def test_obj_scaling(self):
model = AmplModel(self.model_name)
log = config_logger("nlp.der",
"%(name)-10s %(levelname)-8s %(message)s",
level=logging.DEBUG)
dcheck = DerivativeChecker(model, model.x0, tol=1e-5)
dcheck.check(hess=True, chess=True)
assert len(dcheck.grad_errs) == 0
assert len(dcheck.hess_errs) == 0
for j in xrange(model.ncon):
assert (len(dcheck.chess_errs[j]) == 0)
model.compute_scaling_obj(g_max=1.)
assert model.obj(model.x0) == -0.39056208756589972
assert np.allclose(model.grad(model.x0), np.array([0.8, -1.]))
assert model.scale_obj == 1.
model.compute_scaling_obj(reset=True)
assert model.scale_obj is None
model.compute_scaling_obj(g_max=0.5)
assert model.obj(model.x0) == -0.19528104378294986
assert np.allclose(model.grad(model.x0), np.array([0.4, -0.5]))
assert np.allclose(
model.sgrad(model.x0).to_array(), np.array([0.4, -0.5]))
assert model.scale_obj == 0.5
dcheck = DerivativeChecker(model, model.x0, tol=1e-5)
dcheck.check(hess=True, chess=True)
assert len(dcheck.grad_errs) == 0
assert len(dcheck.hess_errs) == 0
for j in xrange(model.ncon):
assert (len(dcheck.chess_errs[j]) == 0)
def test_derivatives(self):
log = config_logger("nlp.der",
"%(name)-10s %(levelname)-8s %(message)s",
level=logging.DEBUG)
dcheck = DerivativeChecker(self.model, self.x)
dcheck.check(chess=False)
assert len(dcheck.grad_errs) == 0
assert len(dcheck.hess_errs) == 0
def test_cons_scaling(self):
model = self.model
log = config_logger("nlp.der",
"%(name)-10s %(levelname)-8s %(message)s",
level=logging.DEBUG)
dcheck = DerivativeChecker(model, model.x0, tol=1e-5)
dcheck.check(hess=True, chess=True)
assert len(dcheck.jac_errs) == 0
assert len(dcheck.hess_errs) == 0
for j in xrange(model.ncon):
assert (len(dcheck.chess_errs[j]) == 0)
model.compute_scaling_cons(g_max=2.)
assert np.allclose(model.scale_con, np.array([1.]))
cons = np.array([0.])
for j in range(model.ncon):
assert np.allclose(model.icons(j, model.x0), cons[j])
assert np.allclose(model.cons(model.x0), cons)
assert np.allclose(
model.jop(model.x0).to_array(), np.array([[1., 2.]]))
assert np.allclose(
ndarray_from_coord(model.ncon,
model.nvar,
*model.A(),
symmetric=False), np.array([[1., 2.]]))
model.compute_scaling_cons(reset=True)
assert model.scale_con is None
model.compute_scaling_cons(g_max=1.)
assert np.allclose(model.scale_con, np.array([0.5]))
assert np.allclose(model.cons(model.x0), np.array([0.]))
assert np.allclose(
model.jop(model.x0).to_array(), np.array([[0.5, 1.]]))
assert np.allclose(
ndarray_from_coord(model.ncon,
model.nvar,
*model.A(),
symmetric=False), np.array([[0.5, 1.]]))
assert np.allclose(model.igrad(0, model.x0), np.array([0.5, 1.]))
assert np.allclose(
model.sigrad(0, model.x0).to_array(), np.array([0.5, 1.]))
v = np.ones(model.nvar)
assert np.allclose(model.hiprod(model.x0, 0, v), np.zeros(model.n))
dcheck = DerivativeChecker(model, model.x0, tol=1e-5)
dcheck.check(hess=True, chess=True)
assert len(dcheck.jac_errs) == 0
assert len(dcheck.hess_errs) == 0
for j in xrange(model.ncon):
assert (len(dcheck.chess_errs[j]) == 0)
print model.display_basic_info()
def test_cons_scaling(self):
model = AmplModel(self.model_name)
log = config_logger("nlp.der",
"%(name)-10s %(levelname)-8s %(message)s",
level=logging.DEBUG)
dcheck = DerivativeChecker(model, model.x0, tol=1e-5)
dcheck.check(hess=True, chess=True)
assert len(dcheck.jac_errs) == 0
assert len(dcheck.hess_errs) == 0
for j in xrange(model.ncon):
assert (len(dcheck.chess_errs[j]) == 0)
model.compute_scaling_cons(g_max=40.)
assert np.allclose(model.scale_con, np.array([1.]))
assert np.allclose(model.cons(model.x0), np.array([29.0]))
assert np.allclose(
model.jop(model.x0).to_array(), np.array([[40., 4.]]))
model.compute_scaling_cons(reset=True)
assert model.scale_con is None
model.compute_scaling_cons(g_max=20.)
assert np.allclose(model.cons(model.x0), np.array([14.5]))
assert np.allclose(model.scale_con, np.array([0.5]))
assert np.allclose(
model.jop(model.x0).to_array(), np.array([[20., 2.]]))
