def uncons_opt(obj, cons, Vars, x0):
"""nlinprog
Parameters
----------
obj :
cons :
Returns
-------
Notes
------
"""
raise NotImplementedError
assert(obj == 0)
for c in cons:
assert(isinstance(c, sym.LessThan))
assert(c.rhs == 0)
obj += c.lhs**2
eval_f = ft.partial(eval_expr, sym.lambdify(Vars, obj))
eval_grad_f = ft.partial(eval_grad_obj, sym.lambdify(Vars, grad(Vars, obj)))
eval_hessian_f = ft.partial(eval_expr, sym.lambdify(Vars, sym.hessian(obj, Vars)))
# Return codes in IpReturnCodes_inc.h
res_x, mL, mU, Lambda, res_obj, status = pyipopt.fmin_unconstrained(
eval_f,
x0,
fprime=eval_grad_f,
fhess=eval_hessian_f,
)
return spec.OPTRES(res_obj, res_x, 'OK', res_obj <= 0)
def _find_assignment(variable_handler, atoms, max_num_resets, tol, verbose=False):
init = np.array(variable_handler.dict_to_vector())
def func(vector):
print "dim:", np.shape(vector), vector
d = variable_handler.vector_to_dict(vector)
return sum(evaluate(atom, d).norm for atom in atoms)
def grad(f, theta):
theta = algopy.UTPM.init_jacobian(theta)
return algopy.UTPM.extract_jacobian(f(theta))
def hess(f, theta):
theta = algopy.UTPM.init_hessian(theta)
return algopy.UTPM.extract_hessian(len(theta), f(theta))
x = None
for i in range(max_num_resets):
results = pyipopt.fmin_unconstrained(func, init, functools.partial(grad, func))
if verbose:
print("iteration %d:" % (i+1))
print(results)
val, zl, zu, constraint_multipliers, obj, status = results
if obj < tol:
x = val
break
init = np.random.rand(len(init))
if x is None:
return None
return variable_handler.vector_to_dict(x)
def main():
pyipopt.set_loglevel(2)
x0 = numpy.array([-1.2, 1], dtype=float)
results = pyipopt.fmin_unconstrained(
scipy.optimize.rosen,
x0,
fprime=scipy.optimize.rosen_der,
fhess=scipy.optimize.rosen_hess,
)
print(results)
def main():
pyipopt.set_loglevel(2)
x0 = numpy.array([-0.27, -0.9], dtype=float)
results = pyipopt.fmin_unconstrained(
himmelblau,
x0,
fprime=functools.partial(eval_grad, himmelblau),
fhess=functools.partial(eval_hess, himmelblau),
)
print results
def main():
pyipopt.set_loglevel(2)
x0 = numpy.array([-3, -1, -3, -1], dtype=float)
results = pyipopt.fmin_unconstrained(
wood,
x0,
fprime=functools.partial(eval_grad, wood),
fhess=functools.partial(eval_hess, wood),
)
print results
def main():
pyipopt.set_loglevel(2)
x0 = numpy.array([-3, -1, -3, -1], dtype=float)
results = pyipopt.fmin_unconstrained(
wood,
x0,
fprime=functools.partial(eval_grad, wood),
fhess=functools.partial(eval_hess, wood),
)
print(results)
def main():
pyipopt.set_loglevel(2)
x0 = numpy.array([-1.2, 1], dtype=float)
results = pyipopt.fmin_unconstrained(
scipy.optimize.rosen,
x0,
fprime=scipy.optimize.rosen_der,
fhess=scipy.optimize.rosen_hess,
)
print results
def solve_unconstr(theta0):
pyipopt.set_loglevel(1)
thetahat , _, _, _, _, fval = pyipopt.fmin_unconstrained(
eval_f,
theta0,
fprime=eval_grad,
fhess=eval_hess,
)
return thetahat
def _find_assignment(variable_handler,
atoms,
max_num_resets,
tol,
verbose=False):
init = np.array(variable_handler.dict_to_vector())
def func(vector):
print "dim:", np.shape(vector), vector
d = variable_handler.vector_to_dict(vector)
return sum(evaluate(atom, d).norm for atom in atoms)
def grad(f, theta):
theta = algopy.UTPM.init_jacobian(theta)
return algopy.UTPM.extract_jacobian(f(theta))
def hess(f, theta):
theta = algopy.UTPM.init_hessian(theta)
return algopy.UTPM.extract_hessian(len(theta), f(theta))
x = None
for i in range(max_num_resets):
results = pyipopt.fmin_unconstrained(func, init,
functools.partial(grad, func))
if verbose:
print("iteration %d:" % (i + 1))
print(results)
val, zl, zu, constraint_multipliers, obj, status = results
if obj < tol:
x = val
break
init = np.random.rand(len(init))
if x is None:
return None
return variable_handler.vector_to_dict(x)
def do_opt(args, f, theta):
"""
@param args: directly parsed from the command line
@param f: function to minimize
@param theta: initial guess of parameter values
"""
#FIXME: remove fmin_args
