def test_infeasible(self):
for i in range(self.num_infeas):
data = genInfeasible(self.K, n=self.m // 3)
yield check_infeasible, scs.solve(data, self.K, **self.opts)
yield check_infeasible, scs.solve(
data, self.K, use_indirect=True, **self.opts)
def solveUnbounded():
K = {'f':10, 'l':15, 'q':[5, 10], 's':[3, 4], 'ep':10, 'ed':10}
m = getConeDims(K)
data = genUnbounded(K, n = m/3)
params = {'EPS':1e-4, 'NORMALIZE':1, 'SCALE':0.5, 'CG_RATE':2}
sol_i = scs.solve(data, K, params, USE_INDIRECT=True)
sol_d = scs.solve(data, K, params)
def solveFeasible():
# cone:
K = {
'f': 10,
'l': 15,
'q': [5, 10, 0, 1],
's': [3, 4, 0, 0, 1],
'ep': 10,
'ed': 10,
'p': [-0.25, 0.5, 0.75, -0.33]
}
m = getConeDims(K)
data, p_star = genFeasible(K, n=m // 3, density=0.01)
params = {'eps': 1e-3, 'normalize': True, 'scale': 5, 'cg_rate': 2}
sol_i = scs.solve(data, K, use_indirect=True, **params)
xi = sol_i['x']
yi = sol_i['y']
print('p* = ', p_star)
print('pri error = ', (dot(data['c'], xi) - p_star) / p_star)
print('dual error = ', (-dot(data['b'], yi) - p_star) / p_star)
# direct:
sol_d = scs.solve(data, K, use_indirect=False, **params)
xd = sol_d['x']
yd = sol_d['y']
print('p* = ', p_star)
print('pri error = ', (dot(data['c'], xd) - p_star) / p_star)
print('dual error = ', (-dot(data['b'], yd) - p_star) / p_star)
def solveInfeasible():
K = {'f':10, 'l':15, 'q':[5, 10, 0 ,1], 's':[3, 4, 0, 0, 1], 'ep':10, 'ed':10, 'p':[-0.25, 0.5, 0.75, -0.33]}
m = getConeDims(K)
data = genInfeasible(K, n = m//3)
params = {'eps':1e-4, 'normalize':True, 'scale':0.5, 'cg_rate':2}
sol_i = scs.solve(data, K, use_indirect=True, **params)
sol_d = scs.solve(data, K, use_indirect=False, **params)
def solveUnbounded():
K = {'f':10, 'l':15, 'q':[5, 10], 's':[3, 4], 'ep':10, 'ed':10}
m = getConeDims(K)
data = genUnbounded(K, n = m/3)
params = {'eps':1e-4, 'normalize':True, 'scale':0.5, 'cg_rate':2}
sol_i = scs.solve(data, K, use_indirect=True, **params)
sol_d = scs.solve(data, K, **params)
def solveUnbounded():
K = {"f": 10, "l": 15, "q": [5, 10, 0, 1], "s": [3, 4, 0, 0, 1], "ep": 10, "ed": 10, "p": [-0.25, 0.5, 0.75, -0.33]}
m = getConeDims(K)
data = genUnbounded(K, n=m / 3)
params = {"eps": 1e-4, "normalize": True, "scale": 0.5, "cg_rate": 2}
sol_i = scs.solve(data, K, use_indirect=True, **params)
sol_d = scs.solve(data, K, **params)
def test_unbounded(self):
for i in range(self.num_unb):
data = genUnbounded(self.K, n=self.m // 2)
yield check_unbounded, scs.solve(data, self.K, **self.opts)
yield check_unbounded, scs.solve(
data, self.K, use_indirect=True, **self.opts)
def test_feasible():
for i in range(num_feas):
data, p_star = genFeasible(K, n = m // 3, density = 0.1)
sol = scs.solve(data, K, **opts)
yield check_solution, dot(data['c'],sol['x']), p_star
yield check_solution, dot(-data['b'],sol['y']), p_star
sol = scs.solve(data, K, use_indirect=True, **opts)
yield check_solution, dot(data['c'],sol['x']), p_star
yield check_solution, dot(-data['b'],sol['y']), p_star
def test_feasible(self):
for i in range(self.num_feas):
data, p_star = genFeasible(self.K, n = self.m // 3, density = 0.1)
sol = scs.solve(data, self.K, **self.opts)
yield check_solution, dot(data['c'],sol['x']), p_star
yield check_solution, dot(-data['b'],sol['y']), p_star
sol = scs.solve(data, self.K, use_indirect=True, **self.opts)
yield check_solution, dot(data['c'],sol['x']), p_star
yield check_solution, dot(-data['b'],sol['y']), p_star
def test_feasible():
for i in range(num_feas):
data, p_star = tools.gen_feasible(K, n=m // 3, density=0.1)
sol = scs.solve(data, K, use_indirect=False, **opts)
yield check_solution, np.dot(data['c'], sol['x']), p_star
yield check_solution, np.dot(-data['b'], sol['y']), p_star
sol = scs.solve(data, K, use_indirect=True, **opts)
yield check_solution, np.dot(data['c'], sol['x']), p_star
yield check_solution, np.dot(-data['b'], sol['y']), p_star
def test_feasible():
for i in range(num_feas):