assert np.allclose(model.igrad(0, model.x0), np.array([20., 2.]))
assert np.allclose(
model.sigrad(0, model.x0).to_array(), np.array([20., 2.]))
v = np.ones(model.nvar)
assert np.allclose(model.hiprod(model.x0, 0, v), np.array([26., 1.]))
dcheck = DerivativeChecker(model, model.x0, tol=1e-5)
dcheck.check(hess=True, chess=True)
assert len(dcheck.jac_errs) == 0
assert len(dcheck.hess_errs) == 0
for j in xrange(model.ncon):
assert (len(dcheck.chess_errs[j]) == 0)
print model.display_basic_info()
def test_obj_scaling(self):
model = self.model
model.x0 = np.zeros(model.nvar)
log = config_logger("nlp.der",
"%(name)-10s %(levelname)-8s %(message)s",
level=logging.DEBUG)
dcheck = DerivativeChecker(model, model.x0, tol=1e-5)
dcheck.check(hess=True, chess=True)
assert len(dcheck.grad_errs) == 0
assert len(dcheck.hess_errs) == 0
for j in xrange(model.ncon):
assert (len(dcheck.chess_errs[j]) == 0)
model.compute_scaling_obj(g_max=1.)
assert model.scale_obj == 1.
assert model.obj(model.x0) == 1.
assert np.allclose(model.grad(model.x0), np.array([1., 0.]))
print model.cost()
assert np.allclose(model.cost().to_array(), np.array([1., 0.]))
model.compute_scaling_obj(reset=True)
assert model.scale_obj is None
model.compute_scaling_obj(g_max=0.5)
assert model.scale_obj == 0.5
assert model.obj(model.x0) == 0.5
assert np.allclose(model.grad(model.x0), np.array([0.5, 0.]))
assert np.allclose(model.cost().to_array(), np.array([0.5, 0.]))
assert np.allclose(
model.sgrad(model.x0).to_array(), np.array([0.5, 0.]))
dcheck = DerivativeChecker(model, model.x0, tol=1e-5)
dcheck.check(hess=True, chess=True)
assert len(dcheck.grad_errs) == 0
assert len(dcheck.hess_errs) == 0
for j in xrange(model.ncon):
assert (len(dcheck.chess_errs[j]) == 0)
from nlp.optimize.funnel import QNFunnel as Funnel
from pykrylov.linop import CompactLSR1Operator as QNOperator
opts["H"] = QNOperator
opts["npairs"] = args.npairs
opts["scaling"] = True
else:
from nlp.model.pysparsemodel import PySparseAmplModel as Model
from nlp.optimize.funnel import Funnel
nprobs = len(other)
if nprobs == 0:
raise ValueError("Please supply problem name as argument")
# Create root logger.
logger = config_logger("nlp", "%(name)-9s %(levelname)-5s %(message)s")
# Create Funnel logger.
funnel_logger = config_logger(
"funnel",
"%(name)-9s %(levelname)-5s %(message)s",
level=logging.WARN if nprobs > 1 else logging.DEBUG)
qn_logger = config_logger("qn",
"%(name)-9s %(levelname)-5s %(message)s",
level=logging.WARN if nprobs > 1 else logging.DEBUG)
if nprobs > 1:
logger.info("%12s %5s %5s %8s %8s %8s %6s %6s %6s %7s", "name", "nvar",
"iter", "f", u"‖c‖", u"‖g+Jᵀy‖", "#f", u"#∇f", "stat", "time")
for problem in other:
except ImportError:
import sys
print "Pyomo is not installed."
sys.exit(0)
from nlp.model.amplmodel import AmplModel
from nlp.optimize.tron import TRON
from nlp.optimize.pcg import TruncatedCG
from nlp.tools.logs import config_logger
from nlp.tools.utils import evaluate_model_methods_at_starting_point
import logging
import numpy as np
tron_logger = config_logger("nlp.tron",
"%(name)-8s %(levelname)-5s %(message)s",
level=logging.INFO)
if __name__ == "__main__":
# Create a Pyomo Rosenbrock model
pyomo_model = ConcreteModel()
pyomo_model.x = Var()
pyomo_model.y = Var(bounds=(-1.5, None))
pyomo_model.o = Objective(expr=(pyomo_model.x - 1)**2 + \
100 * (pyomo_model.y - pyomo_model.x**2)**2)
pyomo_model.x.set_value(-2.)
pyomo_model.y.set_value(1.)
