root / ase / optimize / bfgslinesearch.py @ 14
Historique | Voir | Annoter | Télécharger (14,89 ko)
| 1 |
#__docformat__ = "restructuredtext en"
|
|---|---|
| 2 |
# ******NOTICE***************
|
| 3 |
# optimize.py module by Travis E. Oliphant
|
| 4 |
#
|
| 5 |
# You may copy and use this module as you see fit with no
|
| 6 |
# guarantee implied provided you keep this notice in all copies.
|
| 7 |
# *****END NOTICE************
|
| 8 |
|
| 9 |
import numpy as np |
| 10 |
from numpy import atleast_1d, eye, mgrid, argmin, zeros, shape, empty, \ |
| 11 |
squeeze, vectorize, asarray, absolute, sqrt, Inf, asfarray, isinf |
| 12 |
from ase.utils.linesearch import LineSearch |
| 13 |
from ase.optimize.optimize import Optimizer |
| 14 |
from numpy import arange |
| 15 |
|
| 16 |
|
| 17 |
# These have been copied from Numeric's MLab.py
|
| 18 |
# I don't think they made the transition to scipy_core
|
| 19 |
|
| 20 |
# Modified from scipy_optimize
|
| 21 |
abs = absolute |
| 22 |
import __builtin__ |
| 23 |
pymin = __builtin__.min |
| 24 |
pymax = __builtin__.max |
| 25 |
__version__="0.1"
|
| 26 |
|
| 27 |
class BFGSLineSearch(Optimizer): |
| 28 |
def __init__(self, atoms, restart=None, logfile='-', maxstep=.2, |
| 29 |
trajectory=None, c1=.23, c2=0.46, alpha=10., stpmax=50.): |
| 30 |
"""Minimize a function using the BFGS algorithm.
|
| 31 |
|
| 32 |
Notes:
|
| 33 |
|
| 34 |
Optimize the function, f, whose gradient is given by fprime
|
| 35 |
using the quasi-Newton method of Broyden, Fletcher, Goldfarb,
|
| 36 |
and Shanno (BFGS) See Wright, and Nocedal 'Numerical
|
| 37 |
Optimization', 1999, pg. 198.
|
| 38 |
|
| 39 |
*See Also*:
|
| 40 |
|
| 41 |
scikits.openopt : SciKit which offers a unified syntax to call
|
| 42 |
this and other solvers.
|
| 43 |
|
| 44 |
"""
|
| 45 |
self.maxstep = maxstep
|
| 46 |
self.stpmax = stpmax
|
| 47 |
self.alpha = alpha
|
| 48 |
self.H = None |
| 49 |
self.c1 = c1
|
| 50 |
self.c2 = c2
|
| 51 |
self.force_calls = 0 |
| 52 |
self.function_calls = 0 |
| 53 |
self.r0 = None |
| 54 |
self.g0 = None |
| 55 |
self.e0 = None |
| 56 |
self.load_restart = False |
| 57 |
self.task = 'START' |
| 58 |
self.rep_count = 0 |
| 59 |
self.p = None |
| 60 |
self.alpha_k = None |
| 61 |
self.no_update = False |
| 62 |
self.replay = False |
| 63 |
|
| 64 |
Optimizer.__init__(self, atoms, restart, logfile, trajectory)
|
| 65 |
|
| 66 |
def read(self): |
| 67 |
self.r0, self.g0, self.e0, self.task, self.H = self.load() |
| 68 |
self.load_restart = True |
| 69 |
|
| 70 |
def reset(self): |
| 71 |
print 'reset' |
| 72 |
self.H = None |
| 73 |
self.r0 = None |
| 74 |
self.g0 = None |
| 75 |
self.e0 = None |
| 76 |
self.rep_count = 0 |
| 77 |
|
| 78 |
|
| 79 |
def step(self, f): |
| 80 |
atoms = self.atoms
|
| 81 |
r = atoms.get_positions() |
| 82 |
r = r.reshape(-1)
|
| 83 |
g = -f.reshape(-1) / self.alpha |
| 84 |
#g = -f.reshape(-1)