data, p_star = genFeasible(K, n = m/3, density = 0.01)
sol = scs.solve(data, K, opts)
yield check_solution, dot(data['c'],sol['x']), p_star
yield check_solution, dot(-data['b'],sol['y']), p_star
sol = scs.solve(data, K, opts=dict({'USE_INDIRECT':True}, **opts))
yield check_solution, dot(data['c'],sol['x']), p_star
yield check_solution, dot(-data['b'],sol['y']), p_star
def test_problems_with_longs():
new_cone = {'q': [], 'l': long(2)}
sol = scs.solve(data, new_cone)
yield check_solution, sol['x'][0], 1
sol = scs.solve(data, new_cone, use_indirect=True )
yield check_solution, sol['x'][0], 1
new_cone = {'q':[long(2)], 'l': 0}
sol = scs.solve(data, new_cone)
yield check_solution, sol['x'][0], 0.5
sol = scs.solve(data, new_cone, use_indirect=True )
yield check_solution, sol['x'][0], 0.5
def test_problems_with_longs():
new_cone = {'q': [], 'l': long(2)}
sol = scs.solve(data, new_cone, use_indirect=False)
yield check_solution, sol['x'][0], 1
sol = scs.solve(data, new_cone, use_indirect=True)
yield check_solution, sol['x'][0], 1
new_cone = {'q': [long(2)], 'l': 0}
sol = scs.solve(data, new_cone, use_indirect=False)
yield check_solution, sol['x'][0], 0.5
sol = scs.solve(data, new_cone, use_indirect=True)
yield check_solution, sol['x'][0], 0.5
def test_problems_with_longs():
new_cone = {'q': [], 'l': long(2)}
sol = scs.solve(data, new_cone)
yield check_solution, sol['x'][0], 1
sol = scs.solve(data, new_cone, opts={'USE_INDIRECT':True})
yield check_solution, sol['x'][0], 1
new_cone = {'q':[long(2)], 'l': 0}
sol = scs.solve(data, new_cone)
yield check_solution, sol['x'][0], 0.5
sol = scs.solve(data, new_cone, opts={'USE_INDIRECT':True})
yield check_solution, sol['x'][0], 0.5
def test_problems():
sol = scs.solve(data, cone)
yield check_solution, sol['x'][0], 1
new_cone = {'q':[2], 'l': 0}
sol = scs.solve(data, new_cone)
yield check_solution, sol['x'][0], 0.5
sol = scs.solve(data, cone, use_indirect = True )
yield check_solution, sol['x'][0], 1
sol = scs.solve(data, new_cone, use_indirect = True )
yield check_solution, sol['x'][0], 0.5
def test_problems():
sol = scs.solve(data, cone, use_indirect=False)
yield check_solution, sol['x'][0], 1
new_cone = {'q': [2], 'l': 0}
sol = scs.solve(data, new_cone, use_indirect=False)
yield check_solution, sol['x'][0], 0.5
sol = scs.solve(data, cone, use_indirect=True)
yield check_solution, sol['x'][0], 1
sol = scs.solve(data, new_cone, use_indirect=True)
yield check_solution, sol['x'][0], 0.5
def test_problems():
sol = scs.solve(data, cone)
yield check_solution, sol['x'][0], 1
new_cone = {'q':[2], 'l': 0}
sol = scs.solve(data, new_cone)
yield check_solution, sol['x'][0], 0.5
sol = scs.solve(data, cone, opts={'USE_INDIRECT':True})
yield check_solution, sol['x'][0], 1
sol = scs.solve(data, new_cone, opts={'USE_INDIRECT':True})
yield check_solution, sol['x'][0], 0.5
def solve(_solver_opts):
if scs_version.major < 3:
_results = scs.solve(args,
cones,
verbose=verbose,
**_solver_opts)
_status = self.STATUS_MAP[_results["info"]["statusVal"]]
else:
_results = scs.solve(args,
cones,
verbose=verbose,
**_solver_opts)
_status = self.STATUS_MAP[_results["info"]["status_val"]]
return _results, _status
def solve_via_data(self,
data,
warm_start,
verbose,
solver_opts,
solver_cache=None):
"""Returns the result of the call to the solver.
Parameters
----------
data : dict
Data generated via an apply call.
warm_start : Bool
Whether to warm_start SCS.
verbose : Bool
Control the verbosity.
solver_opts : dict
SCS-specific solver options.
Returns
-------
The result returned by a call to scs.solve().
"""
import scs
args = {"A": data[s.A], "b": data[s.B], "c": data[s.C]}
if warm_start and solver_cache is not None and \
self.name() in solver_cache:
args["x"] = solver_cache[self.name()]["x"]
args["y"] = solver_cache[self.name()]["y"]
args["s"] = solver_cache[self.name()]["s"]
cones = dims_to_solver_dict(data[ConicSolver.DIMS])
