| 1 |
""" |
| 2 |
Approximation of a distribution as a mixture of gaussians with a zero mean but different sigmas |
| 3 |
""" |
| 4 |
|
| 5 |
import numpy.random |
| 6 |
|
| 7 |
import csb.core |
| 8 |
import csb.statistics.ars |
| 9 |
import csb.statistics.rand |
| 10 |
|
| 11 |
from csb.core import typedproperty |
| 12 |
from abc import abstractmethod, ABCMeta |
| 13 |
|
| 14 |
from csb.numeric import log, exp, approx_psi, inv_psi, d_approx_psi |
| 15 |
from scipy.special import psi, kve |
| 16 |
from csb.statistics import harmonic_mean, geometric_mean |
| 17 |
from csb.statistics.pdf import AbstractEstimator, AbstractDensity, BaseDensity |
| 18 |
from csb.statistics.pdf import Gamma, InverseGamma, NullEstimator |
| 19 |
|
| 20 |
|
| 21 |
def inv_digamma_minus_log(y, tol=1e-10, n_iter=100): |
| 22 |
""" |
| 23 |
Solve y = psi(alpha) - log(alpha) for alpha by fixed point |
| 24 |
integration. |
| 25 |
""" |
| 26 |
if y >= -log(6.): |
| 27 |
x = 1 / (2 * (1 - exp(y))) |
| 28 |
else: |
| 29 |
x = 1.e-10 |
| 30 |
for _i in range(n_iter): |
| 31 |
z = approx_psi(x) - log(x) - y |
| 32 |
if abs(z) < tol: |
| 33 |
break |
| 34 |
x -= x * z / (x * d_approx_psi(x) - 1) |
| 35 |
x = abs(x) |
| 36 |
return x |
| 37 |
|
| 38 |
class ScaleMixturePriorEstimator(AbstractEstimator): |
| 39 |
""" |
| 40 |
Prior on the scales of a L{ScaleMixture}, which determines how the scales |
| 41 |
are estimated. |
| 42 |
""" |
| 43 |
|
| 44 |
__metaclass__ = ABCMeta |
| 45 |
|
| 46 |
@abstractmethod |
| 47 |
def get_scale_estimator(self): |
| 48 |
""" |
| 49 |
Return an appropriate estimator for the scales of the mixture distribution |
| 50 |
under this prior. |
| 51 |
""" |
| 52 |
pass |
| 53 |
|
| 54 |
|
| 55 |
class ScaleMixturePrior(object): |
| 56 |
""" |
| 57 |
Prior on the scales of a L{ScaleMixture}, which determines how the scales |
| 58 |
are estimated. |
| 59 |
""" |
| 60 |
|
| 61 |
def __init__(self, *args): |
| 62 |
super(ScaleMixturePrior, self).__init__(*args) |
| 63 |
self._scale = None |
| 64 |
self._scale_estimator = NullEstimator() |
| 65 |
|
| 66 |
@property |
| 67 |
def scale_estimator(self): |
| 68 |
return self._scale_estimator |
| 69 |
|
| 70 |
@property |
| 71 |
def estimator(self): |
| 72 |
return self._estimator |
| 73 |
|
| 74 |
@estimator.setter |
| 75 |
def estimator(self, strategy): |
| 76 |
if not isinstance(strategy, AbstractEstimator): |
| 77 |
raise TypeError(strategy) |
| 78 |
self._estimator = strategy |
| 79 |
if isinstance(strategy, ScaleMixturePriorEstimator): |
| 80 |
self._scale_estimator = strategy.get_scale_estimator() |
| 81 |
else: |
| 82 |
self._scale_estimator = NullEstimator() |
| 83 |
|
| 84 |
|
| 85 |
class ScaleMixture(BaseDensity): |
| 86 |
""" |
| 87 |
Robust probabilistic superposition and comparison of protein structures |
| 88 |
Martin Mechelke and Michael Habeck |
