| 1 |
""" |
| 2 |
Provides various deterministic and stochastic propagators. |
| 3 |
""" |
| 4 |
|
| 5 |
import numpy |
| 6 |
|
| 7 |
from abc import ABCMeta, abstractmethod |
| 8 |
|
| 9 |
from csb.statistics.samplers.mc import TrajectoryBuilder |
| 10 |
from csb.numeric.integrators import FastLeapFrog, VelocityVerlet |
| 11 |
from csb.numeric import InvertibleMatrix |
| 12 |
|
| 13 |
class AbstractPropagator(object): |
| 14 |
""" |
| 15 |
Abstract propagator class. Subclasses serve to propagate |
| 16 |
an inital state by some dynamics to a final state. |
| 17 |
""" |
| 18 |
|
| 19 |
__metaclass__ = ABCMeta |
| 20 |
|
| 21 |
@abstractmethod |
| 22 |
def generate(self, init_state, length, return_trajectory=False): |
| 23 |
""" |
| 24 |
Generate a trajectory, starting from an initial state with a certain length. |
| 25 |
|
| 26 |
@param init_state: Initial state from which to propagate |
| 27 |
@type init_state: L{State} |
| 28 |
|
| 29 |
@param length: Length of the trajectory (in integration steps or stochastic moves) |
| 30 |
@type length: int |
| 31 |
|
| 32 |
@param return_trajectory: Return complete L{Trajectory} instead of the initial |
| 33 |
and final states only (L{PropagationResult}) |
| 34 |
@type return_trajectory: boolean |
| 35 |
|
| 36 |
@rtype: L{AbstractPropagationResult} |
| 37 |
""" |
| 38 |
pass |
| 39 |
|
| 40 |
class MDPropagator(AbstractPropagator): |
| 41 |
""" |
| 42 |
Molecular Dynamics propagator. Generates a trajectory |
| 43 |
by integration of Hamiltionian equations of motion. |
| 44 |
|
| 45 |
@param gradient: Gradient of potential energy. Guides the dynamics. |
| 46 |
@type gradient: L{AbstractGradient} |
| 47 |
|
| 48 |
@param timestep: Timestep to be used for integration |
| 49 |
@type timestep: float |
| 50 |
|
| 51 |
@param mass_matrix: Mass matrix |
| 52 |
@type mass_matrix: n-dimensional L{InvertibleMatrix} with n being the dimension |
| 53 |
of the configuration space, that is, the dimension of |
| 54 |
the position / momentum vectors |
| 55 |
|
| 56 |
@param integrator: Subclass of L{AbstractIntegrator} to be used to integrate |
| 57 |
Hamiltonian equations of motion |
| 58 |
@type integrator: type |
| 59 |
""" |
| 60 |
|
| 61 |
def __init__(self, gradient, timestep, mass_matrix=None, integrator=FastLeapFrog): |
| 62 |
|
| 63 |
super(MDPropagator, self).__init__() |
| 64 |
|
| 65 |
self._gradient = None |
| 66 |
self.gradient = gradient |
| 67 |
self._mass_matrix = mass_matrix |
| 68 |
self._timestep = None |
| 69 |
self.timestep = timestep |
| 70 |
self._integrator = integrator |
| 71 |
|
| 72 |
self._first_run = True |
| 73 |
|
| 74 |
@property |
| 75 |
def gradient(self): |
| 76 |
return self._gradient |
| 77 |
@gradient.setter |
| 78 |
def gradient(self, value): |
| 79 |
self._gradient = value |
| 80 |
|
| 81 |
@property |
| 82 |
def timestep(self): |
| 83 |
return self._timestep |
| 84 |
@timestep.setter |
| 85 |
def timestep(self, value): |
| 86 |
self._timestep = float(value) |
| 87 |
|
| 88 |
@property |
| 89 |
def mass_matrix(self): |
| 90 |
return self._mass_matrix |
| 91 |
@mass_matrix.setter |
| 92 |
def mass_matrix(self, value): |
| 93 |
self._mass_matrix = value |
| 94 |
|
| 95 |
def generate(self, init_state, length, return_trajectory=False): |
| 96 |
|
| 97 |
integrator = self._integrator(self.timestep, self.gradient) |
| 98 |
|
| 99 |