# Default to eps = 1e-4 instead of 1e-3.
solver_opts["eps"] = solver_opts.get("eps", 1e-4)
results = scs.solve(args, cones, verbose=verbose, **solver_opts)
status = self.STATUS_MAP[results["info"]["status"]]
if (status == s.OPTIMAL_INACCURATE
and "acceleration_lookback" not in solver_opts):
# anderson acceleration is sometimes unstable; retry without it
results = scs.solve(args,
cones,
verbose=verbose,
acceleration_lookback=0,
**solver_opts)
if solver_cache is not None:
solver_cache[self.name()] = results
return results
def solve_unbounded():
K = {
'f': 10,
'l': 15,
'q': [5, 10, 0, 1],
's': [3, 4, 0, 0, 1],
'ep': 10,
'ed': 10,
'p': [-0.25, 0.5, 0.75, -0.33]
}
m = get_scs_cone_dims(K)
data = gen_unbounded(K, n=m // 3)
params = {'normalize': True, 'scale': 0.5, 'cg_rate': 2}
sol_i = scs.solve(data, K, use_indirect=True, **params)
sol_d = scs.solve(data, K, use_indirect=False, **params)
def solveInfeasible():
K = {
'f': 10,
'l': 15,
'q': [5, 10, 0, 1],
's': [3, 4, 0, 0, 1],
'ep': 10,
'ed': 10,
'p': [-0.25, 0.5, 0.75, -0.33]
}
m = getConeDims(K)
data = genInfeasible(K, n=m // 3)
params = {'eps': 1e-4, 'normalize': True, 'scale': 0.5, 'cg_rate': 2}
sol_i = scs.solve(data, K, use_indirect=True, **params)
sol_d = scs.solve(data, K, use_indirect=False, **params)
def test_solve_infeasible(use_indirect, gpu):
data = tools.gen_infeasible(K, n=m // 3)
sol = scs.solve(data, K, use_indirect=use_indirect, gpu=gpu, **params)
y = sol['y']
np.testing.assert_array_less(np.linalg.norm(data['A'].T @ y), 1e-3)
np.testing.assert_array_less(data['b'].T @ y, -0.1)
np.testing.assert_almost_equal(y, tools.proj_dual_cone(y, K), decimal=4)
def solve_via_data(self, data, warm_start, verbose, solver_opts, solver_cache=None):
"""Returns the result of the call to the solver.
Parameters
----------
data : dict
Data generated via an apply call.
warm_start : Bool
Whether to warm_start SCS.
verbose : Bool
Control the verbosity.
solver_opts : dict
SCS-specific solver options.
Returns
-------
The result returned by a call to scs.solve().
"""
import scs
args = {"A": data[s.A], "b": data[s.B], "c": data[s.C]}
if warm_start and solver_cache is not None and \
self.name() in solver_cache:
args["x"] = solver_cache[self.name()]["x"]
args["y"] = solver_cache[self.name()]["y"]
args["s"] = solver_cache[self.name()]["s"]
cones = dims_to_solver_dict(data[ConicSolver.DIMS])
results = scs.solve(
args,
cones,
verbose=verbose,
**solver_opts)
if solver_cache is not None:
solver_cache[self.name()] = results
return results
def solve(self, objective, constraints, cached_data, verbose, solver_opts):
"""Returns the result of the call to the solver.
Parameters
----------
objective : LinOp
The canonicalized objective.
constraints : list
The list of canonicalized cosntraints.
cached_data : dict
A map of solver name to cached problem data.
verbose : bool
Should the solver print output?
solver_opts : dict
Additional arguments for the solver.
Returns
-------
tuple
(status, optimal value, primal, equality dual, inequality dual)
"""
(data, dims), obj_offset = self.get_problem_data(objective,
constraints,
cached_data)
# Set the options to be VERBOSE plus any user-specific options.
solver_opts["verbose"] = verbose
results_dict = scs.solve(data, dims, **solver_opts)
return self.format_results(results_dict, dims, obj_offset)
def test_unpack_results(self):
"""Test unpack results method.
"""
with self.assertRaises(Exception) as cm:
Problem(Minimize(exp(self.a))).unpack_results("blah", None)
self.assertEqual(str(cm.exception), "Unknown solver.")
prob = Problem(Minimize(exp(self.a)), [self.a == 0])
args = prob.get_problem_data(s.SCS)
results_dict = scs.solve(*args)
prob = Problem(Minimize(exp(self.a)), [self.a == 0])
prob.unpack_results(s.SCS, results_dict)
self.assertAlmostEqual(self.a.value, 0, places=4)
self.assertAlmostEqual(prob.value, 1, places=3)
self.assertAlmostEqual(prob.status, s.OPTIMAL)
prob = Problem(Minimize(norm(self.x)), [self.x == 0])
args = prob.get_problem_data(s.ECOS)
results_dict = ecos.solve(*args)
prob = Problem(Minimize(norm(self.x)), [self.x == 0])
prob.unpack_results(s.ECOS, results_dict)
self.assertItemsAlmostEqual(self.x.value, [0, 0])
self.assertAlmostEqual(prob.value, 0)
self.assertAlmostEqual(prob.status, s.OPTIMAL)
prob = Problem(Minimize(norm(self.x)), [self.x == 0])
args = prob.get_problem_data(s.CVXOPT)
results_dict = cvxopt.solvers.conelp(*args)
prob = Problem(Minimize(norm(self.x)), [self.x == 0])
prob.unpack_results(s.CVXOPT, results_dict)
self.assertItemsAlmostEqual(self.x.value, [0, 0])
self.assertAlmostEqual(prob.value, 0)
self.assertAlmostEqual(prob.status, s.OPTIMAL)
def test_unpack_results(self):
"""Test unpack results method.
"""
with self.assertRaises(Exception) as cm:
Problem(Minimize(exp(self.a))).unpack_results("blah", None)
self.assertEqual(str(cm.exception), "Unknown solver.")