| 89 |
|
| 90 |
Represenation of a distribution as a mixture of gaussians with a mean of |
| 91 |
zero and different inverse variances/scales. The number of scales equals |
| 92 |
the number of datapoints. |
| 93 |
|
| 94 |
The underlying family is determined by a prior L{ScaleMixturePrior} on the |
| 95 |
scales. Choosing a L{GammaPrior} results in Stundent-t posterior, wheras a |
| 96 |
L{InvGammaPrior} leads to a K-Distribution as posterior. |
| 97 |
""" |
| 98 |
|
| 99 |
def __init__(self, scales=numpy.array([1., 1.]), prior=None, d=3): |
| 100 |
|
| 101 |
super(ScaleMixture, self).__init__() |
| 102 |
|
| 103 |
self._register('scales') |
| 104 |
|
| 105 |
self._d = d |
| 106 |
self.set_params(scales=scales) |
| 107 |
self._prior = prior |
| 108 |
|
| 109 |
if self._prior is not None: |
| 110 |
self.estimator = self._prior.scale_estimator |
| 111 |
|
| 112 |
@property |
| 113 |
def scales(self): |
| 114 |
return numpy.squeeze(self['scales']) |
| 115 |
|
| 116 |
@scales.setter |
| 117 |
def scales(self, value): |
| 118 |
if not isinstance(value, csb.core.string) and \ |
| 119 |
isinstance(value, (numpy.ndarray, list, tuple)): |
| 120 |
self['scales'] = numpy.array(value) |
| 121 |
else: |
| 122 |
raise ValueError("numpy array expected") |
| 123 |
|
| 124 |
@property |
| 125 |
def prior(self): |
| 126 |
return self._prior |
| 127 |
|
| 128 |
@prior.setter |
| 129 |
def prior(self, value): |
| 130 |
if not isinstance(value, ScaleMixturePrior): |
| 131 |
raise TypeError(value) |
| 132 |
self._prior = value |
| 133 |
self.estimator = self._prior.scale_estimator |
| 134 |
|
| 135 |
def log_prob(self, x): |
| 136 |
from csb.numeric import log_sum_exp |
| 137 |
|
| 138 |
dim = self._d |
| 139 |
s = self.scales |
| 140 |
|
| 141 |
log_p = numpy.squeeze(-numpy.multiply.outer(x * x, 0.5 * s)) + \ |
| 142 |
numpy.squeeze(dim * 0.5 * (log(s) - log(2 * numpy.pi))) |
| 143 |
|
| 144 |
if self._prior is not None: |
| 145 |
log_p += numpy.squeeze(self._prior.log_prob(s)) |
| 146 |
return log_sum_exp(log_p.T, 0) |
| 147 |
|
| 148 |
|
| 149 |
def random(self, shape=None): |
| 150 |
|
| 151 |
s = self.scales |
| 152 |
|
| 153 |
if shape is None: |
| 154 |
return numpy.random.normal() * s[numpy.random.randint(len(s))] |
| 155 |
|
| 156 |
else: |
| 157 |
#n = s.shape[0] |
| 158 |
nrv = numpy.random.normal(size=shape) |
| 159 |
indices = numpy.random.randint(len(s), size=shape) |
| 160 |
|
| 161 |
return s[indices] * nrv |
| 162 |
|
| 163 |
|
| 164 |
class ARSPosteriorAlpha(csb.statistics.ars.LogProb): |
| 165 |
""" |
| 166 |
This class represents the posterior distribution of the alpha parameter |
| 167 |
of the Gamma and Inverse Gamma prior, and allows sampling using adaptive |
| 168 |
rejection sampling L{ARS}. |
| 169 |
""" |
| 170 |
|
| 171 |
def __init__(self, a, b, n): |
| 172 |
|
| 173 |