result = integrator.integrate(init_state, length, |
| 100 |
mass_matrix=self.mass_matrix, |
| 101 |
return_trajectory=return_trajectory) |
| 102 |
|
| 103 |
return result |
| 104 |
|
| 105 |
class Looper(object): |
| 106 |
""" |
| 107 |
Implements an iterable list with a ring-like topology, |
| 108 |
that is, if the iterator points on the last element, |
| 109 |
next() returns the first element. |
| 110 |
""" |
| 111 |
|
| 112 |
def __init__(self, items): |
| 113 |
|
| 114 |
self._items = items |
| 115 |
self._n_items = len(self._items) |
| 116 |
self._current = 0 |
| 117 |
|
| 118 |
def __iter__(self): |
| 119 |
|
| 120 |
return self |
| 121 |
|
| 122 |
def next(self): |
| 123 |
|
| 124 |
if self._current == self._n_items: |
| 125 |
self._current = 0 |
| 126 |
|
| 127 |
self._current += 1 |
| 128 |
|
| 129 |
return self._items[self._current - 1] |
| 130 |
|
| 131 |
|
| 132 |
class ThermostattedMDPropagator(MDPropagator): |
| 133 |
""" |
| 134 |
Thermostatted Molecular Dynamics propagator. Employs the Andersen thermostat |
| 135 |
which simulates collision with particles of a heat bath at a given temperature. |
| 136 |
|
| 137 |
@param gradient: Gradient of potential energy. Guides the dynamics. |
| 138 |
@type gradient: L{AbstractGradient} |
| 139 |
|
| 140 |
@param timestep: Timestep to be used for integration |
| 141 |
@type timestep: float |
| 142 |
|
| 143 |
@param mass_matrix: Mass matrix |
| 144 |
@type mass_matrix: n-dimensional L{InvertibleMatrix} with n being the dimension |
| 145 |
of the configuration space, that is, the dimension of |
| 146 |
the position / momentum vectors |
| 147 |
|
| 148 |
@param temperature: Time-dependent temperature |
| 149 |
@type temperature: Real-valued function |
| 150 |
|
| 151 |
@param collision_probability: collision probability within duration of one timestep |
| 152 |
@type collision_probability: float |
| 153 |
|
| 154 |
@param update_interval: Interval with which momenta are redrawn |
| 155 |
@type update_interval: int |
| 156 |
|
| 157 |
@param integrator: Subclass of L{AbstractIntegrator} to be used to perform |
| 158 |
integration steps between momentum updates |
| 159 |
@type integrator: type |
| 160 |
""" |
| 161 |
|
| 162 |
def __init__(self, gradient, timestep, mass_matrix=None, temperature=lambda t: 1., |
| 163 |
collision_probability=0.1, update_interval=1, integrator=VelocityVerlet): |
| 164 |
|
| 165 |
super(ThermostattedMDPropagator, self).__init__(gradient, timestep, |
| 166 |
mass_matrix, integrator) |
| 167 |
|
| 168 |
self._collision_probability = collision_probability |
| 169 |
self._update_interval = update_interval |
| 170 |
self._temperature = temperature |
| 171 |
|
| 172 |
def _update(self, momentum, T, collision_probability): |
| 173 |
""" |
| 174 |
Simulate collision with heat bath particles. |
| 175 |
|
| 176 |
@param momentum: Momentum |
| 177 |
@type momentum: one-dimensional numpy array of numbers |
| 178 |
|
| 179 |
@param T: Temperature of the heat bath |
| 180 |
@type T: float |
| 181 |
|
| 182 |
@param collision_probability: collision probability within duration of one timestep |
| 183 |
@type collision_probability: float |
| 184 |
|
| 185 |
@rtype: tuple (updated momentum, heat induced by the update) |
| 186 |
""" |
| 187 |
|
| 188 |
d = len(momentum) |
| 189 |
|
| 190 |
heat = 0. |
| 191 |
update_list = numpy.where(numpy.random.random(d) < collision_probability)[0] |