#@param fmin_args: data and other precomputed things independent of theta
fmin_args = tuple()
g = functools.partial(eval_grad, f)
h = functools.partial(eval_hess, f)
if args.fmin == 'simplex':
results = scipy.optimize.fmin(
f,
theta,
args=fmin_args,
maxfun=10000,
maxiter=10000,
xtol=1e-8,
ftol=1e-8,
full_output=True,
)
elif args.fmin == 'bfgs':
results = scipy.optimize.fmin_bfgs(
f,
theta,
args=fmin_args,
#fprime=g,
#epsilon=1e-7,
maxiter=10000,
full_output=True,
disp=True,
retall=True,
)
elif args.fmin == 'jeffopt':
results = jeffopt.fmin_jeff_unconstrained(
f,
theta,
args=fmin_args,
#abstol=1e-8,
)
elif args.fmin == 'ncg':
results = scipy.optimize.fmin_ncg(
f,
theta,
args=fmin_args,
fprime=g,
fhess=h,
avextol=1e-6,
maxiter=10000,
full_output=True,
disp=True,
retall=True,
)
elif args.fmin == 'slsqp':
results = scipy.optimize.minimize(
f,
theta,
args=fmin_args,
method='SLSQP',
jac=g,
)
elif args.fmin == 'powell':
results = scipy.optimize.minimize(
f,
theta,
args=fmin_args,
method='Powell',
)
elif args.fmin == 'cg':
results = scipy.optimize.minimize(
f,
theta,
args=fmin_args,
method='CG',
jac=g,
)
elif args.fmin == 'anneal':
results = scipy.optimize.minimize(
f,
theta,
args=fmin_args,
method='Anneal',
)
elif args.fmin == 'ipopt':
results = pyipopt.fmin_unconstrained(
f,
theta,
fprime=g,
fhess=h,
)
else:
raise Exception
print 'results:', results
xopt = results[0]
print 'optimal solution vector:', xopt
print 'exp optimal solution vector:', numpy.exp(xopt)
print
def do_opt(args, f, theta):
"""
@param args: directly parsed from the command line
@param f: function to minimize
@param theta: initial guess of parameter values
"""
#FIXME: remove fmin_args
#@param fmin_args: data and other precomputed things independent of theta
fmin_args = tuple()
g = functools.partial(eval_grad, f)
h = functools.partial(eval_hess, f)
if args.fmin == 'simplex':
results = scipy.optimize.fmin(
f,
theta,
args=fmin_args,
maxfun=10000,
maxiter=10000,
xtol=1e-8,
ftol=1e-8,
full_output=True,
)
elif args.fmin == 'bfgs':
results = scipy.optimize.fmin_bfgs(
f,
theta,
args=fmin_args,
#fprime=g,
#epsilon=1e-7,
maxiter=10000,
full_output=True,
disp=True,
retall=True,
)
elif args.fmin == 'jeffopt':
results = jeffopt.fmin_jeff_unconstrained(
f,
theta,
args=fmin_args,
#abstol=1e-8,
)
elif args.fmin == 'ncg':
results = scipy.optimize.fmin_ncg(
f,
theta,
args=fmin_args,
fprime=g,
fhess=h,
avextol=1e-6,
maxiter=10000,
full_output=True,
disp=True,
retall=True,
)
elif args.fmin == 'slsqp':
results = scipy.optimize.minimize(
f,
theta,
args=fmin_args,
method='SLSQP',
jac=g,
)
elif args.fmin == 'powell':
results = scipy.optimize.minimize(
f,
theta,
args=fmin_args,
method='Powell',
)
elif args.fmin == 'cg':
results = scipy.optimize.minimize(
f,
theta,
args=fmin_args,
method='CG',
jac=g,
)
elif args.fmin == 'anneal':
results = scipy.optimize.minimize(
f,
theta,
args=fmin_args,
method='Anneal',
)
elif args.fmin == 'ipopt':
results = pyipopt.fmin_unconstrained(
f,
theta,
fprime=g,
fhess=h,
)
else:
raise Exception
print 'results:', results
xopt = results[0]
print 'optimal solution vector:', xopt
print 'exp optimal solution vector:', numpy.exp(xopt)
print
multistart = 1000
np.random.seed(1234)
r = np.random.rand(multistart, len(thetalower))
r = r[multistartidx,:]
print '** invoke ipopt'
start = time()
pyipopt.set_loglevel(1)
thetahatnew , _, _, _, _, fval = pyipopt.fmin_unconstrained(
eval_f,
thetalower*r + thetaupper*(1-r),
fprime=eval_grad,
fhess=eval_hess,
)
print '** calculating nlogl and se'
nllnew = eval_f(thetahatnew)
try:
sehatnew = np.diag(np.linalg.inv(eval_hess(thetahatnew)))**0.5
except Exception, e:
sehatnew = np.nan*np.ones(thetahatnew.shape)
print '-- Time =', time()-start, 's'
outname = 'result'+ '_' + specname +'_' "{:03d}".format(int(multistartidx))
def main():
pyipopt.set_loglevel(2)
x0 = numpy.array([-0.27, -0.9], dtype=float)
results = pyipopt.fmin_unconstrained(
himmelblau,
x0,
fprime=functools.partial(eval_grad, himmelblau),
fhess=functools.partial(eval_hess, himmelblau),
)
print(results)
