root / ase / optimize / sciopt.py @ 5
Historique | Voir | Annoter | Télécharger (9,07 ko)
| 1 |
import numpy as np |
|---|---|
| 2 |
try:
|
| 3 |
import scipy.optimize as opt |
| 4 |
except ImportError: |
| 5 |
pass
|
| 6 |
|
| 7 |
from ase.optimize.optimize import Optimizer |
| 8 |
|
| 9 |
|
| 10 |
class Converged(Exception): |
| 11 |
pass
|
| 12 |
|
| 13 |
class OptimizerConvergenceError(Exception): |
| 14 |
pass
|
| 15 |
|
| 16 |
class SciPyOptimizer(Optimizer): |
| 17 |
"""General interface for SciPy optimizers
|
| 18 |
|
| 19 |
Only the call to the optimizer is still needed
|
| 20 |
"""
|
| 21 |
def __init__(self, atoms, logfile='-', trajectory=None, |
| 22 |
callback_always=False, alpha=70.0): |
| 23 |
"""Initialize object
|
| 24 |
|
| 25 |
Parameters:
|
| 26 |
|
| 27 |
callback_always: book
|
| 28 |
Should the callback be run after each force call (also in the
|
| 29 |
linesearch)
|
| 30 |
|
| 31 |
alpha: float
|
| 32 |
Initial guess for the Hessian (curvature of energy surface). A
|
| 33 |
conservative value of 70.0 is the default, but number of needed
|
| 34 |
steps to converge might be less if a lower value is used. However,
|
| 35 |
a lower value also means risk of instability.
|
| 36 |
|
| 37 |
"""
|
| 38 |
restart = None
|
| 39 |
Optimizer.__init__(self, atoms, restart, logfile, trajectory)
|
| 40 |
self.force_calls = 0 |
| 41 |
self.callback_always = callback_always
|
| 42 |
self.H0 = alpha
|
| 43 |
|
| 44 |
def x0(self): |
| 45 |
"""Return x0 in a way SciPy can use
|
| 46 |
|
| 47 |
This class is mostly usable for subclasses wanting to redefine the
|
| 48 |
parameters (and the objective function)"""
|
| 49 |
return self.atoms.get_positions().reshape(-1) |
| 50 |
|
| 51 |
def f(self, x): |
| 52 |
"""Objective function for use of the optimizers"""
|
| 53 |
self.atoms.set_positions(x.reshape(-1, 3)) |
| 54 |
# Scale the problem as SciPy uses I as initial Hessian.
|
| 55 |
return self.atoms.get_potential_energy() / self.H0 |
| 56 |
|
| 57 |
def fprime(self, x): |
| 58 |
"""Gradient of the objective function for use of the optimizers"""
|
| 59 |
self.atoms.set_positions(x.reshape(-1, 3)) |
| 60 |
self.force_calls += 1 |
| 61 |
|
| 62 |
if self.callback_always: |
| 63 |
self.callback(x)
|
| 64 |
|
| 65 |
# Remember that forces are minus the gradient!
|
| 66 |
# Scale the problem as SciPy uses I as initial Hessian.
|
| 67 |
return - self.atoms.get_forces().reshape(-1) / self.H0 |
| 68 |
|
| 69 |
def callback(self, x): |
| 70 |
"""Callback function to be run after each iteration by SciPy
|
| 71 |
|
| 72 |
This should also be called once before optimization starts, as SciPy
|
| 73 |
optimizers only calls it after each iteration, while ase optimizers
|
| 74 |
call something similar before as well.
|
| 75 |
"""
|
| 76 |
f = self.atoms.get_forces()
|
| 77 |
self.log(f)
|
| 78 |
self.call_observers()
|
| 79 |
if self.converged(f): |
| 80 |
raise Converged
|
| 81 |
self.nsteps += 1 |
| 82 |
|
| 83 |
def run(self, fmax=0.05, steps=100000000): |
| 84 |
self.fmax = fmax
|
| 85 |
# As SciPy does not log the zeroth iteration, we do that manually
|
| 86 |
self.callback(None) |
| 87 |
try:
|
| 88 |
# Scale the problem as SciPy uses I as initial Hessian.
|
| 89 |
self.call_fmin(fmax / self.H0, steps) |
| 90 |
except Converged:
|
| 91 |
pass
|
| 92 |
|
| 93 |
def dump(self, data): |
| 94 |
pass
|
| 95 |
|
| 96 |
def load(self): |
| 97 |
pass
|
| 98 |
|
| 99 |
def call_fmin(self, fmax, steps): |
| 100 |
raise NotImplementedError |
| 101 |
|
| 102 |
class SciPyFminCG(SciPyOptimizer): |
| 103 |
"""Non-linear (Polak-Ribiere) conjugate gradient algorithm"""
|
| 104 |
def call_fmin(self, fmax, steps): |
| 105 |
output = opt.fmin_cg(self.f,
|
| 106 |
self.x0(),
|
| 107 |
fprime=self.fprime,
|
| 108 |
#args=(),
|
| 109 |
gtol=fmax * 0.1, #Should never be reached |
| 110 |
norm=np.inf, |
| 111 |
#epsilon=
|
| 112 |
maxiter=steps, |
| 113 |
full_output=1,
|
| 114 |
disp=0,
|
| 115 |
#retall=0,
|
| 116 |
callback=self.callback
|
| 117 |
) |
| 118 |
warnflag = output[-1]
|
| 119 |
if warnflag == 2: |
| 120 |
raise OptimizerConvergenceError('Warning: Desired error not necessarily achieved ' \ |
| 121 |
'due to precision loss')
|
| 122 |
|
| 123 |
class SciPyFminBFGS(SciPyOptimizer): |
| 124 |
"""Quasi-Newton method (Broydon-Fletcher-Goldfarb-Shanno)"""
|
| 125 |
def call_fmin(self, fmax, steps): |
| 126 |
output = opt.fmin_bfgs(self.f,
|
| 127 |
self.x0(),
|
| 128 |
fprime=self.fprime,
|
| 129 |
#args=(),
|
| 130 |
gtol=fmax * 0.1, #Should never be reached |
| 131 |
norm=np.inf, |
| 132 |
#epsilon=1.4901161193847656e-08,
|
| 133 |
maxiter=steps, |
| 134 |
full_output=1,
|
| 135 |
disp=0,
|
| 136 |
#retall=0,
|
| 137 |
callback=self.callback
|
| 138 |
) |
| 139 |
warnflag = output[-1]
|
| 140 |
if warnflag == 2: |
| 141 |
raise OptimizerConvergenceError('Warning: Desired error not necessarily achieved' \ |
| 142 |
'due to precision loss')
|
| 143 |
|
| 144 |
class SciPyGradientlessOptimizer(Optimizer): |
| 145 |
"""General interface for gradient less SciPy optimizers
|
| 146 |
|
| 147 |
Only the call to the optimizer is still needed
|
| 148 |
|
| 149 |
Note: If you redefien x0() and f(), you don't even need an atoms object.
|
| 150 |
Redefining these also allows you to specify an arbitrary objective
|
| 151 |
function.
|
| 152 |
|
| 153 |
XXX: This is still a work in progress
|
| 154 |
"""
|
| 155 |
def __init__(self, atoms, logfile='-', trajectory=None, |
| 156 |
callback_always=False):
|
| 157 |
"""Parameters:
|
| 158 |
|
| 159 |
callback_always: book
|
| 160 |
Should the callback be run after each force call (also in the
|
| 161 |
linesearch)
|
| 162 |
"""
|
| 163 |
restart = None
|
| 164 |
Optimizer.__init__(self, atoms, restart, logfile, trajectory)
|
| 165 |
self.function_calls = 0 |
| 166 |
self.callback_always = callback_always
|
| 167 |
|
| 168 |
def x0(self): |
| 169 |
"""Return x0 in a way SciPy can use
|
| 170 |
|
| 171 |
This class is mostly usable for subclasses wanting to redefine the
|
| 172 |
parameters (and the objective function)"""
|
| 173 |
return self.atoms.get_positions().reshape(-1) |
| 174 |
|
| 175 |
def f(self, x): |
| 176 |
"""Objective function for use of the optimizers"""
|
| 177 |
self.atoms.set_positions(x.reshape(-1, 3)) |
| 178 |
self.function_calls += 1 |
| 179 |
# Scale the problem as SciPy uses I as initial Hessian.
|
| 180 |
return self.atoms.get_potential_energy() |
| 181 |
|
| 182 |
def callback(self, x): |
| 183 |
"""Callback function to be run after each iteration by SciPy
|
| 184 |
|
| 185 |
This should also be called once before optimization starts, as SciPy
|
| 186 |
optimizers only calls it after each iteration, while ase optimizers
|
| 187 |
call something similar before as well.
