root / ase / optimize / lbfgs.py @ 5
Historique | Voir | Annoter | Télécharger (10,85 ko)
| 1 |
# -*- coding: utf-8 -*-
|
|---|---|
| 2 |
import sys |
| 3 |
import numpy as np |
| 4 |
from ase.optimize.optimize import Optimizer |
| 5 |
from ase.utils.linesearch import LineSearch |
| 6 |
|
| 7 |
class LBFGS(Optimizer): |
| 8 |
"""Limited memory BFGS optimizer.
|
| 9 |
|
| 10 |
A limited memory version of the bfgs algorithm. Unlike the bfgs algorithm
|
| 11 |
used in bfgs.py, the inverse of Hessian matrix is updated. The inverse
|
| 12 |
Hessian is represented only as a diagonal matrix to save memory
|
| 13 |
|
| 14 |
"""
|
| 15 |
def __init__(self, atoms, restart=None, logfile='-', trajectory=None, |
| 16 |
maxstep=None, memory=100, damping = 1.0, alpha = 10.0, |
| 17 |
use_line_search=False):
|
| 18 |
"""
|
| 19 |
Parameters:
|
| 20 |
|
| 21 |
restart: string
|
| 22 |
Pickle file used to store vectors for updating the inverse of Hessian
|
| 23 |
matrix. If set, file with such a name will be searched and information
|
| 24 |
stored will be used, if the file exists.
|
| 25 |
|
| 26 |
logfile: string
|
| 27 |
Where should output go. None for no output, '-' for stdout.
|
| 28 |
|
| 29 |
trajectory: string
|
| 30 |
Pickle file used to store trajectory of atomic movement.
|
| 31 |
|
| 32 |
maxstep: float
|
| 33 |
How far is a single atom allowed to move. This is useful for DFT
|
| 34 |
calculations where wavefunctions can be reused if steps are small.
|
| 35 |
Default is 0.04 Angstrom.
|
| 36 |
|
| 37 |
memory: int
|
| 38 |
Number of steps to be stored. Default value is 100. Three numpy
|
| 39 |
arrays of this length containing floats are stored.
|
| 40 |
|
| 41 |
damping: float
|
| 42 |
The calculated step is multiplied with this number before added to
|
| 43 |
the positions.
|
| 44 |
|
| 45 |
alpha: float
|
| 46 |
Initial guess for the Hessian (curvature of energy surface). A
|
| 47 |
conservative value of 70.0 is the default, but number of needed
|
| 48 |
steps to converge might be less if a lower value is used. However,
|
| 49 |
a lower value also means risk of instability.
|
| 50 |
|
| 51 |
"""
|
| 52 |
Optimizer.__init__(self, atoms, restart, logfile, trajectory)
|
| 53 |
|
| 54 |
if maxstep is not None: |
| 55 |
if maxstep > 1.0: |
| 56 |
raise ValueError('You are using a much too large value for ' + |
| 57 |
'the maximum step size: %.1f Angstrom' % maxstep)
|
| 58 |
self.maxstep = maxstep
|
| 59 |
else:
|
| 60 |
self.maxstep = 0.04 |
| 61 |
|
| 62 |
self.memory = memory
|
| 63 |
self.H0 = 1. / alpha # Initial approximation of inverse Hessian |
| 64 |
# 1./70. is to emulate the behaviour of BFGS
|
| 65 |
# Note that this is never changed!
|
| 66 |
self.damping = damping
|
| 67 |
self.use_line_search = use_line_search
|
| 68 |
self.p = None |
| 69 |
self.function_calls = 0 |
| 70 |
self.force_calls = 0 |
| 71 |
|
| 72 |
def initialize(self): |
| 73 |
"""Initalize everything so no checks have to be done in step"""
|
| 74 |
self.iteration = 0 |
| 75 |
self.s = []
|
| 76 |
self.y = []
|
| 77 |
self.rho = [] # Store also rho, to avoid calculationg the dot product |
| 78 |
# again and again
|
| 79 |
|
| 80 |
self.r0 = None |
| 81 |
self.f0 = None |
| 82 |
self.e0 = None |
| 83 |
self.task = 'START' |
| 84 |
self.load_restart = False |
| 85 |
|
| 86 |
def read(self): |
| 87 |
"""Load saved arrays to reconstruct the Hessian"""
|
| 88 |
self.iteration, self.s, self.y, self.rho, \ |
| 89 |
self.r0, self.f0, self.e0, self.task = self.load() |
| 90 |
self.load_restart = True |
| 91 |
|
| 92 |
def step(self, f): |
| 93 |
"""Take a single step
|
| 94 |
|
| 95 |
Use the given forces, update the history and calculate the next step --
|
| 96 |
then take it"""
|
| 97 |
r = self.atoms.get_positions()
|
| 98 |
p0 = self.p
|
| 99 |
|
| 100 |
self.update(r, f, self.r0, self.f0) |
| 101 |
|
| 102 |
s = self.s
|
| 103 |
y = self.y
|
| 104 |
rho = self.rho
|
| 105 |
H0 = self.H0
|
| 106 |
|
| 107 |
loopmax = np.min([self.memory, self.iteration]) |
| 108 |
a = np.empty((loopmax,), dtype=np.float64) |
| 109 |
|
| 110 |
### The algorithm itself:
|
| 111 |
q = - f.reshape(-1)
|
| 112 |
for i in range(loopmax - 1, -1, -1): |
| 113 |
a[i] = rho[i] * np.dot(s[i], q) |
| 114 |
q -= a[i] * y[i] |
| 115 |
z = H0 * q |
| 116 |
|
| 117 |
for i in range(loopmax): |
| 118 |
b = rho[i] * np.dot(y[i], z) |
| 119 |
z += s[i] * (a[i] - b) |
| 120 |
|
| 121 |
self.p = - z.reshape((-1, 3)) |
| 122 |
###
|
| 123 |
|
| 124 |
g = -f |
| 125 |
if self.use_line_search == True: |
| 126 |
e = self.func(r)
|
| 127 |
self.line_search(r, g, e)
|
| 128 |
dr = (self.alpha_k * self.p).reshape(len(self.atoms),-1) |
| 129 |
else:
|
| 130 |
self.force_calls += 1 |
| 131 |
self.function_calls += 1 |
| 132 |
dr = self.determine_step(self.p) * self.damping |
| 133 |
self.atoms.set_positions(r+dr)
|
| 134 |
|
| 135 |
self.iteration += 1 |
| 136 |
self.r0 = r
|
| 137 |
self.f0 = -g
|
| 138 |
self.dump((self.iteration, self.s, self.y, |
| 139 |
self.rho, self.r0, self.f0, self.e0, self.task)) |
| 140 |
|
| 141 |
def determine_step(self, dr): |
| 142 |
"""Determine step to take according to maxstep
|
| 143 |
|
| 144 |
Normalize all steps as the largest step. This way
|
| 145 |
we still move along the eigendirection.
|
| 146 |
"""
|
| 147 |
steplengths = (dr**2).sum(1)**0.5 |
| 148 |
longest_step = np.max(steplengths) |
| 149 |
if longest_step >= self.maxstep: |
| 150 |
dr *= self.maxstep / longest_step
|
| 151 |
|
| 152 |
return dr
|
| 153 |
|
| 154 |
def update(self, r, f, r0, f0): |
| 155 |
"""Update everything that is kept in memory
|
| 156 |
|
| 157 |
This function is mostly here to allow for replay_trajectory.
|
| 158 |
"""
|
| 159 |
if self.iteration > 0: |
| 160 |
s0 = r.reshape(-1) - r0.reshape(-1) |
| 161 |
self.s.append(s0)
|
| 162 |
|
| 163 |
# We use the gradient which is minus the force!
|
| 164 |
y0 = f0.reshape(-1) - f.reshape(-1) |
| 165 |
self.y.append(y0)
|
| 166 |
|
| 167 |
rho0 = 1.0 / np.dot(y0, s0)
|
| 168 |
self.rho.append(rho0)
|
| 169 |
|
| 170 |
if self.iteration > self.memory: |
| 171 |
self.s.pop(0) |
| 172 |
self.y.pop(0) |
| 173 |
self.rho.pop(0) |
| 174 |
|
| 175 |
|
| 176 |
def replay_trajectory(self, traj): |
| 177 |
"""Initialize history from old trajectory."""