self.a = float(a) |
| 174 |
self.b = float(b) |
| 175 |
self.n = float(n) |
| 176 |
|
| 177 |
def __call__(self, x): |
| 178 |
|
| 179 |
from scipy.special import gammaln |
| 180 |
|
| 181 |
return self.a * x - \ |
| 182 |
self.n * gammaln(numpy.clip(x, 1e-308, 1e308)) + \ |
| 183 |
self.b * log(x), \ |
| 184 |
self.a - self.n * psi(x) + self.b / x |
| 185 |
|
| 186 |
def initial_values(self, tol=1e-10): |
| 187 |
""" |
| 188 |
Generate initial values by doing fixed point |
| 189 |
iterations to solve for alpha |
| 190 |
""" |
| 191 |
n, a, b = self.n, self.a, self.b |
| 192 |
|
| 193 |
if abs(b) < 1e-10: |
| 194 |
alpha = inv_psi(a / n) |
| 195 |
else: |
| 196 |
alpha = 1. |
| 197 |
|
| 198 |
z = tol + 1. |
| 199 |
|
| 200 |
while abs(z) > tol: |
| 201 |
|
| 202 |
z = n * psi(alpha) - \ |
| 203 |
b / numpy.clip(alpha, 1e-300, 1e300) - a |
| 204 |
|
| 205 |
alpha -= z / (n * d_approx_psi(alpha) - b |
| 206 |
/ (alpha ** 2 + 1e-300)) |
| 207 |
alpha = numpy.clip(alpha, 1e-100, 1e300) |
| 208 |
|
| 209 |
return numpy.clip(alpha - 1 / (n + 1e-300), 1e-100, 1e300), \ |
| 210 |
alpha + 1 / (n + 1e-300), alpha |
| 211 |
|
| 212 |
|
| 213 |
class GammaPosteriorSampler(ScaleMixturePriorEstimator): |
| 214 |
|
| 215 |
def __init__(self): |
| 216 |
super(GammaPosteriorSampler, self).__init__() |
| 217 |
self.n_samples = 2 |
| 218 |
|
| 219 |
def get_scale_estimator(self): |
| 220 |
return GammaScaleSampler() |
| 221 |
|
| 222 |
def estimate(self, context, data): |
| 223 |
""" |
| 224 |
Generate samples from the posterior of alpha and beta. |
| 225 |
|
| 226 |
For beta the posterior is a gamma distribution and analytically |
| 227 |
acessible. |
| 228 |
|
| 229 |
The posterior of alpha can not be expressed analytically and is |
| 230 |
aproximated using adaptive rejection sampling. |
| 231 |
""" |
| 232 |
pdf = GammaPrior() |
| 233 |
|
| 234 |
## sufficient statistics |
| 235 |
|
| 236 |
a = numpy.mean(data) |
| 237 |
b = exp(numpy.mean(log(data))) |
| 238 |
v = numpy.std(data) ** 2 |
| 239 |
n = len(data) |
| 240 |
|
| 241 |
beta = a / v |
| 242 |
alpha = beta * a |
| 243 |
samples = [] |
| 244 |
|
| 245 |
for _i in range(self.n_samples): |
| 246 |
|
| 247 |
## sample beta from Gamma distribution |
| 248 |
beta = numpy.random.gamma(n * alpha + context._hyper_beta.alpha, |
| 249 |
1 / (n * a + context._hyper_beta.beta)) |
| 250 |
|
| 251 |
## sample alpha with ARS |
| 252 |
logp = ARSPosteriorAlpha(n * log(beta * b)\ |
| 253 |
- context.hyper_alpha.beta, |
| 254 |
context.hyper_alpha.alpha - 1., n) |
| 255 |
ars = csb.statistics.ars.ARS(logp) |
| 256 |
ars.initialize(logp.initial_values()[:2], z0=0.) |
| 257 |
alpha = ars.sample() |
| 258 |
|
| 259 |
if alpha is None: |
| 260 |
raise ValueError("ARS failed") |
| 261 |
|
| 262 |
samples.append((alpha, beta)) |
| 263 |
|
| 264 |
pdf.alpha, pdf.beta = samples[-1] |