| 192 |
|
| 193 |
if len(update_list) > 0: |
| 194 |
K = None |
| 195 |
if self.mass_matrix.is_unity_multiple: |
| 196 |
K = lambda x: 0.5 * sum(x ** 2) / self.mass_matrix[0][0] |
| 197 |
else: |
| 198 |
K = lambda x: 0.5 * numpy.dot(x.T, numpy.dot(self.mass_matrix.inverse, x)) |
| 199 |
|
| 200 |
ke_old = K(momentum) |
| 201 |
|
| 202 |
updated_momentum = [numpy.sqrt(T) * self._random_loopers[i].next() for i in update_list] |
| 203 |
momentum[update_list] = updated_momentum |
| 204 |
heat = (K(momentum) - ke_old) / T |
| 205 |
|
| 206 |
return momentum, heat |
| 207 |
|
| 208 |
def _step(self, i, state, heat, integrator): |
| 209 |
""" |
| 210 |
Performs one step consisting of an integration step |
| 211 |
and possibly a momentum update |
| 212 |
|
| 213 |
@param i: integration step count |
| 214 |
@type i: int |
| 215 |
|
| 216 |
@param state: state to be updated |
| 217 |
@type state: L{State} |
| 218 |
|
| 219 |
@param heat: heat produced up to the current integration step |
| 220 |
@type heat: float |
| 221 |
|
| 222 |
@param integrator: integration scheme used to evolve the state deterministically |
| 223 |
@type integrator: L{AbstractIntegrator} |
| 224 |
""" |
| 225 |
|
| 226 |
state = integrator.integrate_once(state, i, mass_matrix=self.mass_matrix) |
| 227 |
|
| 228 |
if i % self._update_interval == 0: |
| 229 |
state.momentum, stepheat = self._update(state.momentum, |
| 230 |
self._temperature(i * self.timestep), |
| 231 |
self._collision_probability) |
| 232 |
|
| 233 |
heat += stepheat |
| 234 |
|
| 235 |
return state, heat |
| 236 |
|
| 237 |
def generate(self, init_state, length, return_trajectory=False): |
| 238 |
|
| 239 |
if self._first_run == True and self.mass_matrix is None: |
| 240 |
d = len(init_state.position) |
| 241 |
self.mass_matrix = InvertibleMatrix(numpy.eye(d), numpy.eye(d)) |
| 242 |
|
| 243 |
integrator = self._integrator(self.timestep, self.gradient) |
| 244 |
builder = TrajectoryBuilder.create(full=return_trajectory) |
| 245 |
|
| 246 |
builder.add_initial_state(init_state) |
| 247 |
|
| 248 |
heat = 0. |
| 249 |
state = init_state.clone() |
| 250 |
|
| 251 |
d = len(state.position) |
| 252 |
|
| 253 |
n_randoms = int(1.5 * length * self._collision_probability / float(self._update_interval)) |
| 254 |
|
| 255 |
if n_randoms < 5: |
| 256 |
n_randoms = 5 |
| 257 |
|
| 258 |
if not self.mass_matrix.is_unity_multiple: |
| 259 |
randoms = numpy.random.multivariate_normal(mean=numpy.zeros(d), |
| 260 |
cov=self.mass_matrix, |
| 261 |
size=n_randoms).T |
| 262 |
else: |
| 263 |
randoms = numpy.random.normal(scale=numpy.sqrt(self.mass_matrix[0][0]), |
| 264 |
size=(d, n_randoms)) |
| 265 |
self._random_loopers = [Looper(x) for x in randoms] |
| 266 |
|
| 267 |
for i in range(length - 1): |
| 268 |
state, heat = self._step(i, state, heat, integrator) |
| 269 |
builder.add_intermediate_state(state) |
| 270 |
|
| 271 |
state, heat = self._step(length - 1, state, heat, integrator) |
| 272 |
builder.add_final_state(state) |
| 273 |
|
| 274 |
traj = builder.product |
| 275 |
traj.heat = heat |
| 276 |
|
| 277 |
return traj |
| 278 |
|
| 279 |
class AbstractMCPropagator(AbstractPropagator): |
| 280 |
""" |
| 281 |
Provides the interface for MC trajectory generators. Implementations |
| 282 |
generate a sequence of states according to some implementation of |