prob = Problem(Minimize(exp(self.a)), [self.a == 0])
args = prob.get_problem_data(s.SCS)
results_dict = scs.solve(*args)
prob = Problem(Minimize(exp(self.a)), [self.a == 0])
prob.unpack_results(s.SCS, results_dict)
self.assertAlmostEqual(self.a.value, 0, places=4)
self.assertAlmostEqual(prob.value, 1, places=3)
self.assertAlmostEqual(prob.status, s.OPTIMAL)
prob = Problem(Minimize(norm(self.x)), [self.x == 0])
args = prob.get_problem_data(s.ECOS)
results_dict = ecos.solve(*args)
prob = Problem(Minimize(norm(self.x)), [self.x == 0])
prob.unpack_results(s.ECOS, results_dict)
self.assertItemsAlmostEqual(self.x.value, [0,0])
self.assertAlmostEqual(prob.value, 0)
self.assertAlmostEqual(prob.status, s.OPTIMAL)
prob = Problem(Minimize(norm(self.x)), [self.x == 0])
args = prob.get_problem_data(s.CVXOPT)
results_dict = cvxopt.solvers.conelp(*args)
prob = Problem(Minimize(norm(self.x)), [self.x == 0])
prob.unpack_results(s.CVXOPT, results_dict)
self.assertItemsAlmostEqual(self.x.value, [0,0])
self.assertAlmostEqual(prob.value, 0)
self.assertAlmostEqual(prob.status, s.OPTIMAL)
def test_feasible(use_indirect):
for i in range(num_feas):
data, p_star = tools.gen_feasible(K, n=m // 3, density=0.1)
sol = scs.solve(data, K, use_indirect=use_indirect, **opts)
assert_almost_equal(np.dot(data['c'], sol['x']), p_star, decimal=2)
assert_almost_equal(np.dot(-data['b'], sol['y']), p_star, decimal=2)
def test_failures():
yield assert_raises, TypeError, scs.solve
yield assert_raises, ValueError, scs.solve, data, {'q': [4], 'l': -2}
yield check_keyword, ValueError, 'max_iters', -1
yield check_keyword, TypeError, 'max_iters', 1.1
yield check_failure, scs.solve(data, {'q': [1], 'l': 0})
def test_failures():
yield assert_raises, TypeError, scs.solve
yield assert_raises, ValueError, scs.solve, data, {'q':[4], 'l':-2}
yield check_keyword, ValueError, 'max_iters', -1
yield check_keyword, TypeError, 'max_iters', 1.1
yield check_failure, scs.solve( data, {'q':[1], 'l': 0} )
def solveFeasible():
# cone:
K = {
'f': 10,
'l': 15,
'q': [5, 10, 0, 1],
's': [3, 4, 0, 0, 1],
'ep': 10,
'ed': 10
}
K = validateCone(K)
density = 0.01 # A matrix density
m = getConeDims(K)
n = round(m / 3)
params = {'EPS': 1e-3, 'NORMALIZE': 1, 'SCALE': 5, 'CG_RATE': 1.5}
z = randn(m, )
z = symmetrizeSDP(z, K) # for SD cones
y = proj_dual_cone(z, K) # y = s - z;
s = y - z # s = proj_cone(z,K)
A = sparse.rand(m, n, density, format='csc')
A.data = randn(A.nnz)
x = randn(n)
c = -transpose(A).dot(y)
b = A.dot(x) + s
data = {'A': A, 'b': b, 'c': c}
# indirect
sol_i = scs.solve(data, K, params, USE_INDIRECT=True)
xi = sol_i['x']
yi = sol_i['y']
print 'c\'x* = ', dot(c, x)
print '% error = ', (dot(c, xi) - dot(c, x)) / dot(c, x)
print 'b\'y* = ', dot(b, y)
print '% error = ', (dot(b, yi) - dot(b, y)) / dot(b, y)
# direct:
sol_d = scs.solve(data, K, params)
xd = sol_d['x']
yd = sol_d['y']
print 'c\'x* = ', dot(c, x)
print '% error = ', (dot(c, xd) - dot(c, x)) / dot(c, x)
print 'b\'y* = ', dot(b, y)
print '% error = ', (dot(b, yd) - dot(b, y)) / dot(b, y)
def test_solve_unbounded(use_indirect, gpu):
data = tools.gen_unbounded(K, n=m // 3)
sol = scs.solve(data, K, use_indirect=use_indirect, gpu=gpu, **params)
x = sol['x']
s = sol['s']
np.testing.assert_array_less(np.linalg.norm(data['A'] @ x + s), 1e-3)
np.testing.assert_array_less(data['c'].T @ x, -0.1)
np.testing.assert_almost_equal(s, tools.proj_cone(s, K), decimal=4)
def test_solve_feasible(use_indirect, gpu):
data, p_star = tools.gen_feasible(K, n=m // 3, density=0.01)
sol = scs.solve(data, K, use_indirect=use_indirect, gpu=gpu, **params)
x = sol['x']
y = sol['y']
print('p* = ', p_star)
print('pri error = ', (np.dot(data['c'], x) - p_star) / p_star)
print('dual error = ', (-np.dot(data['b'], y) - p_star) / p_star)
def test_failures():
yield assert_raises, TypeError, scs.solve
yield assert_raises, ValueError, scs.solve, data, {'q': [4], 'l': -2}
yield check_keyword, ValueError, 'max_iters', -1
# python 2.6 and before just cast float to int
if platform.python_version_tuple() >= ('2', '7', '0'):
yield check_keyword, TypeError, 'max_iters', 1.1
yield check_failure, scs.solve(data, {'q': [1], 'l': 0})
def test_failures():
yield assert_raises, TypeError, scs.solve
yield assert_raises, ValueError, scs.solve, data, {'q':[4], 'l':-2}
yield check_keyword, ValueError, 'max_iters', -1
# python 2.6 and before just cast float to int
if platform.python_version_tuple() >= ('2', '7', '0'):
yield check_keyword, TypeError, 'max_iters', 1.1
yield check_failure, scs.solve( data, {'q':[1], 'l': 0} )
def sdp(omega, fmin, warm_start=True):
'''
Solves the following SDP with the first-order primal-dual solver SCS:
---------Primal Form----------
minimize <M, \omega>
s.t. M - C_i = positive semidefinite for all i = 0...k
Dual Form:
---------Dual Form------------
minimize \sum_{i=1}^{k} <Y_i, C_i> - fmin
s.t. Y_i positive semidefinite for all i = 0...k
\sum_{i=0}^{k} Y_i = \omega
------------------------------
in order to get the value and the gradient of the acquisiton function.
