| 1 |
""" |
| 2 |
Random number generators |
| 3 |
""" |
| 4 |
|
| 5 |
def probability_transform(shape, inv_cum, cum_min=0., cum_max=1.): |
| 6 |
""" |
| 7 |
Generic sampler based on the probability transform. |
| 8 |
|
| 9 |
@param shape: shape of the random sample |
| 10 |
@param inv_cum: inversion of the cumulative density function from which one seeks to sample |
| 11 |
@param cum_min: lower value of the cumulative distribution |
| 12 |
@param cum_max: upper value of the cumulative distribution |
| 13 |
@return: random variates of the PDF implied by the inverse cumulative distribution |
| 14 |
""" |
| 15 |
from numpy.random import random |
| 16 |
|
| 17 |
return inv_cum(cum_min + random(shape) * (cum_max - cum_min)) |
| 18 |
|
| 19 |
def truncated_gamma(shape=None, alpha=1., beta=1., x_min=None, x_max=None): |
| 20 |
""" |
| 21 |
Generate random variates from a lower-and upper-bounded gamma distribution. |
| 22 |
|
| 23 |
@param shape: shape of the random sample |
| 24 |
@param alpha: shape parameter (alpha > 0.) |
| 25 |
@param beta: scale parameter (beta >= 0.) |
| 26 |
@param x_min: lower bound of variate |
| 27 |
@param x_max: upper bound of variate |
| 28 |
@return: random variates of lower-bounded gamma distribution |
| 29 |
""" |
| 30 |
from scipy.special import gammainc, gammaincinv |
| 31 |
from numpy.random import gamma |
| 32 |
from numpy import inf |
| 33 |
|
| 34 |
if x_min is None and x_max is None: |
| 35 |
return gamma(alpha, 1 / beta, shape) |
| 36 |
elif x_min is None: |
| 37 |
x_min = 0. |
| 38 |
elif x_max is None: |
| 39 |
x_max = inf |
| 40 |
|
| 41 |
x_min = max(0., x_min) |
| 42 |
x_max = min(1e300, x_max) |
| 43 |
|
| 44 |
a = gammainc(alpha, beta * x_min) |
| 45 |
b = gammainc(alpha, beta * x_max) |
| 46 |
|
| 47 |
return probability_transform(shape, |
| 48 |
lambda x, alpha=alpha: gammaincinv(alpha, x), |
| 49 |
a, b) / beta |
| 50 |
|
| 51 |
def truncated_normal(shape=None, mu=0., sigma=1., x_min=None, x_max=None): |
| 52 |
""" |
| 53 |
Generates random variates from a lower-and upper-bounded normal distribution |
| 54 |
|
| 55 |
@param shape: shape of the random sample |
| 56 |
@param mu: location parameter |
| 57 |
@param sigma: width of the distribution (sigma >= 0.) |
| 58 |
@param x_min: lower bound of variate |
| 59 |
@param x_max: upper bound of variate |
| 60 |
@return: random variates of lower-bounded normal distribution |
| 61 |
""" |
| 62 |
from scipy.special import erf, erfinv |
| 63 |
from numpy.random import standard_normal |
| 64 |
from numpy import inf, sqrt |
| 65 |
|
| 66 |
if x_min is None and x_max is None: |
| 67 |
return standard_normal(shape) * sigma + mu |
| 68 |
elif x_min is None: |
| 69 |
x_min = -inf |
| 70 |
elif x_max is None: |
| 71 |
x_max = inf |
| 72 |
|
| 73 |
x_min = max(-1e300, x_min) |
| 74 |
x_max = min(+1e300, x_max) |
| 75 |
var = sigma ** 2 + 1e-300 |
| 76 |
