# fmt: off
from typing import IO, Optional, Union
import numpy as np
from ase import Atoms
from ase.optimize.optimize import Optimizer
from ase.utils.linesearch import LineSearch
[docs]
class LBFGS(Optimizer):
"""Limited memory BFGS optimizer.
A limited memory version of the bfgs algorithm. Unlike the bfgs algorithm
used in bfgs.py, the inverse of Hessian matrix is updated. The inverse
Hessian is represented only as a diagonal matrix to save memory
"""
def __init__(
self,
atoms: Atoms,
restart: Optional[str] = None,
logfile: Union[IO, str] = '-',
trajectory: Optional[str] = None,
maxstep: Optional[float] = None,
memory: int = 100,
damping: float = 1.0,
alpha: float = 70.0,
use_line_search: bool = False,
**kwargs,
):
"""
Parameters
----------
atoms: :class:`~ase.Atoms`
The Atoms object to relax.
restart: str
JSON file used to store vectors for updating the inverse of
Hessian matrix. If set, file with such a name will be searched
and information stored will be used, if the file exists.
logfile: file object or str
If *logfile* is a string, a file with that name will be opened.
Use '-' for stdout.
trajectory: string
Trajectory file used to store optimisation path.
maxstep: float
How far is a single atom allowed to move. This is useful for DFT
calculations where wavefunctions can be reused if steps are small.
Default is 0.2 Angstrom.
memory: int
Number of steps to be stored. Default value is 100. Three numpy
arrays of this length containing floats are stored.
damping: float
The calculated step is multiplied with this number before added to
the positions.
alpha: float
Initial guess for the Hessian (curvature of energy surface). A
conservative value of 70.0 is the default, but number of needed
steps to converge might be less if a lower value is used. However,
a lower value also means risk of instability.
kwargs : dict, optional
Extra arguments passed to
:class:`~ase.optimize.optimize.Optimizer`.
"""
super().__init__(atoms, restart, logfile, trajectory, **kwargs)
if maxstep is not None:
self.maxstep = maxstep
else:
self.maxstep = self.defaults['maxstep']
if self.maxstep > 1.0:
raise ValueError('You are using a much too large value for ' +
'the maximum step size: %.1f Angstrom' %
self.maxstep)
self.memory = memory
# Initial approximation of inverse Hessian 1./70. is to emulate the
# behaviour of BFGS. Note that this is never changed!
self.H0 = 1. / alpha
self.damping = damping
self.use_line_search = use_line_search
self.p = None
self.function_calls = 0
self.force_calls = 0
def initialize(self):
"""Initialize everything so no checks have to be done in step"""
self.iteration = 0
self.s = []
self.y = []
# Store also rho, to avoid calculating the dot product again and
# again.
self.rho = []
self.r0 = None
self.f0 = None
self.e0 = None
self.task = 'START'
self.load_restart = False
def read(self):
"""Load saved arrays to reconstruct the Hessian"""
self.iteration, self.s, self.y, self.rho, \
self.r0, self.f0, self.e0, self.task = self.load()
self.load_restart = True
def step(self, forces=None):
"""Take a single step
Use the given forces, update the history and calculate the next step --
then take it"""
forces = self._get_gradient(forces)
pos = self.optimizable.get_x()
self.update(pos, forces, self.r0, self.f0)
s = self.s
y = self.y
rho = self.rho
H0 = self.H0
loopmax = np.min([self.memory, self.iteration])
a = np.empty((loopmax,), dtype=np.float64)
# ## The algorithm itself:
q = -forces
for i in range(loopmax - 1, -1, -1):
a[i] = rho[i] * np.dot(s[i], q)
q -= a[i] * y[i]
z = H0 * q
for i in range(loopmax):
b = rho[i] * np.dot(y[i], z)
z += s[i] * (a[i] - b)
self.p = - z
# ##
g = -forces
if self.use_line_search:
e = self.func(pos)
self.line_search(pos, g, e)
dr = self.alpha_k * self.p
else:
self.force_calls += 1
self.function_calls += 1
dr = self.determine_step(self.p) * self.damping
self.optimizable.set_x(pos + dr)
self.iteration += 1
self.r0 = pos
self.f0 = -g
self.dump((self.iteration, self.s, self.y,
self.rho, self.r0, self.f0, self.e0, self.task))
def determine_step(self, dr):
"""Determine step to take according to maxstep
Normalize all steps as the largest step. This way
we still move along the eigendirection.
"""
longest_step = self.optimizable.gradient_norm(dr)
if longest_step >= self.maxstep:
dr *= self.maxstep / longest_step
return dr
def update(self, pos, forces, r0, f0):
"""Update everything that is kept in memory
This function is mostly here to allow for replay_trajectory.
"""
if self.iteration > 0:
s0 = pos - r0
self.s.append(s0)
# We use the gradient which is minus the force!
y0 = f0 - forces
self.y.append(y0)
rho0 = 1.0 / np.dot(y0, s0)
self.rho.append(rho0)
if self.iteration > self.memory:
self.s.pop(0)
self.y.pop(0)
self.rho.pop(0)
def replay_trajectory(self, traj):
"""Initialize history from old trajectory."""
if isinstance(traj, str):
from ase.io.trajectory import Trajectory
traj = Trajectory(traj, 'r')
r0 = None
f0 = None
# The last element is not added, as we get that for free when taking
# the first qn-step after the replay
for i in range(len(traj) - 1):
pos = traj[i].get_positions()
forces = traj[i].get_forces()
self.update(pos, forces, r0, f0)
r0 = pos.copy()
f0 = forces.copy()
self.iteration += 1
self.r0 = r0
self.f0 = f0
def func(self, x):
"""Objective function for use of the optimizers"""
self.optimizable.set_x(x)
self.function_calls += 1
return self.optimizable.get_value()
def fprime(self, x):
"""Gradient of the objective function for use of the optimizers"""
self.optimizable.set_x(x)
self.force_calls += 1
# Remember that forces are minus the gradient!
return -self.optimizable.get_gradient()
def line_search(self, r, g, e):
p_size = np.sqrt((self.p**2).sum())
if p_size <= np.sqrt(self.optimizable.ndofs() / 3 * 1e-10):
self.p /= (p_size / np.sqrt(self.optimizable.ndofs() / 3 * 1e-10))
ls = LineSearch(get_gradient_norm=self.optimizable.gradient_norm)
self.alpha_k, e, self.e0, self.no_update = \
ls._line_search(self.func, self.fprime, r, self.p, g, e, self.e0,
maxstep=self.maxstep, c1=.23,
c2=.46, stpmax=50.)
if self.alpha_k is None:
raise RuntimeError('LineSearch failed!')
[docs]
class LBFGSLineSearch(LBFGS):
"""This optimizer uses the LBFGS algorithm, but does a line search that
fulfills the Wolff conditions.
"""
def __init__(self, *args, **kwargs):
kwargs['use_line_search'] = True
super().__init__(*args, **kwargs)