Inputs:
omega: Second order moment matrix
fmin: min value achieved so far, i.e. min(y0)
warm_start: Whether or not to warm-start the solution
Outpus:
opt_val: Optimal value of the SDP
M: Optimizer of the SDP
Y: List of the Optimal Lagrange Multipliers (Dual Optimizers) of the SDP
C: List of C_i, i = 0 ... k
'''
omega = (omega + omega.T)/2
# Express the problem in the format required by scs
data = create_scs_data(omega, fmin)
k_ = omega.shape[0]
cone = {'s': [k_]*k_}
if not 'past_omegas' in globals() or len(past_omegas) == 0 or \
past_omegas[0].shape[0] != k_ or warm_start == False:
# Clear the saved solutions, as they are of different size
# than the one we are currently trying to solve.
reset_warm_starting()
else:
# Update data with warm-started solution
data = get_warm_start(omega, data)
# Call SCS
sol = scs.solve(data, cone, eps=1e-5, use_indirect=False, verbose=OUTPUT_LEVEL==1)
# print(sol['info']['solveTime']) # Prints solution time for SDP.
if sol['info']['status'] != 'Solved':
logging.getLogger('opt').warning(
'SCS solution status:' + sol['info']['status']
)
# Extract solution from SCS' structures
M, Y = unpack_solution(sol['x'], sol['y'], k_)
objective = -sol['info']['pobj']
sol['C'] = data['C']
if warm_start:
# Save solution for warm starting
past_solutions.append(sol); past_omegas.append(omega)
return objective, M, Y, data['C']
def _solve_relaxation(A, B, eps=1e-9, max_iters=2500, verbose=False):
"""Given the linear system formed by the problem's geometric constraints,
computes all possible poses.
Arguments:
A -- the matrix defining the homogeneous linear system formed from the problem's
geometric constraints, such that A r = 0.
B -- an auxiliary matrix which allows to retrieve the translation vector
from an optimal rotation t = - B @ r.
eps -- numerical precision of the solver
max_iters -- maximum number of iterations the solver is allowed to perform
verbose -- print additional information to the console
"""
# Solve the QCQP using shor's relaxation
# Construct Q
Q = np.block([[A.T @ A, np.zeros((9, 1))], [np.zeros((1, 9)), 0]])
# Invoke solver
results = scs.solve(
{"A": _A, "b": _b, "c": _vech10(Q, 2)}, # data
{"f": 22, "l": 0, "q": [], "ep": 0, "s": [10]}, # cones
verbose=verbose,
eps=eps,
max_iters=max_iters,
)
# Invoke solver
Z = _vech10_inv(results["x"])
if np.any(np.isnan(Z)):
if verbose:
warnings.warn(
"The SDP solver did not return a valid solution. Increasing max_iters might solve the issue."
)
return [(np.full((3, 3), np.nan), np.full(3, np.nan))]
vals, vecs = np.linalg.eigh(Z)
# check for rank
rank = np.sum(vals > 1e-3)
r_c = None
if rank == 1:
r_c = (vecs[:-1, -1] / vecs[-1, -1])[None, :]
else:
r_c = _constraint_ortho_det(vecs, rank)
# project to valid rotation spaces
U, _, Vh = np.linalg.svd(r_c.reshape(-1, 3, 3))
R = U @ Vh
r = R.reshape(-1, 9)
t = -r @ B.T
# Optimality guarantees
res = r @ A.T
if np.any(np.abs(np.sum(res * res, axis=-1) - results["info"]["dobj"]) > eps):
not_certifiable = "The solution is not certifiably optimal."
warnings.warn(not_certifiable)
return list(zip(R.transpose(0, 2, 1), t))
def solveInfeasible():
K = {'f':10, 'l':15, 'q':[5, 10], 's':[3, 4], 'ep':10, 'ed':10}
K = validateCone(K)
m = getConeDims(K)
n = round(m / 3)
params = {'EPS':1e-4, 'NORMALIZE':1, 'SCALE':0.5, 'CG_RATE':1.5}
z = randn(m,)
z = symmetrizeSDP(z, K) # for SD cones
y = proj_dual_cone(z, K) # y = s - z;
A = randn(m, n)
A = A - outer(y, transpose(A).dot(y)) / linalg.norm(y) ** 2 # dense...
b = randn(m);
b = -b / dot(b, y);
data = {'A':sparse.csc_matrix(A), 'b':b, 'c':randn(n)}
sol_i = scs.solve(data, K, params, USE_INDIRECT=True)
sol_d = scs.solve(data, K, params)
def solveFeasible():
# cone:
K = {"f": 10, "l": 15, "q": [5, 10, 0, 1], "s": [3, 4, 0, 0, 1], "ep": 10, "ed": 10, "p": [-0.25, 0.5, 0.75, -0.33]}
m = getConeDims(K)
data, p_star = genFeasible(K, n=m / 3, density=0.01)
params = {"eps": 1e-3, "normalize": True, "scale": 5, "cg_rate": 2}
sol_i = scs.solve(data, K, use_indirect=True, **params)
xi = sol_i["x"]
yi = sol_i["y"]
print("p* = ", p_star)
print("pri error = ", (dot(data["c"], xi) - p_star) / p_star)
print("dual error = ", (-dot(data["b"], yi) - p_star) / p_star)
# direct:
sol_d = scs.solve(data, K, **params)
xd = sol_d["x"]
yd = sol_d["y"]
print("p* = ", p_star)
print("pri error = ", (dot(data["c"], xd) - p_star) / p_star)
print("dual error = ", (-dot(data["b"], yd) - p_star) / p_star)
def solveInfeasible():
K = {'f': 10, 'l': 15, 'q': [5, 10], 's': [3, 4], 'ep': 10, 'ed': 10}
K = validateCone(K)
m = getConeDims(K)
n = round(m / 3)
params = {'EPS': 1e-4, 'NORMALIZE': 1, 'SCALE': 0.5, 'CG_RATE': 1.5}
z = randn(m, )
z = symmetrizeSDP(z, K) # for SD cones
y = proj_dual_cone(z, K) # y = s - z;
A = randn(m, n)
A = A - outer(y, transpose(A).dot(y)) / linalg.norm(y)**2 # dense...