|
| 85 |
p0 = self.p
|
| 86 |
self.update(r, g, self.r0, self.g0, p0) |
| 87 |
e = atoms.get_potential_energy() / self.alpha
|
| 88 |
#e = self.func(r)
|
| 89 |
|
| 90 |
self.p = -np.dot(self.H,g) |
| 91 |
p_size = np.sqrt((self.p **2).sum()) |
| 92 |
if self.nsteps != 0: |
| 93 |
p0_size = np.sqrt((p0 **2).sum())
|
| 94 |
delta_p = self.p/p_size + p0/p0_size
|
| 95 |
if p_size <= np.sqrt(len(atoms) * 1e-10): |
| 96 |
self.p /= (p_size / np.sqrt(len(atoms)*1e-10)) |
| 97 |
ls = LineSearch() |
| 98 |
self.alpha_k, e, self.e0, self.no_update = \ |
| 99 |
ls._line_search(self.func, self.fprime, r, self.p, g, e, self.e0, |
| 100 |
maxstep=self.maxstep, c1=self.c1, |
| 101 |
c2=self.c2, stpmax=self.stpmax) |
| 102 |
#if alpha_k is None: # line search failed try different one.
|
| 103 |
# alpha_k, fc, gc, e, e0, gfkp1 = \
|
| 104 |
# line_search(self.func, self.fprime,r,p,g,
|
| 105 |
# e,self.e0)
|
| 106 |
#if abs(e - self.e0) < 0.000001:
|
| 107 |
# self.rep_count += 1
|
| 108 |
#else:
|
| 109 |
# self.rep_count = 0
|
| 110 |
|
| 111 |
#if (alpha_k is None) or (self.rep_count >= 3):
|
| 112 |
# # If the line search fails, reset the Hessian matrix and
|
| 113 |
# # start a new line search.
|
| 114 |
# self.reset()
|
| 115 |
# return
|
| 116 |
|
| 117 |
dr = self.alpha_k * self.p |
| 118 |
atoms.set_positions((r+dr).reshape(len(atoms),-1)) |
| 119 |
self.r0 = r
|
| 120 |
self.g0 = g
|
| 121 |
self.dump((self.r0, self.g0, self.e0, self.task, self.H)) |
| 122 |
|
| 123 |
def update(self, r, g, r0, g0, p0): |
| 124 |
self.I = eye(len(self.atoms) * 3, dtype=int) |
| 125 |
if self.H is None: |
| 126 |
self.H = eye(3 * len(self.atoms)) |
| 127 |
#self.H = eye(3 * len(self.atoms)) / self.alpha
|
| 128 |
return
|
| 129 |
else:
|
| 130 |
dr = r - r0 |
| 131 |
dg = g - g0 |
| 132 |
if not ((self.alpha_k > 0 and abs(np.dot(g,p0))-abs(np.dot(g0,p0)) < 0) \ |
| 133 |
or self.replay): |
| 134 |
return
|
| 135 |
if self.no_update == True: |
| 136 |
print 'skip update' |
| 137 |
return
|
| 138 |
|
| 139 |
try: # this was handled in numeric, let it remaines for more safety |
| 140 |
rhok = 1.0 / (np.dot(dg,dr))
|
| 141 |
except ZeroDivisionError: |
| 142 |
rhok = 1000.0
|
| 143 |
print "Divide-by-zero encountered: rhok assumed large" |
| 144 |
if isinf(rhok): # this is patch for np |
| 145 |
rhok = 1000.0
|
| 146 |
print "Divide-by-zero encountered: rhok assumed large" |
| 147 |
A1 = self.I - dr[:, np.newaxis] * dg[np.newaxis, :] * rhok
|
| 148 |
A2 = self.I - dg[:, np.newaxis] * dr[np.newaxis, :] * rhok
|
| 149 |
H0 = self.H
|
| 150 |
self.H = np.dot(A1, np.dot(self.H, A2)) + rhok * dr[:, np.newaxis] \ |
| 151 |
* dr[np.newaxis, :] |
| 152 |
#self.B = np.linalg.inv(self.H)
|
| 153 |
#omega, V = np.linalg.eigh(self.B)
|
| 154 |
#eighfile = open('eigh.log','w')
|
| 155 |
|
| 156 |
def func(self, x): |
| 157 |
"""Objective function for use of the optimizers"""