| 283 |
L{AbstractSingleChainMC}. |
| 284 |
|
| 285 |
@param pdf: PDF to sample from |
| 286 |
@type pdf: L{AbstractDensity} |
| 287 |
|
| 288 |
@param temperature: See documentation of L{AbstractSingleChainMC} |
| 289 |
@type temperature: float |
| 290 |
""" |
| 291 |
|
| 292 |
__metaclass__ = ABCMeta |
| 293 |
|
| 294 |
def __init__(self, pdf, temperature=1.): |
| 295 |
|
| 296 |
self._pdf = pdf |
| 297 |
self._temperature = temperature |
| 298 |
self._acceptance_rate = 0.0 |
| 299 |
|
| 300 |
def generate(self, init_state, length, return_trajectory=True): |
| 301 |
|
| 302 |
self._init_sampler(init_state) |
| 303 |
self._sampler.state = init_state |
| 304 |
|
| 305 |
builder = TrajectoryBuilder.create(full=return_trajectory) |
| 306 |
|
| 307 |
builder.add_initial_state(init_state) |
| 308 |
|
| 309 |
for i in range(length): |
| 310 |
self._sampler.sample() |
| 311 |
if i != length - 1: |
| 312 |
builder.add_intermediate_state(self._sampler.state) |
| 313 |
|
| 314 |
builder.add_final_state(self._sampler.state) |
| 315 |
|
| 316 |
self._acceptance_rate = self._sampler.acceptance_rate |
| 317 |
|
| 318 |
return builder.product |
| 319 |
|
| 320 |
@abstractmethod |
| 321 |
def _init_sampler(self, init_state): |
| 322 |
""" |
| 323 |
Initializes the sampler with which to obtain the MC state |
| 324 |
trajectory. |
| 325 |
""" |
| 326 |
|
| 327 |
pass |
| 328 |
|
| 329 |
@property |
| 330 |
def acceptance_rate(self): |
| 331 |
""" |
| 332 |
Acceptance rate of the MC sampler that generated the |
| 333 |
trajectory. |
| 334 |
""" |
| 335 |
return self._acceptance_rate |
| 336 |
|
| 337 |
|
| 338 |
class RWMCPropagator(AbstractMCPropagator): |
| 339 |
""" |
| 340 |
Draws a number of samples from a PDF using the L{RWMCSampler} and |
| 341 |
returns them as a L{Trajectory}. |
| 342 |
|
| 343 |
@param pdf: PDF to sample from |
| 344 |
@type pdf: L{AbstractDensity} |
| 345 |
@param stepsize: Serves to set the step size in |
| 346 |
proposal_density, e.g. for automatic acceptance |
| 347 |
rate adaption |
| 348 |
@type stepsize: float |
| 349 |
@param proposal_density: The proposal density as a function f(x, s) |
| 350 |
of the current state x and the stepsize s. |
| 351 |
By default, the proposal density is uniform, |
| 352 |
centered around x, and has width s. |
| 353 |
@type proposal_density: callable |
| 354 |
|
| 355 |
@param temperature: See documentation of L{AbstractSingleChainMC} |
| 356 |
@type temperature: float |
| 357 |
""" |
| 358 |
|
| 359 |
def __init__(self, pdf, stepsize=1., proposal_density=None, temperature=1.): |
| 360 |
|
| 361 |
super(RWMCPropagator, self).__init__(pdf, temperature) |
| 362 |
|
| 363 |
self._stepsize = stepsize |
| 364 |
self._proposal_density = proposal_density |
| 365 |
|
| 366 |
def _init_sampler(self, init_state): |
| 367 |
|
| 368 |
from csb.statistics.samplers.mc.singlechain import RWMCSampler |
| 369 |
|
| 370 |
self._sampler = RWMCSampler(self._pdf, init_state, self._stepsize, |
| 371 |
self._proposal_density, self._temperature) |
| 372 |
|
| 373 |
class HMCPropagator(AbstractMCPropagator): |
| 374 |
""" |
| 375 |
Draws a number of samples from a PDF using the L{HMCSampler} and |
| 376 |
returns them as a L{Trajectory}. |
| 377 |
|
| 378 |
@param pdf: PDF to sample from |
| 379 |
@type pdf: L{AbstractDensity} |
| 380 |
@param gradient: Gradient of the negative log-probability |