b = randn(m)
b = -b / dot(b, y)
data = {'A': sparse.csc_matrix(A), 'b': b, 'c': randn(n)}
sol_i = scs.solve(data, K, params, USE_INDIRECT=True)
sol_d = scs.solve(data, K, params)
def solveUnbounded():
K = {'f':10, 'l':15, 'q':[5, 10], 's':[3, 4], 'ep':10, 'ed':10}
K = validateCone(K)
m = getConeDims(K)
n = round(m / 3)
params = {'EPS':1e-4, 'NORMALIZE':1, 'SCALE':0.5, 'CG_RATE':1.5}
z = randn(m);
z = symmetrizeSDP(z, K); # for SD cones
s = proj_cone(z, K);
A = randn(m, n);
x = randn(n);
A = A - outer(s + A.dot(x), x) / linalg.norm(x) ** 2; # dense...
c = randn(n);
c = -c / dot(c, x);
data = {'A':sparse.csc_matrix(A), 'b':randn(m), 'c':c}
sol_i = scs.solve(data, K, params, USE_INDIRECT=True)
sol_d = scs.solve(data, K, params)
def solveFeasible():
# cone:
K = {'f':10, 'l':15, 'q':[5, 10, 0 ,1], 's':[3, 4, 0, 0, 1], 'ep':10, 'ed':10}
m = getConeDims(K)
data, p_star = genFeasible(K, n = m/3, density = 0.01)
params = {'eps':1e-3, 'normalize':True, 'scale':5, 'cg_rate':2}
sol_i = scs.solve(data, K, use_indirect=True, **params)
xi = sol_i['x']
yi = sol_i['y']
print('p* = ', p_star)
print('pri error = ', (dot(data['c'], xi) - p_star) / p_star)
print('dual error = ', (-dot(data['b'], yi) - p_star) / p_star)
# direct:
sol_d = scs.solve(data, K, **params)
xd = sol_d['x']
yd = sol_d['y']
print('p* = ', p_star)
print('pri error = ', (dot(data['c'], xd) - p_star) / p_star)
print('dual error = ', (-dot(data['b'], yd) - p_star) / p_star)
def solveFeasible():
# cone:
K = {'f':10, 'l':15, 'q':[5, 10, 0 ,1], 's':[3, 4, 0, 0, 1], 'ep':10, 'ed':10}
m = getConeDims(K)
data, p_star = genFeasible(K, n = m/3, density = 0.01)
params = {'EPS':1e-3, 'NORMALIZE':1, 'SCALE':5, 'CG_RATE':2}
sol_i = scs.solve(data, K, params, USE_INDIRECT=True)
xi = sol_i['x']
yi = sol_i['y']
print('p* = ', p_star)
print('pri % error = ', (dot(data['c'], xi) - p_star) / p_star)
print('dual % error = ', (-dot(data['b'], yi) - p_star) / p_star)
# direct:
sol_d = scs.solve(data, K, params)
xd = sol_d['x']
yd = sol_d['y']
print('p* = ', p_star)
print('pri % error = ', (dot(data['c'], xd) - p_star) / p_star)
print('dual % error = ', (-dot(data['b'], yd) - p_star) / p_star)
def solveFeasible():
# cone:
K = {'f':10, 'l':50000, 'q':[5, 10, 0 ,1], 's':[3, 4, 0, 0, 1], 'ep':10, 'ed':10, 'p':[-0.25, 0.5, 0.75, -0.33]}
m = getConeDims(K)
data, p_star = genFeasible(K, n = m/3, density = 0.01)
params = {'eps':1e-3, 'normalize':True, 'scale':5, 'cg_rate':2}
sol_i = scs.solve(data, K, gpu=True, **params)
xi = sol_i['x']
yi = sol_i['y']
print('p* = ', p_star)
print('pri error = ', (dot(data['c'], xi) - p_star) / p_star)
print('dual error = ', (-dot(data['b'], yi) - p_star) / p_star)
def solveUnbounded():
K = {'f': 10, 'l': 15, 'q': [5, 10], 's': [3, 4], 'ep': 10, 'ed': 10}
K = validateCone(K)
m = getConeDims(K)
n = round(m / 3)
params = {'EPS': 1e-4, 'NORMALIZE': 1, 'SCALE': 0.5, 'CG_RATE': 1.5}
z = randn(m)
z = symmetrizeSDP(z, K)
# for SD cones
s = proj_cone(z, K)
A = randn(m, n)
x = randn(n)