sigma = sqrt(2 * var) |
| 77 |
|
| 78 |
a = erf((x_min - mu) / sigma) |
| 79 |
b = erf((x_max - mu) / sigma) |
| 80 |
|
| 81 |
return probability_transform(shape, erfinv, a, b) * sigma + mu |
| 82 |
|
| 83 |
def sample_dirichlet(alpha, n_samples=1): |
| 84 |
""" |
| 85 |
Sample points from a dirichlet distribution with parameter alpha. |
| 86 |
|
| 87 |
@param alpha: alpha parameter of a dirichlet distribution |
| 88 |
@type alpha: array |
| 89 |
""" |
| 90 |
from numpy import array, sum, transpose, ones |
| 91 |
from numpy.random import gamma |
| 92 |
|
| 93 |
alpha = array(alpha, ndmin=1) |
| 94 |
X = gamma(alpha, |
| 95 |
ones(len(alpha)), |
| 96 |
[n_samples, len(alpha)]) |
| 97 |
|
| 98 |
return transpose(transpose(X) / sum(X, -1)) |
| 99 |
|
| 100 |
def sample_sphere3d(radius=1., n_samples=1): |
| 101 |
""" |
| 102 |
Sample points from 3D sphere. |
| 103 |
|
| 104 |
@param radius: radius of the sphere |
| 105 |
@type radius: float |
| 106 |
|
| 107 |
@param n_samples: number of samples to return |
| 108 |
@type n_samples: int |
| 109 |
|
| 110 |
@return: n_samples times random cartesian coordinates inside the sphere |
| 111 |
@rtype: numpy array |
| 112 |
""" |
| 113 |
from numpy.random import random |
| 114 |
from numpy import arccos, transpose, cos, sin, pi, power |
| 115 |
|
| 116 |
r = radius * power(random(n_samples), 1 / 3.) |
| 117 |
theta = arccos(2. * (random(n_samples) - 0.5)) |
| 118 |
phi = 2 * pi * random(n_samples) |
| 119 |
|
| 120 |
x = cos(phi) * sin(theta) * r |
| 121 |
y = sin(phi) * sin(theta) * r |
| 122 |
z = cos(theta) * r |
| 123 |
|
| 124 |
return transpose([x, y, z]) |
| 125 |
|
| 126 |
def sample_from_histogram(p, n_samples=1): |
| 127 |
""" |
| 128 |
returns the indice of bin according to the histogram p |
| 129 |
|
| 130 |
@param p: histogram |
| 131 |
@type p: numpy.array |
| 132 |
@param n_samples: number of samples to generate |
| 133 |
@type n_samples: integer |
| 134 |
""" |
| 135 |
|
| 136 |
from numpy import add, less, argsort, take, arange |
| 137 |
from numpy.random import random |
| 138 |
|
| 139 |
indices = argsort(p) |
| 140 |
indices = take(indices, arange(len(p) - 1, -1, -1)) |
| 141 |
|
| 142 |
c = add.accumulate(take(p, indices)) / add.reduce(p) |
| 143 |
|
| 144 |
return indices[add.reduce(less.outer(c, random(n_samples)), 0)] |
| 145 |
|
| 146 |
def gen_inv_gaussian(a, b, p, burnin=10): |
| 147 |
""" |
| 148 |
Sampler based on Gibbs sampling. |
| 149 |
Assumes scalar p. |
| 150 |
""" |
| 151 |
from numpy.random import gamma |
| 152 |
from numpy import sqrt |
| 153 |
|
| 154 |
s = a * 0. + 1. |
| 155 |
|
| 156 |
if p < 0: |
| 157 |
a, b = b, a |
| 158 |
|
| 159 |
for i in range(burnin): |
| 160 |
|
| 161 |
l = b + 2 * s |
| 162 |
m = sqrt(l / a) |
| 163 |
|
| 164 |
x = inv_gaussian(m, l, shape=m.shape) |
| 165 |
s = gamma(abs(p) + 0.5, x) |
| 166 |
|
| 167 |
if p >= 0: |
| 168 |
return x |