|
| 158 |
self.atoms.set_positions(x.reshape(-1, 3)) |
| 159 |
self.function_calls += 1 |
| 160 |
# Scale the problem as SciPy uses I as initial Hessian.
|
| 161 |
return self.atoms.get_potential_energy() / self.alpha |
| 162 |
#return self.atoms.get_potential_energy()
|
| 163 |
|
| 164 |
def fprime(self, x): |
| 165 |
"""Gradient of the objective function for use of the optimizers"""
|
| 166 |
self.atoms.set_positions(x.reshape(-1, 3)) |
| 167 |
self.force_calls += 1 |
| 168 |
# Remember that forces are minus the gradient!
|
| 169 |
# Scale the problem as SciPy uses I as initial Hessian.
|
| 170 |
return - self.atoms.get_forces().reshape(-1) / self.alpha |
| 171 |
#return - self.atoms.get_forces().reshape(-1)
|
| 172 |
|
| 173 |
def replay_trajectory(self, traj): |
| 174 |
"""Initialize hessian from old trajectory."""
|
| 175 |
self.replay = True |
| 176 |
if isinstance(traj, str): |
| 177 |
from ase.io.trajectory import PickleTrajectory |
| 178 |
traj = PickleTrajectory(traj, 'r')
|
| 179 |
atoms = traj[0]
|
| 180 |
r0 = None
|
| 181 |
g0 = None
|
| 182 |
for i in range(0, len(traj) - 1): |
| 183 |
r = traj[i].get_positions().ravel() |
| 184 |
g = - traj[i].get_forces().ravel() / self.alpha
|
| 185 |
self.update(r, g, r0, g0, self.p) |
| 186 |
self.p = -np.dot(self.H,g) |
| 187 |
r0 = r.copy() |
| 188 |
g0 = g.copy() |
| 189 |
self.r0 = r0
|
| 190 |
self.g0 = g0
|
| 191 |
#self.r0 = traj[-2].get_positions().ravel()
|
| 192 |
#self.g0 = - traj[-2].get_forces().ravel()
|
| 193 |
|
| 194 |
def wrap_function(function, args): |
| 195 |
ncalls = [0]
|
| 196 |
def function_wrapper(x): |
| 197 |
ncalls[0] += 1 |
| 198 |
return function(x, *args)
