| 1 |
""" |
| 2 |
provides various integration schemes and an abstract gradient class. |
| 3 |
""" |
| 4 |
|
| 5 |
import numpy |
| 6 |
|
| 7 |
from abc import ABCMeta, abstractmethod |
| 8 |
from csb.statistics.samplers.mc import State, TrajectoryBuilder |
| 9 |
from csb.numeric import InvertibleMatrix |
| 10 |
|
| 11 |
|
| 12 |
class AbstractIntegrator(object): |
| 13 |
""" |
| 14 |
Abstract integrator class. Subclasses implement different integration |
| 15 |
schemes for solving deterministic equations of motion. |
| 16 |
|
| 17 |
@param timestep: Integration timestep |
| 18 |
@type timestep: float |
| 19 |
|
| 20 |
@param gradient: Gradient of potential energy |
| 21 |
@type gradient: L{AbstractGradient} |
| 22 |
""" |
| 23 |
|
| 24 |
__metaclass__ = ABCMeta |
| 25 |
|
| 26 |
def __init__(self, timestep, gradient): |
| 27 |
|
| 28 |
self._timestep = timestep |
| 29 |
self._gradient = gradient |
| 30 |
|
| 31 |
def integrate(self, init_state, length, mass_matrix=None, return_trajectory=False): |
| 32 |
""" |
| 33 |
Integrates equations of motion starting from an initial state a certain |
| 34 |
number of steps. |
| 35 |
|
| 36 |
@param init_state: Initial state from which to start integration |
| 37 |
@type init_state: L{State} |
| 38 |
|
| 39 |
@param length: Nubmer of integration steps to be performed |
| 40 |
@type length: int |
| 41 |
|
| 42 |
@param mass_matrix: Mass matrix |
| 43 |
@type mass_matrix: n-dimensional L{InvertibleMatrix} with n being the dimension |
| 44 |
of the configuration space, that is, the dimension of |
| 45 |
the position / momentum vectors |
| 46 |
|
| 47 |
@param return_trajectory: Return complete L{Trajectory} instead of the initial |
| 48 |
and final states only (L{PropagationResult}). This reduces |
| 49 |
performance. |
| 50 |
@type return_trajectory: boolean |
| 51 |
|
| 52 |
@rtype: L{AbstractPropagationResult} |
| 53 |
""" |
| 54 |
|
| 55 |
builder = TrajectoryBuilder.create(full=return_trajectory) |
| 56 |
|
| 57 |
builder.add_initial_state(init_state) |
| 58 |
state = init_state.clone() |
| 59 |
|
| 60 |
for i in range(length - 1): |
| 61 |
state = self.integrate_once(state, i, mass_matrix=mass_matrix) |
| 62 |
builder.add_intermediate_state(state) |
| 63 |
|
| 64 |
state = self.integrate_once(state, length - 1, mass_matrix=mass_matrix) |
| 65 |
builder.add_final_state(state) |
| 66 |
|
| 67 |
return builder.product |
| 68 |
|
| 69 |
@abstractmethod |
| 70 |
def integrate_once(self, state, current_step, mass_matrix=None): |
| 71 |
""" |
| 72 |
Integrates one step starting from an initial state and an initial time |
| 73 |
given by the product of the timestep and the current_step parameter. |
| 74 |
The input C{state} is changed in place. |
| 75 |
|
| 76 |
@param state: State which to evolve one integration step |
| 77 |
@type state: L{State} |
| 78 |
|
| 79 |
@param current_step: Current integration step |
| 80 |
@type current_step: int |
| 81 |
|
| 82 |
@param mass_matrix: mass matrix |
| 83 |
@type mass_matrix: n-dimensional numpy array with n being the dimension |
| 84 |
of the configuration space, that is, the dimension of |
| 85 |
the position / momentum vectors |
| 86 |
@return: the altered state |
| 87 |
@rtype: L{State} |
| 88 |
""" |
| 89 |
pass |
| 90 |
|
| 91 |
def _get_inverse(self, mass_matrix): |
| 92 |
|
| 93 |
inverse_mass_matrix = None |
| 94 |
if mass_matrix is None: |
| 95 |
inverse_mass_matrix = 1.0 |
| 96 |
else: |
| 97 |
if mass_matrix.is_unity_multiple: |
| 98 |
inverse_mass_matrix = mass_matrix.inverse[0][0] |
| 99 |
else: |
| 100 |
inverse_mass_matrix = mass_matrix.inverse |
| 101 |
|
| 102 |
return inverse_mass_matrix |
| 103 |
|
| 104 |
class LeapFrog(AbstractIntegrator): |
| 105 |
""" |
| 106 |
Leap Frog integration scheme implementation that calculates position and |
| 107 |
momenta at equal times. Slower than FastLeapFrog, but intermediate points |
| 108 |
in trajectories obtained using |
| 109 |
LeapFrog.integrate(init_state, length, return_trajectoy=True) are physical. |
| 110 |
""" |
| 111 |
|
| 112 |
def integrate_once(self, state, current_step, mass_matrix=None): |
| 113 |
|
| 114 |
inverse_mass_matrix = self._get_inverse(mass_matrix) |