# Writes a Pyomo model in NL file format."""
nl_filename = "rosenbrock.nl"
help="maximum number of iterations")
# Parse command-line arguments.
(args, other) = parser.parse_known_args()
if args.armijo:
from nlp.optimize.lbfgs import LBFGS
else:
from nlp.optimize.lbfgs import WolfeLBFGS as LBFGS
nprobs = len(other)
if nprobs == 0:
raise ValueError("Please supply problem name as argument")
# Create root logger.
logger = config_logger("nlp", "%(name)-3s %(levelname)-5s %(message)s")
# Create LBFGS logger.
slv_log = config_logger("nlp.lbfgs",
"%(name)-9s %(levelname)-5s %(message)s",
level=logging.WARN if nprobs > 1 else logging.INFO)
logger.info("%10s %5s %6s %8s %8s %6s %6s %5s %7s", "name", "nvar", "iter",
"f", u"‖∇f‖", "#f", u"#∇f", "stat", "time")
for problem in other:
model = QNAmplModel(problem,
H=InverseLBFGSOperator,
npairs=args.npairs,
scaling=True)
model.compute_scaling_obj()
# Argument parsing, take list of names from the command line
parser = argparse.ArgumentParser()
parser.add_argument("name_list", nargs='+', help="list of SIF files to process")
parser.add_argument("--qn_type", type=str, default="bfgs", choices=["bfgs","sr1"],
help="the type of quasi-Newton approximation to be used")
parser.add_argument("--approx_type", type=str, default=None, choices=[None,"full","struct"],
help="choose exact Hessian (None), full approximate Hessian (full), or structured approximation (struct)")
args = parser.parse_args()
nprobs = len(args.name_list)
# Create root logger.
# logger = config_logger("nlp", "%(name)-3s %(levelname)-5s %(message)s")
logger = config_logger("nlp", "%(name)-3s %(levelname)-5s %(message)s",
stream=None,
filename="cutest_rough.txt", filemode="a")
# Create Auglag logger.
auglag_logger = config_logger("nlp.auglag",
"%(name)-8s %(levelname)-5s %(message)s",
level=logging.WARN if nprobs > 1 else logging.INFO)
# Create TRON logger.
tron_logger = config_logger("nlp.tron",
"%(name)-8s %(levelname)-5s %(message)s",
level=logging.WARN if nprobs > 1 else logging.INFO)
if nprobs > 1:
logger.info("%9s %5s %5s %6s %8s %8s %6s %6s %5s %7s",
"name", "nvar", "ncon", "iter", "f", u"‖P∇L‖", "#f", u"#∇f", "stat",
#!/usr/bin/env python
"""Simple AMPL driver for the derivative checker."""
from nlp.model.pysparsemodel import PySparseAmplModel
from nlp.tools.dercheck import DerivativeChecker
from nlp.tools.logs import config_logger
import sys
if len(sys.argv) == 1:
raise ValueError("Please supply problem name as argument")
# Create root logger.
log = config_logger("nlp.der", "%(name)-10s %(levelname)-8s %(message)s")
nlp = PySparseAmplModel(sys.argv[1])
dcheck = DerivativeChecker(nlp, nlp.x0)
dcheck.check()
pg = auglag.pgnorm
ts = auglag.tsolve
else:
it = -auglag.iter
fc, gc = -auglag.model.obj.ncalls, -auglag.model.grad.ncalls
pg = -1.0 if auglag.pgnorm is None else -auglag.pgnorm
ts = -1.0 if auglag.tsolve is None else -auglag.tsolve
return (it, fc, gc, pg, ts)
nprobs = len(sys.argv) - 1
if nprobs == 0:
raise ValueError("Please supply problem name as argument")
# Create root logger.
logger = config_logger("nlp", "%(name)-3s %(levelname)-5s %(message)s")
# Create Auglag logger.
auglag_logger = config_logger("nlp.auglag",
"%(name)-8s %(levelname)-5s %(message)s",
level=logging.WARN if nprobs > 1 else logging.INFO)
# Create TRON logger.
tron_logger = config_logger("nlp.tron",
"%(name)-8s %(levelname)-5s %(message)s",
level=logging.WARN if nprobs > 1 else logging.INFO)
if nprobs > 1:
logger.info("%12s %5s %6s %8s %8s %6s %6s %5s %7s",
"name", "nvar", "iter", "f", u"‖P∇f‖", "#f", u"#∇f", "stat",
"time")
pg = cqp.pgnorm
ts = cqp.tsolve
else:
it = -cqp.iter
fc, gc = -cqp.model.obj.ncalls, -cqp.model.grad.ncalls
pg = -1.0 if cqp.pgnorm is None else -cqp.pgnorm
ts = -1.0 if cqp.tsolve is None else -cqp.tsolve
return (it, fc, gc, pg, ts)
nprobs = len(sys.argv) - 1
if nprobs == 0:
raise ValueError("Please supply problem name as argument")
# Create root logger.
logger = config_logger("nlp", "%(name)-3s %(levelname)-5s %(message)s")
# Create CQP logger.
cqp_logger = config_logger("nlp.cqp",
"%(name)-8s %(levelname)-5s %(message)s",
level=logging.WARN if nprobs > 1 else logging.INFO)
if nprobs > 1:
logger.info("%12s %5s %6s %8s %8s %6s %6s %5s %7s", "name", "nvar", "iter",
"f", u"‖P∇f‖", "#f", u"#∇f", "stat", "time")
for problem in sys.argv[1:]:
model = PySparseAmplModel(problem)
model.compute_scaling_obj()
# Check for inequality- or equality-constrained problem.
