def __init__(self, coords, potential, eigenvec0=None,
rotational_steps=20,
translational_steps=10,
maxiter=500,
leig_kwargs=None,
translator_kwargs=None,
dimer=True,
):
coords = coords.copy()
self.rotational_steps = rotational_steps
self.translational_steps = translational_steps
self.maxiter = maxiter
self.iter_number = 0
# check the keyword dictionaries
if translator_kwargs is None: translator_kwargs = {}
if leig_kwargs is None: leig_kwargs = {}
# set up the initial guess for the eigenvector
if eigenvec0 is None:
eigenvec0 = rotations.vec_random_ndim(coords.shape)
eigenvec0 /= np.linalg.norm(eigenvec0)
assert coords.shape == eigenvec0.shape
# set up the object that will maintain the rotation of the dimer
self.rotator = FindLowestEigenVector(coords, potential, eigenvec0=eigenvec0, **leig_kwargs)
# set up the object that will translate the dimer
if dimer:
self.translator = _DimerTranslator(coords, potential, eigenvec0, **translator_kwargs)
else:
self.translator = _HybridEigenvectorWalker(coords, potential, eigenvec0, **translator_kwargs)
def test(self):
x = self.system.get_random_configuration()
evec = vec_random_ndim(x.size)
evec /= np.linalg.norm(evec)
opt = _DimerTranslator(x, self.pot, evec)
ret = opt.run(10)
self.assertEqual(ret.nfev, self.pot.nfev)
self.assertGreater(ret.nfev, 0)
def test(self):
x = self.system.get_random_configuration()
evec = vec_random_ndim(x.size)
evec /= np.linalg.norm(evec)
opt = _DimerTranslator(x, self.pot, evec)
ret = opt.run(10)
self.assertEqual(ret.nfev, self.pot.nfev)
self.assertGreater(ret.nfev, 0)
def _minimizeTangentSpace(self, coords, energy=None, gradient=None):
"""minimize the energy in the space perpendicular to eigenvec.
Parameters
----------
coords : the current position
energy, gradient : the energy and gradient at the current position
"""
assert gradient is not None
if self._transverse_walker is None:
if self.invert_gradient:
# note: if we pass transverse energy and gradient here we can save 1 potential call
self._transverse_walker = _DimerTranslator(
coords, self.pot, self.eigenvec,
**self.tangent_space_quench_params)
else:
self._transverse_walker = _TransverseWalker(
coords, self.pot, self.eigenvec, energy, gradient,
**self.tangent_space_quench_params)
else:
self._transverse_walker.update_eigenvec(self.eigenvec,
self.eigenval)
self._transverse_walker.update_coords(coords, energy, gradient)
# determine the number of steps
# i.e. if the eigenvector is deemed to have converged or is changing slowly
eigenvec_converged = self.overlap > .999
if eigenvec_converged:
nstepsperp = self.nsteps_tangent2
else:
nstepsperp = self.nsteps_tangent1
# reduce the maximum step size if necessary
maxstep = self.maxstep_tangent
if self.reduce_step > 0:
maxstep *= self.step_factor**self.reduce_step
self._transverse_walker.update_maxstep(maxstep)
coords_old = coords.copy()
ret = self._transverse_walker.run(nstepsperp)
coords = ret.coords
self.tangent_move_step = np.linalg.norm(coords - coords_old)
self.tangent_result = ret
if self.tangent_move_step > 1e-16:
try:
self.energy, self.gradient = self._transverse_walker.get_true_energy_gradient(
coords)
except AttributeError:
raise Exception(
"was tspot was never called? use the same gradient")
return ret
def _minimizeTangentSpace(self, coords, energy=None, gradient=None):
"""minimize the energy in the space perpendicular to eigenvec.
Parameters
----------
coords : the current position
energy, gradient : the energy and gradient at the current position
"""
assert gradient is not None
if self._transverse_walker is None:
if self.invert_gradient:
# note: if we pass transverse energy and gradient here we can save 1 potential call
self._transverse_walker = _DimerTranslator(coords, self.pot, self.eigenvec,
# energy=transverse_energy, gradient=transverse_gradient,
**self.tangent_space_quench_params)
else:
self._transverse_walker = _TransverseWalker(coords, self.pot, self.eigenvec, energy, gradient,
**self.tangent_space_quench_params)
else:
self._transverse_walker.update_eigenvec(self.eigenvec, self.eigenval)
self._transverse_walker.update_coords(coords, energy, gradient)
#determine the number of steps
#i.e. if the eigenvector is deemed to have converged
eigenvec_converged = self.overlap > .999
nstepsperp = self.nsteps_tangent1
if eigenvec_converged:
nstepsperp = self.nsteps_tangent2
# reduce the maximum step size if necessary
maxstep = self.maxstep_tangent
if self.reduce_step > 0:
maxstep *= (self.step_factor)**self.reduce_step
self._transverse_walker.update_maxstep(maxstep)
coords_old = coords.copy()
ret = self._transverse_walker.run(nstepsperp)
coords = ret.coords
self.tangent_move_step = np.linalg.norm(coords - coords_old)
self.tangent_result = ret
if self.tangent_move_step > 1e-16:
try:
self.energy, self.gradient = self._transverse_walker.get_true_energy_gradient(coords)
except AttributeError:
print "was tspot was never called? use the same gradient"
raise
return ret