|
| 178 |
if isinstance(traj, str): |
| 179 |
from ase.io.trajectory import PickleTrajectory |
| 180 |
traj = PickleTrajectory(traj, 'r')
|
| 181 |
r0 = None
|
| 182 |
f0 = None
|
| 183 |
# The last element is not added, as we get that for free when taking
|
| 184 |
# the first qn-step after the replay
|
| 185 |
for i in range(0, len(traj) - 1): |
| 186 |
r = traj[i].get_positions() |
| 187 |
f = traj[i].get_forces() |
| 188 |
self.update(r, f, r0, f0)
|
| 189 |
r0 = r.copy() |
| 190 |
f0 = f.copy() |
| 191 |
self.iteration += 1 |
| 192 |
self.r0 = r0
|
| 193 |
self.f0 = f0
|
| 194 |
|
| 195 |
def func(self, x): |
| 196 |
"""Objective function for use of the optimizers"""
|
| 197 |
self.atoms.set_positions(x.reshape(-1, 3)) |
| 198 |
self.function_calls += 1 |
| 199 |
return self.atoms.get_potential_energy() |
| 200 |
|
| 201 |
def fprime(self, x): |
| 202 |
"""Gradient of the objective function for use of the optimizers"""
|
| 203 |
self.atoms.set_positions(x.reshape(-1, 3)) |
| 204 |
self.force_calls += 1 |
| 205 |
# Remember that forces are minus the gradient!
|
| 206 |
return - self.atoms.get_forces().reshape(-1) |
| 207 |
|
| 208 |
def line_search(self, r, g, e): |
| 209 |
self.p = self.p.ravel() |
| 210 |
p_size = np.sqrt((self.p **2).sum()) |
| 211 |
if p_size <= np.sqrt(len(self.atoms) * 1e-10): |
| 212 |
self.p /= (p_size / np.sqrt(len(self.atoms)*1e-10)) |
| 213 |
g = g.ravel() |
| 214 |
r = r.ravel() |
| 215 |
ls = LineSearch() |
| 216 |
self.alpha_k, e, self.e0, self.no_update = \ |
| 217 |
ls._line_search(self.func, self.fprime, r, self.p, g, e, self.e0, |
| 218 |
maxstep=self.maxstep, c1=.23, |
| 219 |
c2=.46, stpmax=50.) |
| 220 |
|
| 221 |
class LBFGSLineSearch(LBFGS): |
| 222 |
"""This optimizer uses the LBFGS algorithm, but does a line search that fulfills
|
| 223 |
the Wolff conditions.
|
| 224 |
"""
|
| 225 |
|
| 226 |
def __init__(self, *args, **kwargs): |
| 227 |
kwargs['use_line_search'] = True |
| 228 |
LBFGS.__init__(self, *args, **kwargs)