| 265 |
|
| 266 |
return pdf |
| 267 |
|
| 268 |
class GammaPosteriorMAP(ScaleMixturePriorEstimator): |
| 269 |
|
| 270 |
def __init__(self): |
| 271 |
super(GammaPosteriorMAP, self).__init__() |
| 272 |
|
| 273 |
def get_scale_estimator(self): |
| 274 |
return GammaScaleMAP() |
| 275 |
|
| 276 |
def estimate(self, context, data): |
| 277 |
""" |
| 278 |
Estimate alpha and beta from their posterior |
| 279 |
""" |
| 280 |
|
| 281 |
pdf = GammaPrior() |
| 282 |
|
| 283 |
s = data[0].mean() |
| 284 |
y = data[1].mean() - log(s) - 1. |
| 285 |
|
| 286 |
alpha = abs(inv_digamma_minus_log(numpy.clip(y, |
| 287 |
- 1e308, |
| 288 |
- 1e-300))) |
| 289 |
beta = alpha / s |
| 290 |
|
| 291 |
|
| 292 |
pdf.alpha, pdf.beta = alpha, beta |
| 293 |
return pdf |
| 294 |
|
| 295 |
|
| 296 |
|
| 297 |
|
| 298 |
class InvGammaPosteriorSampler(ScaleMixturePriorEstimator): |
| 299 |
|
| 300 |
def __init__(self): |
| 301 |
super(InvGammaPosteriorSampler, self).__init__() |
| 302 |
self.n_samples = 2 |
| 303 |
|
| 304 |
def get_scale_estimator(self): |
| 305 |
return InvGammaScaleSampler() |
| 306 |
|
| 307 |
def estimate(self, context, data): |
| 308 |
""" |
| 309 |
Generate samples from the posterior of alpha and beta. |
| 310 |
|
| 311 |
For beta the posterior is a gamma distribution and analytically |
| 312 |
acessible. |
| 313 |
|
| 314 |
The posterior of alpha can not be expressed analytically and is |
| 315 |
aproximated using adaptive rejection sampling. |
| 316 |
""" |
| 317 |
pdf = GammaPrior() |
| 318 |
|
| 319 |
## sufficient statistics |
| 320 |
|
| 321 |
h = harmonic_mean(numpy.clip(data, 1e-308, 1e308)) |
| 322 |
g = geometric_mean(numpy.clip(data, 1e-308, 1e308)) |
| 323 |
|
| 324 |
n = len(data) |
| 325 |
|
| 326 |
samples = [] |
| 327 |
|
| 328 |
a = numpy.mean(1 / data) |
| 329 |
v = numpy.std(1 / data) ** 2 |
| 330 |
|
| 331 |
beta = a / v |
| 332 |
alpha = beta * a |
| 333 |
|
| 334 |
for i in range(self.n_samples): |
| 335 |
|
| 336 |
## sample alpha with ARS |
| 337 |
|
| 338 |
logp = ARSPosteriorAlpha(n * (log(beta) - log(g)) - context.hyper_alpha.beta, |
| 339 |
context.hyper_alpha.alpha - 1., n) |
| 340 |
ars = csb.statistics.ars.ARS(logp) |
| 341 |
ars.initialize(logp.initial_values()[:2], z0=0.) |
| 342 |
|
| 343 |
alpha = numpy.abs(ars.sample()) |
| 344 |
|
| 345 |
if alpha is None: |
| 346 |
raise ValueError("Sampling failed") |
| 347 |
|
| 348 |
|
| 349 |
## sample beta from Gamma distribution |
| 350 |
|
| 351 |
beta = numpy.random.gamma(n * alpha + context.hyper_beta.alpha, \ |
| 352 |
1 / (n / h + context.hyper_beta.beta)) |
| 353 |
|
| 354 |
samples.append((alpha, beta)) |
| 355 |
|
| 356 |
pdf.alpha, pdf.beta = samples[-1] |
| 357 |
return pdf |
| 358 |
|
| 359 |
|
| 360 |
class InvGammaPosteriorMAP(ScaleMixturePriorEstimator): |
| 361 |
|
| 362 |
def __init__(self): |