|
| 199 |
return ncalls, function_wrapper
|
| 200 |
|
| 201 |
def _cubicmin(a,fa,fpa,b,fb,c,fc): |
| 202 |
# finds the minimizer for a cubic polynomial that goes through the
|
| 203 |
# points (a,fa), (b,fb), and (c,fc) with derivative at a of fpa.
|
| 204 |
#
|
| 205 |
# if no minimizer can be found return None
|
| 206 |
#
|
| 207 |
# f(x) = A *(x-a)^3 + B*(x-a)^2 + C*(x-a) + D
|
| 208 |
|
| 209 |
C = fpa |
| 210 |
D = fa |
| 211 |
db = b-a |
| 212 |
dc = c-a |
| 213 |
if (db == 0) or (dc == 0) or (b==c): return None |
| 214 |
denom = (db*dc)**2 * (db-dc)
|
| 215 |
d1 = empty((2,2)) |
| 216 |
d1[0,0] = dc**2 |
| 217 |
d1[0,1] = -db**2 |
| 218 |
d1[1,0] = -dc**3 |
| 219 |
d1[1,1] = db**3 |
| 220 |
[A,B] = np.dot(d1,asarray([fb-fa-C*db,fc-fa-C*dc]).flatten()) |
| 221 |
A /= denom |
| 222 |
B /= denom |
| 223 |
radical = B*B-3*A*C
|
| 224 |
if radical < 0: return None |
| 225 |
if (A == 0): return None |
| 226 |
xmin = a + (-B + sqrt(radical))/(3*A)
|
| 227 |
return xmin
|
| 228 |
|
| 229 |
def _quadmin(a,fa,fpa,b,fb): |
| 230 |
# finds the minimizer for a quadratic polynomial that goes through
|
| 231 |
# the points (a,fa), (b,fb) with derivative at a of fpa
|
| 232 |
# f(x) = B*(x-a)^2 + C*(x-a) + D
|
| 233 |
D = fa |
| 234 |
C = fpa |
| 235 |
db = b-a*1.0
|
| 236 |
if (db==0): return None |
| 237 |
B = (fb-D-C*db)/(db*db) |
| 238 |
if (B <= 0): return None |
| 239 |
xmin = a - C / (2.0*B)