| 169 |
else: |
| 170 |
return 1 / x |
| 171 |
|
| 172 |
def inv_gaussian(mu=1., _lambda=1., shape=None): |
| 173 |
""" |
| 174 |
Generate random samples from inverse gaussian. |
| 175 |
""" |
| 176 |
from numpy.random import standard_normal, random |
| 177 |
from numpy import sqrt, less_equal, clip |
| 178 |
|
| 179 |
mu_2l = mu / _lambda / 2. |
| 180 |
Y = mu * standard_normal(shape) ** 2 |
| 181 |
X = mu + mu_2l * (Y - sqrt(4 * _lambda * Y + Y ** 2)) |
| 182 |
U = random(shape) |
| 183 |
|
| 184 |
m = less_equal(U, mu / (mu + X)) |
| 185 |
|
| 186 |
return clip(m * X + (1 - m) * mu ** 2 / X, 1e-308, 1e308) |
| 187 |
|
| 188 |
def random_rotation(A, n_iter=10, initial_values=None): |
| 189 |
""" |
| 190 |
Generation of three-dimensional random rotations in |
| 191 |
fitting and matching problems, Habeck 2009. |
| 192 |
|
| 193 |
Generate random rotation R from:: |
| 194 |
|
| 195 |
exp(trace(dot(transpose(A), R))) |
| 196 |
|
| 197 |
@param A: generating parameter |
| 198 |
@type A: 3 x 3 numpy array |
| 199 |
|
| 200 |
@param n_iter: number of gibbs sampling steps |
| 201 |
@type n_iter: integer |
| 202 |
|
| 203 |
@param initial_values: initial euler angles alpha, beta and gamma |
| 204 |
@type initial_values: tuple |
| 205 |
|
| 206 |
@rtype: 3 x 3 numpy array |
| 207 |
""" |
| 208 |
from numpy import cos, sin, dot, pi, clip |
| 209 |
from numpy.linalg import svd, det |
| 210 |
from random import vonmisesvariate, randint |
| 211 |
from csb.numeric import euler |
| 212 |
|
| 213 |
|
| 214 |
def sample_beta(kappa, n=1): |
| 215 |
from numpy import arccos |
| 216 |
from csb.numeric import log, exp |
| 217 |
from numpy.random import random |
| 218 |
|
| 219 |
u = random(n) |
| 220 |
|
| 221 |
if kappa != 0.: |
| 222 |
x = clip(1 + 2 * log(u + (1 - u) * exp(-kappa)) / kappa, -1., 1.) |
| 223 |
else: |
| 224 |
x = 2 * u - 1 |
| 225 |
|
| 226 |
if n == 1: |
| 227 |
return arccos(x)[0] |
| 228 |
else: |
| 229 |
return arccos(x) |
| 230 |
|
| 231 |
|
| 232 |
U, L, V = svd(A) |
| 233 |
|
| 234 |
if det(U) < 0: |
| 235 |
L[2] *= -1 |
| 236 |
U[:, 2] *= -1 |
| 237 |
if det(V) < 0: |
| 238 |
L[2] *= -1 |
| 239 |
V[2] *= -1 |
| 240 |
|
| 241 |
if initial_values is None: |
| 242 |
beta = 0. |
| 243 |
else: |
| 244 |
alpha, beta, gamma = initial_values |
| 245 |
|
| 246 |
for _i in range(n_iter): |
| 247 |
|
| 248 |
## sample alpha and gamma |
| 249 |
phi = vonmisesvariate(0., clip(cos(beta / 2) ** 2 * (L[0] + L[1]), 1e-308, 1e10)) |
| 250 |
psi = vonmisesvariate(pi, sin(beta / 2) ** 2 * (L[0] - L[1])) |
| 251 |
u = randint(0, 1) |
| 252 |
|
| 253 |
alpha = 0.5 * (phi + psi) + pi * u |
| 254 |
gamma = 0.5 * (phi - psi) + pi * u |
| 255 |
|
| 256 |
## sample beta |
| 257 |
kappa = cos(phi) * (L[0] + L[1]) + cos(psi) * (L[0] - L[1]) + 2 * L[2] |
| 258 |
beta = sample_beta(kappa) |
| 259 |
|
| 260 |
return dot(U, dot(euler(alpha, beta, gamma), V)) |