| 381 |
@type gradient: L{AbstractGradient} |
| 382 |
|
| 383 |
@param timestep: Timestep used for integration |
| 384 |
@type timestep: float |
| 385 |
|
| 386 |
@param nsteps: Number of integration steps to be performed in |
| 387 |
each iteration |
| 388 |
@type nsteps: int |
| 389 |
|
| 390 |
@param mass_matrix: Mass matrix |
| 391 |
@type mass_matrix: n-dimensional L{InvertibleMatrix} with n being the dimension |
| 392 |
of the configuration space, that is, the dimension of |
| 393 |
the position / momentum vectors |
| 394 |
|
| 395 |
@param integrator: Subclass of L{AbstractIntegrator} to be used for |
| 396 |
integrating Hamiltionian equations of motion |
| 397 |
@type integrator: type |
| 398 |
|
| 399 |
@param temperature: See documentation of L{AbstractSingleChainMC} |
| 400 |
@type temperature: float |
| 401 |
""" |
| 402 |
|
| 403 |
def __init__(self, pdf, gradient, timestep, nsteps, mass_matrix=None, |
| 404 |
integrator=FastLeapFrog, temperature=1.): |
| 405 |
|
| 406 |
super(HMCPropagator, self).__init__(pdf, temperature) |
| 407 |
|
| 408 |
self._gradient = gradient |
| 409 |
self._timestep = timestep |
| 410 |
self._nsteps = nsteps |
| 411 |
self._mass_matrix = mass_matrix |
| 412 |
self._integrator = integrator |
| 413 |
|
| 414 |
def _init_sampler(self, init_state): |
| 415 |
|
| 416 |
from csb.statistics.samplers.mc.singlechain import HMCSampler |
| 417 |
|
| 418 |
self._sampler = HMCSampler(self._pdf, init_state, self._gradient, |
| 419 |
self._timestep, self._nsteps, |
| 420 |
mass_matrix=self.mass_matrix, |
| 421 |
integrator=self._integrator, temperature=self._temperature) |
| 422 |
|
| 423 |
@property |
| 424 |
def mass_matrix(self): |
| 425 |
return self._mass_matrix |
| 426 |
@mass_matrix.setter |
| 427 |
def mass_matrix(self, value): |
| 428 |
self._mass_matrix = value |
| 429 |
|
| 430 |
|
| 431 |
class AbstractNCMCPropagator(AbstractMCPropagator): |
| 432 |
""" |
| 433 |
Draws a number of samples from a PDF using the L{AbstractNCMCSampler}. |
| 434 |
|
| 435 |
@param protocol: The nonequilibrium protocol specifying a sequence of |
| 436 |
perturbation and propagation steps |
| 437 |
@type protocol: L{Protocol} |
| 438 |
|
| 439 |
@param reverse_protocol: The protocol with the order of perturbation and |
| 440 |
propagation reversed in each step. |
| 441 |
@type reverse_protocol: L{Protocol} |
| 442 |
""" |
| 443 |
|
| 444 |
__metaclass__ = ABCMeta |
| 445 |
|
| 446 |
def __init__(self, protocol, reverse_protocol): |
| 447 |
|
| 448 |
self._protocol = None |
| 449 |
self.protocol = protocol |
| 450 |
self._reverse_protocol = None |
| 451 |
self.reverse_protocol = reverse_protocol |
| 452 |
|
| 453 |
pdf = self.protocol.steps[0].perturbation.sys_before.hamiltonian |
| 454 |
temperature = self.protocol.steps[0].perturbation.sys_before.hamiltonian.temperature |
| 455 |
|
| 456 |
super(AbstractNCMCPropagator, self).__init__(pdf, temperature) |
| 457 |
|
| 458 |
@property |
| 459 |
def protocol(self): |
| 460 |
return self._protocol |
| 461 |
@protocol.setter |
| 462 |
def protocol(self, value): |
| 463 |
self._protocol = value |
| 464 |
|
| 465 |
@property |
| 466 |
def reverse_protocol(self): |
| 467 |
return self._reverse_protocol |
| 468 |
@reverse_protocol.setter |
| 469 |
def reverse_protocol(self, value): |
| 470 |
self._reverse_protocol = value |