|
| 240 |
return xmin
|
| 241 |
|
| 242 |
def zoom(a_lo, a_hi, phi_lo, phi_hi, derphi_lo, |
| 243 |
phi, derphi, phi0, derphi0, c1, c2): |
| 244 |
maxiter = 10
|
| 245 |
i = 0
|
| 246 |
delta1 = 0.2 # cubic interpolant check |
| 247 |
delta2 = 0.1 # quadratic interpolant check |
| 248 |
phi_rec = phi0 |
| 249 |
a_rec = 0
|
| 250 |
while 1: |
| 251 |
# interpolate to find a trial step length between a_lo and a_hi
|
| 252 |
# Need to choose interpolation here. Use cubic interpolation and then
|
| 253 |
#if the result is within delta * dalpha or outside of the interval
|
| 254 |
#bounded by a_lo or a_hi then use quadratic interpolation, if the
|
| 255 |
#result is still too close, then use bisection
|
| 256 |
|
| 257 |
dalpha = a_hi-a_lo; |
| 258 |
if dalpha < 0: a,b = a_hi,a_lo |
| 259 |
else: a,b = a_lo, a_hi
|
| 260 |
|
| 261 |
# minimizer of cubic interpolant
|
| 262 |
# (uses phi_lo, derphi_lo, phi_hi, and the most recent value of phi)
|
| 263 |
# if the result is too close to the end points (or out of the
|
| 264 |
# interval) then use quadratic interpolation with phi_lo,
|
| 265 |
# derphi_lo and phi_hi
|
| 266 |
# if the result is stil too close to the end points (or out of
|
| 267 |
# the interval) then use bisection
|
| 268 |
|
| 269 |
if (i > 0): |
| 270 |
cchk = delta1*dalpha |
| 271 |
a_j = _cubicmin(a_lo, phi_lo, derphi_lo, a_hi, phi_hi, a_rec, |
| 272 |
phi_rec) |
| 273 |
if (i==0) or (a_j is None) or (a_j > b-cchk) or (a_j < a+cchk): |
| 274 |
qchk = delta2*dalpha |
| 275 |
a_j = _quadmin(a_lo, phi_lo, derphi_lo, a_hi, phi_hi) |
| 276 |
if (a_j is None) or (a_j > b-qchk) or (a_j < a+qchk): |
| 277 |
a_j = a_lo + 0.5*dalpha
|
| 278 |
# print "Using bisection."
|
| 279 |
# else: print "Using quadratic."
|
| 280 |
# else: print "Using cubic."
|
| 281 |
|
| 282 |
# Check new value of a_j
|
| 283 |
|
| 284 |
phi_aj = phi(a_j) |
| 285 |
if (phi_aj > phi0 + c1*a_j*derphi0) or (phi_aj >= phi_lo): |
| 286 |
phi_rec = phi_hi |
| 287 |
a_rec = a_hi |
| 288 |
a_hi = a_j |
| 289 |
phi_hi = phi_aj |
| 290 |
else:
|
| 291 |
derphi_aj = derphi(a_j) |
| 292 |
if abs(derphi_aj) <= -c2*derphi0: |
| 293 |
a_star = a_j |
| 294 |
val_star = phi_aj |
| 295 |
valprime_star = derphi_aj |
| 296 |
break
|
| 297 |
if derphi_aj*(a_hi - a_lo) >= 0: |
| 298 |
phi_rec = phi_hi |
| 299 |
a_rec = a_hi |
| 300 |
a_hi = a_lo |
| 301 |
phi_hi = phi_lo |
| 302 |
else:
|
| 303 |
phi_rec = phi_lo |
| 304 |
a_rec = a_lo |
| 305 |
a_lo = a_j |
| 306 |
phi_lo = phi_aj |
| 307 |
derphi_lo = derphi_aj |
| 308 |
i += 1
|
| 309 |
if (i > maxiter):
|
| 310 |
a_star = a_j |
| 311 |
val_star = phi_aj |
| 312 |
valprime_star = None
|
| 313 |
break
|
| 314 |
return a_star, val_star, valprime_star
|
| 315 |
|
| 316 |
def line_search(f, myfprime, xk, pk, gfk, old_fval, old_old_fval, |
| 317 |
args=(), c1=1e-4, c2=0.9, amax=50): |
| 318 |
"""Find alpha that satisfies strong Wolfe conditions.
|
| 319 |
|
| 320 |
Parameters:
|
| 321 |
|
| 322 |
f : callable f(x,*args)
|
| 323 |
Objective function.
|
| 324 |
myfprime : callable f'(x,*args)
|
| 325 |
Objective function gradient (can be None).
|
| 326 |
xk : ndarray
|
| 327 |
Starting point.
|
| 328 |
pk : ndarray
|
| 329 |
Search direction.
|
| 330 |
gfk : ndarray
|
| 331 |
Gradient value for x=xk (xk being the current parameter
|
| 332 |
estimate).
|
| 333 |
args : tuple
|
| 334 |
Additional arguments passed to objective function.
|
| 335 |
c1 : float
|
| 336 |
Parameter for Armijo condition rule.
|
| 337 |
c2 : float
|
| 338 |
Parameter for curvature condition rule.
|
| 339 |
|
| 340 |
Returns:
|
| 341 |
|
| 342 |
alpha0 : float
|
| 343 |
Alpha for which ``x_new = x0 + alpha * pk``.
|
| 344 |
fc : int
|
| 345 |
Number of function evaluations made.
|
| 346 |
gc : int
|
| 347 |
Number of gradient evaluations made.
|
| 348 |
|
| 349 |
Notes:
|
| 350 |
|
| 351 |
Uses the line search algorithm to enforce strong Wolfe
|
| 352 |
conditions. See Wright and Nocedal, 'Numerical Optimization',
|
| 353 |
1999, pg. 59-60.
|
| 354 |
|
| 355 |
For the zoom phase it uses an algorithm by [...].
|
| 356 |
|
| 357 |
"""
|
| 358 |
|
| 359 |
global _ls_fc, _ls_gc, _ls_ingfk
|
| 360 |
_ls_fc = 0
|
| 361 |
_ls_gc = 0
|
| 362 |
_ls_ingfk = None
|
| 363 |
def phi(alpha): |
| 364 |
global _ls_fc
|
| 365 |
_ls_fc += 1
|
| 366 |
return f(xk+alpha*pk,*args)
|
| 367 |
|
| 368 |
if isinstance(myfprime,type(())): |
| 369 |
def phiprime(alpha): |
| 370 |
global _ls_fc, _ls_ingfk
|
| 371 |
_ls_fc += len(xk)+1 |
| 372 |
eps = myfprime[1]
|
| 373 |
fprime = myfprime[0]
|
| 374 |
newargs = (f,eps) + args |
| 375 |
_ls_ingfk = fprime(xk+alpha*pk,*newargs) # store for later use
|
| 376 |
return np.dot(_ls_ingfk,pk)
|
| 377 |
else:
|
| 378 |
fprime = myfprime |
| 379 |
def phiprime(alpha): |
| 380 |
global _ls_gc, _ls_ingfk
|
| 381 |
_ls_gc += 1
|
| 382 |
_ls_ingfk = fprime(xk+alpha*pk,*args) # store for later use
|
| 383 |
return np.dot(_ls_ingfk,pk)
