever, the problem does follow DQCP rules. "
"Consider calling solve() with `qcp=True`.")
raise DCPError("Problem does not follow DCP rules. Specifically:\n" +
append)
elif gp and not problem.is_dgp():
append = build_non_disciplined_error_msg(problem, 'DGP')
if problem.is_dcp():
append += ("\nHowever, the problem does follow DCP rules. "
"Consider calling solve() with `gp=False`.")
elif problem.is_dqcp():
append += ("\nHowever, the problem does follow DQCP rules. "
"Consider calling solve() with `qcp=True`.")
raise DGPError("Problem does not follow DGP rules." + append)
# Dcp2Cone and Qp2SymbolicQp require problems to minimize their objectives.
if type(problem.objective) == Maximize:
reductions += [FlipObjective()]
if _solve_as_qp(problem, candidates):
reductions += [CvxAttr2Constr(), qp2symbolic_qp.Qp2SymbolicQp()]
else:
# Canonicalize it to conic problem.
if not candidates['conic_solvers']:
raise SolverError("Problem could not be reduced to a QP, and no "
"conic solvers exist among candidate solvers "
"(%s)." % candidates)
else:
reductions += [Dcp2Cone(), CvxAttr2Constr()]
constr_types = {type(c) for c in problem.constraints}
if FiniteSet in constr_types:
reductions += [Valinvec2mixedint()]
return reductions
def _solve(self,
solver=None,
warm_start=True,
verbose=False,
parallel=False,
gp=False,
qcp=False,
**kwargs):
"""Solves a DCP compliant optimization problem.
Saves the values of primal and dual variables in the variable
and constraint objects, respectively.
Parameters
----------
solver : str, optional
The solver to use. Defaults to ECOS.
warm_start : bool, optional
Should the previous solver result be used to warm start?
verbose : bool, optional
Overrides the default of hiding solver output.
parallel : bool, optional
If problem is separable, solve in parallel.
gp : bool, optional
If True, parses the problem as a disciplined geometric program.
qcp : bool, optional
If True, parses the problem as a disciplined quasiconvex program.
kwargs : dict, optional
A dict of options that will be passed to the specific solver.
In general, these options will override any default settings
imposed by cvxpy.
Returns
-------
float
The optimal value for the problem, or a string indicating
why the problem could not be solved.
"""
if gp and qcp:
raise ValueError("At most one of `gp` and `qcp` can be True.")
if qcp and not self.is_dcp():
if not self.is_dqcp():
raise error.DQCPError("The problem is not DQCP.")
reductions = [dqcp2dcp.Dqcp2Dcp()]
if type(self.objective) == Maximize:
reductions = [FlipObjective()] + reductions
chain = Chain(problem=self, reductions=reductions)
soln = bisection.bisect(chain.reduce(),
solver=solver,
verbose=verbose,
**kwargs)
self.unpack(chain.retrieve(soln))
return self.value
if parallel:
from cvxpy.transforms.separable_problems import get_separable_problems
self._separable_problems = (get_separable_problems(self))
if len(self._separable_problems) > 1:
return self._parallel_solve(solver, warm_start, verbose,
**kwargs)
self._construct_chains(solver=solver, gp=gp)
data, solving_inverse_data = self._solving_chain.apply(
self._intermediate_problem)
solution = self._solving_chain.solve_via_data(self, data, warm_start,
verbose, kwargs)
full_chain = self._solving_chain.prepend(self._intermediate_chain)
inverse_data = self._intermediate_inverse_data + solving_inverse_data
self.unpack_results(solution, full_chain, inverse_data)
return self.value