A = A - outer(s + A.dot(x), x) / linalg.norm(x)**2
# dense...
c = randn(n)
c = -c / dot(c, x)
data = {'A': sparse.csc_matrix(A), 'b': randn(m), 'c': c}
sol_i = scs.solve(data, K, params, USE_INDIRECT=True)
sol_d = scs.solve(data, K, params)
def solveFeasible():
# cone:
K = {'f':10, 'l':15, 'q':[5, 10, 0 ,1], 's':[3, 4, 0, 0, 1], 'ep':10, 'ed':10}
K = validateCone(K)
density = 0.01 # A matrix density
m = getConeDims(K)
n = round(m / 3)
params = {'EPS':1e-3, 'NORMALIZE':1, 'SCALE':5, 'CG_RATE':1.5}
z = randn(m,)
z = symmetrizeSDP(z, K) # for SD cones
y = proj_dual_cone(z, K) # y = s - z;
s = y - z # s = proj_cone(z,K)
A = sparse.rand(m, n, density, format='csc')
A.data = randn(A.nnz)
x = randn(n)
c = -transpose(A).dot(y)
b = A.dot(x) + s
data = {'A': A, 'b': b, 'c': c}
# indirect
sol_i = scs.solve(data, K, params, USE_INDIRECT=True)
xi = sol_i['x']
yi = sol_i['y']
print 'c\'x* = ', dot(c, x)
print '% error = ', (dot(c, xi) - dot(c, x)) / dot(c, x)
print 'b\'y* = ', dot(b, y)
print '% error = ', (dot(b, yi) - dot(b, y)) / dot(b, y)
# direct:
sol_d = scs.solve(data, K, params)
xd = sol_d['x']
yd = sol_d['y']
print 'c\'x* = ', dot(c, x)
print '% error = ', (dot(c, xd) - dot(c, x)) / dot(c, x)
print 'b\'y* = ', dot(b, y)
print '% error = ', (dot(b, yd) - dot(b, y)) / dot(b, y)
def solve_unbounded(use_indirect, gpu):
K = {
'f': 10,
'l': 15,
'q': [5, 10, 0, 1],
's': [3, 4, 0, 0, 1],
'ep': 10,
'ed': 10,
'p': [-0.25, 0.5, 0.75, -0.33]
}
m = tools.get_scs_cone_dims(K)
data = tools.gen_unbounded(K, n=m // 3)
params = {'normalize': True, 'scale': 0.5, 'cg_rate': 2}
sol = scs.solve(data, K, use_indirect=use_indirect, gpu=gpu, **params)
def solve_infeasible():
K = {
'f': 10,
'l': 15,
'q': [5, 10, 0, 1],
's': [3, 4, 0, 0, 1],
'ep': 10,
'ed': 10,
'p': [-0.25, 0.5, 0.75, -0.33]
}
m = get_scs_cone_dims(K)
data = gen_infeasible(K, n=m // 3)
params = {'eps': 1e-4, 'normalize': True, 'scale': 0.5, 'cg_rate': 2}
sol_i = scs.solve(data, K, gpu=True, **params)
def solve_unbounded(use_indirect, gpu):
K = {
"f": 10,
"l": 15,
"q": [5, 10, 0, 1],
"s": [3, 4, 0, 0, 1],
"ep": 10,
"ed": 10,
"p": [-0.25, 0.5, 0.75, -0.33],
}
m = tools.get_scs_cone_dims(K)
data = tools.gen_unbounded(K, n=m // 3)
params = {"normalize": True, "scale": 0.5}
sol = scs.solve(data, K, use_indirect=use_indirect, gpu=gpu, **params)
def solveUnbounded():
K = {
'f': 10,
'l': 15,
'q': [5, 10, 0, 1],
's': [3, 4, 0, 0, 1],
'ep': 10,
'ed': 10,
'p': [-0.25, 0.5, 0.75, -0.33]
}
m = getConeDims(K)
data = genUnbounded(K, n=m // 3)
params = {'eps': 1e-4, 'normalize': True, 'scale': 0.5, 'cg_rate': 2}
sol_i = scs.solve(data, K, gpu=True, **params)
def do_prox(data, indmap, solmap, x0_vals, rho):
# don't modify original dict
x0_vals = dict(x0_vals)
# set tau in x0_vals
x0_vals['__tau'] = rho/2.0
restuff(data, indmap, x0_vals)
out = scs.solve(data, data['dims'], verbose=False)
scs_x = out['x']
x_vals = extract_sol(scs_x, solmap)
return x_vals
def scs_solve(A, b, c, dim_dict, init_z=None, **kwargs):
"""Wraps scs.solve for convenience."""
scs_cones = {
'l': dim_dict['l'] if 'l' in dim_dict else 0,
'q': dim_dict['q'] if 'q' in dim_dict else [],
's': dim_dict['s'] if 's' in dim_dict else [],
'ep': dim_dict['ep'] if 'ep' in dim_dict else 0,
'ed': dim_dict['ed'] if 'ed' in dim_dict else 0,
'f': dim_dict['z'] if 'z' in dim_dict else 0
}
#print('scs_cones', scs_cones)
sol = scs.solve({'A': A, 'b': b, 'c': c}, cone=scs_cones, **kwargs)
info = sol['info']
if info['statusVal'] > 0:
z = xsy2z(sol['x'], sol['s'], sol['y'], tau=1., kappa=0.)
if info['statusVal'] < 0:
x = np.zeros_like(sol['x']) \
if np.any(np.isnan(sol['x'])) else sol['x']
s = np.zeros_like(sol['s']) \
if np.any(np.isnan(sol['s'])) else sol['s']
y = np.zeros_like(sol['y']) \
if np.any(np.isnan(sol['y'])) else sol['y']
if np.allclose(y, 0.) and c @ x < 0:
obj = c @ x
# assert obj < 0
x /= -obj
s /= -obj
# print('primal res:', np.linalg.norm(A@x + s))
if np.allclose(s, 0.) and b @ y < 0:
obj = b @ y
# assert obj < 0
y /= -obj
# print('dual res:', np.linalg.norm(A.T@y))
# print('SCS NONSOLVED')
# print('x', x)
# print('s', s)
# print('y', y)
z = xsy2z(x, s, y, tau=0., kappa=1.)
return z, info
def _scs_solve(self, objective, constr_map, dims,
var_offsets, x_length,
verbose, opts):
"""Calls the SCS solver and returns the result.