|
| 188 |
"""
|
| 189 |
# We can't assume that forces are available!
|
| 190 |
#f = self.atoms.get_forces()
|
| 191 |
#self.log(f)
|
| 192 |
self.call_observers()
|
| 193 |
#if self.converged(f):
|
| 194 |
# raise Converged
|
| 195 |
self.nsteps += 1 |
| 196 |
|
| 197 |
def run(self, ftol=0.01, xtol=0.01, steps=100000000): |
| 198 |
self.xtol = xtol
|
| 199 |
self.ftol = ftol
|
| 200 |
# As SciPy does not log the zeroth iteration, we do that manually
|
| 201 |
self.callback(None) |
| 202 |
try:
|
| 203 |
# Scale the problem as SciPy uses I as initial Hessian.
|
| 204 |
self.call_fmin(xtol, ftol, steps)
|
| 205 |
except Converged:
|
| 206 |
pass
|
| 207 |
|
| 208 |
def dump(self, data): |
| 209 |
pass
|
| 210 |
|
| 211 |
def load(self): |
| 212 |
pass
|
| 213 |
|
| 214 |
def call_fmin(self, fmax, steps): |
| 215 |
raise NotImplementedError |
| 216 |
|
| 217 |
class SciPyFmin(SciPyGradientlessOptimizer): |
| 218 |
"""Nelder-Mead Simplex algorithm
|
| 219 |
|
| 220 |
Uses only function calls.
|
| 221 |
|
| 222 |
XXX: This is still a work in progress
|
| 223 |
"""
|
| 224 |
def call_fmin(self, xtol, ftol, steps): |
| 225 |
output = opt.fmin(self.f,
|
| 226 |
self.x0(),
|
| 227 |
#args=(),
|
| 228 |
xtol=xtol, |
| 229 |
ftol=ftol, |
| 230 |
maxiter=steps, |
| 231 |
#maxfun=None,
|
| 232 |
#full_output=1,
|
| 233 |
disp=0,
|
| 234 |
#retall=0,
|
| 235 |
callback=self.callback
|
| 236 |
) |
| 237 |
|
| 238 |
class SciPyFminPowell(SciPyGradientlessOptimizer): |
| 239 |
"""Powell's (modified) level set method
|
| 240 |
|
| 241 |
Uses only function calls.
|
| 242 |
|
| 243 |
XXX: This is still a work in progress
|
| 244 |
"""
|
| 245 |
def __init__(self, *args, **kwargs): |
| 246 |
"""Parameters:
|
| 247 |
|
| 248 |
direc: float
|
| 249 |
How much to change x to initially. Defaults to 0.04.
|
| 250 |
"""
|
| 251 |
direc = kwargs.pop('direc', None) |
| 252 |
SciPyGradientlessOptimizer.__init__(self, *args, **kwargs)
|
| 253 |
|
| 254 |
if direc is None: |
| 255 |
self.direc = np.eye(len(self.x0()), dtype=float) * 0.04 |
| 256 |
else:
|
| 257 |
self.direc = np.eye(len(self.x0()), dtype=float) * direc |
| 258 |
|
| 259 |
def call_fmin(self, xtol, ftol, steps): |
| 260 |
output = opt.fmin_powell(self.f,
|
| 261 |
self.x0(),
|
| 262 |
#args=(),
|
| 263 |
xtol=xtol, |
| 264 |
ftol=ftol, |
| 265 |
maxiter=steps, |
| 266 |
#maxfun=None,
|
| 267 |
#full_output=1,
|
| 268 |
disp=0,
|
| 269 |
#retall=0,
|
| 270 |
callback=self.callback,
|
| 271 |
direc=self.direc
|
| 272 |
) |