|
| 229 |
|
| 230 |
# """Modified version of LBFGS.
|
| 231 |
#
|
| 232 |
# This optimizer uses the LBFGS algorithm, but does a line search for the
|
| 233 |
# minimum along the search direction. This is done by issuing an additional
|
| 234 |
# force call for each step, thus doubling the number of calculations.
|
| 235 |
#
|
| 236 |
# Additionally the Hessian is reset if the new guess is not sufficiently
|
| 237 |
# better than the old one.
|
| 238 |
# """
|
| 239 |
# def __init__(self, *args, **kwargs):
|
| 240 |
# self.dR = kwargs.pop('dR', 0.1)
|
| 241 |
# LBFGS.__init__(self, *args, **kwargs)
|
| 242 |
#
|
| 243 |
# def update(self, r, f, r0, f0):
|
| 244 |
# """Update everything that is kept in memory
|
| 245 |
#
|
| 246 |
# This function is mostly here to allow for replay_trajectory.
|
| 247 |
# """
|
| 248 |
# if self.iteration > 0:
|
| 249 |
# a1 = abs(np.dot(f.reshape(-1), f0.reshape(-1)))
|
| 250 |
# a2 = np.dot(f0.reshape(-1), f0.reshape(-1))
|
| 251 |
# if not (a1 <= 0.5 * a2 and a2 != 0):
|
| 252 |
# # Reset optimization
|
| 253 |
# self.initialize()
|
| 254 |
#
|
| 255 |
# # Note that the reset above will set self.iteration to 0 again
|
| 256 |
# # which is why we should check again
|
| 257 |
# if self.iteration > 0:
|
| 258 |
# s0 = r.reshape(-1) - r0.reshape(-1)
|
| 259 |
# self.s.append(s0)
|
| 260 |
#
|
| 261 |
# # We use the gradient which is minus the force!
|
| 262 |
# y0 = f0.reshape(-1) - f.reshape(-1)
|
| 263 |
# self.y.append(y0)
|
| 264 |
#
|
| 265 |
# rho0 = 1.0 / np.dot(y0, s0)
|
| 266 |
# self.rho.append(rho0)
|
| 267 |
#
|
| 268 |
# if self.iteration > self.memory:
|
| 269 |
# self.s.pop(0)
|
| 270 |
# self.y.pop(0)
|
| 271 |
# self.rho.pop(0)
|
| 272 |
#
|
| 273 |
# def determine_step(self, dr):
|
| 274 |
# f = self.atoms.get_forces()
|
| 275 |
#
|
| 276 |
# # Unit-vector along the search direction
|
| 277 |
# du = dr / np.sqrt(np.dot(dr.reshape(-1), dr.reshape(-1)))
|
| 278 |
#
|
| 279 |
# # We keep the old step determination before we figure
|
| 280 |
# # out what is the best to do.
|
| 281 |
# maxstep = self.maxstep * np.sqrt(3 * len(self.atoms))
|
| 282 |
#
|
| 283 |
# # Finite difference step using temporary point
|
| 284 |
# self.atoms.positions += (du * self.dR)
|
| 285 |
# # Decide how much to move along the line du
|
| 286 |
# Fp1 = np.dot(f.reshape(-1), du.reshape(-1))
|
| 287 |
# Fp2 = np.dot(self.atoms.get_forces().reshape(-1), du.reshape(-1))
|
| 288 |
# CR = (Fp1 - Fp2) / self.dR
|
| 289 |
# #RdR = Fp1*0.1
|
| 290 |
# if CR < 0.0:
|
| 291 |
# #print "negcurve"
|
| 292 |
# RdR = maxstep
|
| 293 |
# #if(abs(RdR) > maxstep):
|
| 294 |
# # RdR = self.sign(RdR) * maxstep
|
| 295 |
# else:
|
| 296 |
# Fp = (Fp1 + Fp2) * 0.5
|
| 297 |
# RdR = Fp / CR
|
| 298 |
# if abs(RdR) > maxstep:
|
| 299 |
# RdR = np.sign(RdR) * maxstep
|
| 300 |
# else:
|
| 301 |
# RdR += self.dR * 0.5
|
| 302 |
# return du * RdR
|
| 303 |
|
| 304 |
class HessLBFGS(LBFGS): |
| 305 |
"""Backwards compatibiliyt class"""
|
| 306 |
def __init__(self, *args, **kwargs): |
| 307 |
if 'method' in kwargs: |
| 308 |
del kwargs['method'] |
| 309 |
sys.stderr.write('Please use LBFGS instead of HessLBFGS!')
|
| 310 |
LBFGS.__init__(self, *args, **kwargs)
|
| 311 |
|
| 312 |
class LineLBFGS(LBFGSLineSearch): |
| 313 |
"""Backwards compatibiliyt class"""
|
| 314 |
def __init__(self, *args, **kwargs): |
| 315 |
if 'method' in kwargs: |
| 316 |
del kwargs['method'] |
| 317 |
sys.stderr.write('Please use LBFGSLineSearch instead of LineLBFGS!')
|
| 318 |
LBFGSLineSearch.__init__(self, *args, **kwargs)
|