Parameters
----------
objective: LinExpr
The canonicalized objective.
constr_map: dict
A dict of the canonicalized constraints.
dims: dict
A dict with information about the types of constraints.
var_offsets: dict
A dict mapping variable id to offset in the stacked variable x.
x_length: int
The height of x.
verbose: bool
Should the solver show output?
opts: dict
A dict of the solver parameters passed to scs
Returns
-------
tuple
(status, optimal objective, optimal x,
optimal equality constraint dual,
optimal inequality constraint dual)
"""
prob_data = self._scs_problem_data(objective, constr_map, dims,
var_offsets, x_length)
obj_offset = prob_data[1]
# Set the options to be VERBOSE plus any user-specific options.
opts = {k.upper():v for k, v in opts.items()}
opts["VERBOSE"] = verbose
use_indirect = opts.get("USE_INDIRECT", False)
results = scs.solve(*prob_data[0], opts=opts, USE_INDIRECT = use_indirect)
status = s.SOLVER_STATUS[s.SCS][results["info"]["status"]]
if status in s.SOLUTION_PRESENT:
primal_val = results["info"]["pobj"]
value = self.objective._primal_to_result(primal_val - obj_offset)
eq_dual = results["y"][0:dims["f"]]
ineq_dual = results["y"][dims["f"]:]
return (status, value, results["x"], eq_dual, ineq_dual)
else:
return (status, None, None, None, None)
def time_scs(mapper, **kwargs):
""" Map `scs.solve` over the global `data` and return timing results
Maps with `mapper`, which may be a parallel map, such as
`concurrent.futures.ThreadPoolExecutor.map`
Pass `kwargs` onto `scs.solve` to, for example, toggle verbose output
"""
ts = []
for _ in range(repeat):
start = time.time()
# `mapper` will usually return a generator instantly
# need to consume the entire generator to find the actual compute time
# calling `list` consumes the generator. an empty `for` loop would also work
a = list(mapper(lambda x: scs.solve(*x, **kwargs), data))
end = time.time()
ts.append(end-start)
return min(ts)
def solve(self, objective, constraints, cached_data,
warm_start, verbose, solver_opts):
"""Returns the result of the call to the solver.
Parameters
----------
objective : LinOp
The canonicalized objective.
constraints : list
The list of canonicalized cosntraints.
cached_data : dict
A map of solver name to cached problem data.
warm_start : bool
Should the previous solver result be used to warm_start?
verbose : bool
Should the solver print output?
solver_opts : dict
Additional arguments for the solver.
Returns
-------
tuple
(status, optimal value, primal, equality dual, inequality dual)
"""
import scs
data = self.get_problem_data(objective,
constraints,
cached_data)
# Set the options to be VERBOSE plus any user-specific options.
solver_opts["verbose"] = verbose
scs_args = {"c": data[s.C], "A": data[s.A], "b": data[s.B]}
# If warm_starting, add old primal and dual variables.
solver_cache = cached_data[self.name()]
if warm_start and solver_cache.prev_result is not None:
scs_args["x"] = solver_cache.prev_result["x"]
scs_args["y"] = solver_cache.prev_result["y"]
scs_args["s"] = solver_cache.prev_result["s"]
results_dict = scs.solve(scs_args, data[s.DIMS], **solver_opts)
return self.format_results(results_dict, data[s.DIMS],
data[s.OFFSET], cached_data)
def solve(self, objective, constraints, cached_data, verbose, solver_opts):
"""Returns the result of the call to the solver.
Parameters
----------
objective : LinOp
The canonicalized objective.
constraints : list
The list of canonicalized cosntraints.
cached_data : dict
A map of solver name to cached problem data.
verbose : bool
Should the solver print output?
solver_opts : dict
Additional arguments for the solver.
Returns
-------
tuple
(status, optimal value, primal, equality dual, inequality dual)
"""
(data, dims), obj_offset = self.get_problem_data(objective,
constraints,
cached_data)
# Set the options to be VERBOSE plus any user-specific options.
solver_opts["verbose"] = verbose
results = scs.solve(data, dims, **solver_opts)
status = s.SOLVER_STATUS[s.SCS][results["info"]["status"]]
if status in s.SOLUTION_PRESENT:
primal_val = results["info"]["pobj"]
value = primal_val + obj_offset
eq_dual = results["y"][0:dims["f"]]
ineq_dual = results["y"][dims["f"]:]
return (status, value, results["x"], eq_dual, ineq_dual)
else:
return (status, None, None, None, None)
def solveUnbounded():
K = {'f':10, 'l':15, 'q':[5, 10, 0 ,1], 's':[3, 4, 0, 0, 1], 'ep':10, 'ed':10, 'p':[-0.25, 0.5, 0.75, -0.33]}
m = getConeDims(K)
data = genUnbounded(K, n = m/3)
params = {'eps':1e-4, 'normalize':True, 'scale':0.5, 'cg_rate':2}
sol_i = scs.solve(data, K, gpu=True, **params)
def test_unbounded():
for i in range(num_unb):
data = genUnbounded(K, n = m // 2)
yield check_unbounded, scs.solve(data, K, **opts)
yield check_unbounded, scs.solve(data, K, use_indirect=True, **opts)