|
| 384 |
|
| 385 |
|
| 386 |
alpha0 = 0
|
| 387 |
phi0 = old_fval |
| 388 |
derphi0 = np.dot(gfk,pk) |
| 389 |
|
| 390 |
alpha1 = pymin(1.,1.01*2*(phi0-old_old_fval)/derphi0) |
| 391 |
|
| 392 |
if alpha1 == 0: |
| 393 |
# This shouldn't happen. Perhaps the increment has slipped below
|
| 394 |
# machine precision? For now, set the return variables skip the
|
| 395 |
# useless while loop, and raise warnflag=2 due to possible imprecision.
|
| 396 |
alpha_star = None
|
| 397 |
fval_star = old_fval |
| 398 |
old_fval = old_old_fval |
| 399 |
fprime_star = None
|
| 400 |
|
| 401 |
phi_a1 = phi(alpha1) |
| 402 |
#derphi_a1 = phiprime(alpha1) evaluated below
|
| 403 |
|
| 404 |
phi_a0 = phi0 |
| 405 |
derphi_a0 = derphi0 |
| 406 |
|
| 407 |
i = 1
|
| 408 |
maxiter = 10
|
| 409 |
while 1: # bracketing phase |
| 410 |
if alpha1 == 0: |
| 411 |
break
|
| 412 |
if (phi_a1 > phi0 + c1*alpha1*derphi0) or \ |
| 413 |
((phi_a1 >= phi_a0) and (i > 1)): |
| 414 |
alpha_star, fval_star, fprime_star = \ |
| 415 |
zoom(alpha0, alpha1, phi_a0, |
| 416 |
phi_a1, derphi_a0, phi, phiprime, |
| 417 |
phi0, derphi0, c1, c2) |
| 418 |
break
|
| 419 |
|
| 420 |
derphi_a1 = phiprime(alpha1) |
| 421 |
if (abs(derphi_a1) <= -c2*derphi0): |
| 422 |
alpha_star = alpha1 |
| 423 |
fval_star = phi_a1 |
| 424 |
fprime_star = derphi_a1 |
| 425 |
break
|
| 426 |
|
| 427 |
if (derphi_a1 >= 0): |
| 428 |
alpha_star, fval_star, fprime_star = \ |
| 429 |
zoom(alpha1, alpha0, phi_a1, |
| 430 |
phi_a0, derphi_a1, phi, phiprime, |
| 431 |
phi0, derphi0, c1, c2) |
| 432 |
break
|
| 433 |
|
| 434 |
alpha2 = 2 * alpha1 # increase by factor of two on each iteration |
| 435 |
i = i + 1
|
| 436 |
alpha0 = alpha1 |
| 437 |
alpha1 = alpha2 |
| 438 |
phi_a0 = phi_a1 |
| 439 |
phi_a1 = phi(alpha1) |
| 440 |
derphi_a0 = derphi_a1 |
| 441 |
|
| 442 |
# stopping test if lower function not found
|
| 443 |
if (i > maxiter):
|
| 444 |
alpha_star = alpha1 |
| 445 |
fval_star = phi_a1 |
| 446 |
fprime_star = None
|
| 447 |
break
|
| 448 |
|
| 449 |
if fprime_star is not None: |
| 450 |
# fprime_star is a number (derphi) -- so use the most recently
|
| 451 |
# calculated gradient used in computing it derphi = gfk*pk
|
| 452 |
# this is the gradient at the next step no need to compute it
|
| 453 |
# again in the outer loop.
|
| 454 |
fprime_star = _ls_ingfk |
| 455 |
|
| 456 |
return alpha_star, _ls_fc, _ls_gc, fval_star, old_fval, fprime_star
|
| 457 |
|