| 115 |
|
| 116 |
i = current_step |
| 117 |
|
| 118 |
if i == 0: |
| 119 |
self._oldgrad = self._gradient(state.position, 0.) |
| 120 |
|
| 121 |
momentumhalf = state.momentum - 0.5 * self._timestep * self._oldgrad |
| 122 |
state.position = state.position + self._timestep * numpy.dot(inverse_mass_matrix, momentumhalf) |
| 123 |
self._oldgrad = self._gradient(state.position, (i + 1) * self._timestep) |
| 124 |
state.momentum = momentumhalf - 0.5 * self._timestep * self._oldgrad |
| 125 |
|
| 126 |
return state |
| 127 |
|
| 128 |
class FastLeapFrog(LeapFrog): |
| 129 |
""" |
| 130 |
Leap Frog integration scheme implementation that calculates position and |
| 131 |
momenta at unequal times by concatenating the momentum updates of two |
| 132 |
successive integration steps. |
| 133 |
WARNING: intermediate points in trajectories obtained by |
| 134 |
FastLeapFrog.integrate(init_state, length, return_trajectories=True) |
| 135 |
are NOT to be interpreted as phase-space trajectories, because |
| 136 |
position and momenta are not given at equal times! In the initial and the |
| 137 |
final state, positions and momenta are given at equal times. |
| 138 |
""" |
| 139 |
|
| 140 |
def integrate(self, init_state, length, mass_matrix=None, return_trajectory=False): |
| 141 |
|
| 142 |
inverse_mass_matrix = self._get_inverse(mass_matrix) |
| 143 |
|
| 144 |
builder = TrajectoryBuilder.create(full=return_trajectory) |
| 145 |
|
| 146 |
builder.add_initial_state(init_state) |
| 147 |
state = init_state.clone() |
| 148 |
|
| 149 |
state.momentum = state.momentum - 0.5 * self._timestep * self._gradient(state.position, 0.) |
| 150 |
|
| 151 |
for i in range(length-1): |
| 152 |
state.position = state.position + self._timestep * numpy.dot(inverse_mass_matrix, state.momentum) |
| 153 |
state.momentum = state.momentum - self._timestep * \ |
| 154 |
self._gradient(state.position, (i + 1) * self._timestep) |
| 155 |
builder.add_intermediate_state(state) |
| 156 |
|
| 157 |
state.position = state.position + self._timestep * numpy.dot(inverse_mass_matrix, state.momentum) |
| 158 |
state.momentum = state.momentum - 0.5 * self._timestep * \ |
| 159 |
self._gradient(state.position, length * self._timestep) |
| 160 |
builder.add_final_state(state) |
| 161 |
|
| 162 |
return builder.product |
| 163 |
|
| 164 |
class VelocityVerlet(AbstractIntegrator): |
| 165 |
""" |
| 166 |
Velocity Verlet integration scheme implementation. |
| 167 |
""" |
| 168 |
|
| 169 |
def integrate_once(self, state, current_step, mass_matrix=None): |
| 170 |
|
| 171 |
inverse_mass_matrix = self._get_inverse(mass_matrix) |
| 172 |
|
| 173 |
i = current_step |
| 174 |
|
| 175 |
if i == 0: |
| 176 |
self._oldgrad = self._gradient(state.position, 0.) |
| 177 |
|
| 178 |
state.position = state.position + self._timestep * numpy.dot(inverse_mass_matrix, state.momentum) \ |
| 179 |
- 0.5 * self._timestep ** 2 * numpy.dot(inverse_mass_matrix, self._oldgrad) |
| 180 |
newgrad = self._gradient(state.position, (i + 1) * self._timestep) |
| 181 |
state.momentum = state.momentum - 0.5 * self._timestep * (self._oldgrad + newgrad) |
| 182 |
self._oldgrad = newgrad |
| 183 |
|
| 184 |
return state |
| 185 |
|
| 186 |
class AbstractGradient(object): |
| 187 |
""" |
| 188 |
Abstract gradient class. Implementations evaluate the gradient of an energy |
| 189 |
function. |
| 190 |
""" |
| 191 |
|
| 192 |
__metaclass__ = ABCMeta |
| 193 |
|
| 194 |
@abstractmethod |
| 195 |
def evaluate(self, q, t): |
| 196 |
""" |
| 197 |
Evaluates the gradient at position q and time t. |
| 198 |
|
| 199 |
@param q: Position array |
| 200 |
@type q: One-dimensional numpy array |
| 201 |
|
| 202 |
@param t: Time |
| 203 |
@type t: float |
| 204 |
|
| 205 |
@rtype: numpy array |
| 206 |
""" |
| 207 |
pass |
| 208 |
|
| 209 |
def __call__(self, q, t): |
| 210 |
""" |
| 211 |
Evaluates the gradient at position q and time t. |
| 212 |
|
| 213 |
@param q: Position array |
| 214 |
@type q: One-dimensional numpy array |
| 215 |
|
| 216 |
@param t: Time |
| 217 |
@type t: float |
| 218 |
|
| 219 |
@rtype: numpy array |
| 220 |
""" |
| 221 |
State.check_flat_array(q) |
| 222 |
return self.evaluate(q, t) |