| 363 |
super(InvGammaPosteriorMAP, self).__init__() |
| 364 |
|
| 365 |
def get_scale_estimator(self): |
| 366 |
return InvGammaScaleMAP() |
| 367 |
|
| 368 |
def estimate(self, context, data): |
| 369 |
""" |
| 370 |
Generate samples from the posterior of alpha and beta. |
| 371 |
|
| 372 |
For beta the posterior is a gamma distribution and analytically |
| 373 |
acessible. |
| 374 |
|
| 375 |
The posterior of alpha can not be expressed analytically and is |
| 376 |
aproximated using adaptive rejection sampling. |
| 377 |
""" |
| 378 |
pdf = GammaPrior() |
| 379 |
|
| 380 |
|
| 381 |
y = log(data).mean() - log((data ** -1).mean()) |
| 382 |
|
| 383 |
alpha = inv_digamma_minus_log(numpy.clip(y, |
| 384 |
- 1e308, |
| 385 |
- 1e-300)) |
| 386 |
alpha = abs(alpha) |
| 387 |
|
| 388 |
beta = numpy.clip(alpha / |
| 389 |
(data ** (-1)).mean(), |
| 390 |
1e-100, 1e100) |
| 391 |
|
| 392 |
pdf.alpha, pdf.beta = alpha, beta |
| 393 |
return pdf |
| 394 |
|
| 395 |
|
| 396 |
|
| 397 |
class GammaScaleSampler(AbstractEstimator): |
| 398 |
""" |
| 399 |
Sample the scalies given the data |
| 400 |
""" |
| 401 |
|
| 402 |
def estimate(self, context, data): |
| 403 |
pdf = ScaleMixture() |
| 404 |
alpha = context.prior.alpha |
| 405 |
beta = context.prior.beta |
| 406 |
d = context._d |
| 407 |
|
| 408 |
if len(data.shape) == 1: |
| 409 |
data = data[:, numpy.newaxis] |
| 410 |
|
| 411 |
a = alpha + 0.5 * d * len(data.shape) |
| 412 |
b = beta + 0.5 * data.sum(-1) ** 2 |
| 413 |
|
| 414 |
s = numpy.clip(numpy.random.gamma(a, 1. / b), 1e-20, 1e10) |
| 415 |
pdf.scales = s |
| 416 |
|
| 417 |
context.prior.estimate(s) |
| 418 |
pdf.prior = context.prior |
| 419 |
|
| 420 |
return pdf |
| 421 |
|
| 422 |
|
| 423 |
class GammaScaleMAP(AbstractEstimator): |
| 424 |
""" |
| 425 |
MAP estimator of the scales |
| 426 |
""" |
| 427 |
|
| 428 |
def estimate(self, context, data): |
| 429 |
pdf = ScaleMixture() |
| 430 |
alpha = context.prior.alpha |
| 431 |
beta = context.prior.beta |
| 432 |
d = context._d |
| 433 |
|
| 434 |
if len(data.shape) == 1: |
| 435 |
data = data[:, numpy.newaxis] |
| 436 |
|
| 437 |
a = alpha + 0.5 * d * len(data.shape) |
| 438 |
b = beta + 0.5 * data.sum(-1) ** 2 |
| 439 |
|
| 440 |
s = a / b |
| 441 |
log_s = psi(a) - log(b) |
| 442 |
|
| 443 |
pdf.scales = s |
| 444 |
context.prior.estimate([s, log_s]) |
| 445 |
pdf.prior = context.prior |
| 446 |
|
| 447 |
return pdf |
| 448 |
|
| 449 |
|
| 450 |
|
| 451 |
class InvGammaScaleSampler(AbstractEstimator): |
| 452 |
""" |
| 453 |
Sample the scales given the data |
| 454 |
""" |
| 455 |
|
| 456 |
def estimate(self, context, data): |
| 457 |
pdf = ScaleMixture() |
| 458 |
alpha = context.prior.alpha |
| 459 |
beta = context.prior.beta |
| 460 |
d = context._d |
| 461 |
|
| 462 |
if len(data.shape) == 1: |
| 463 |
data = data[:, numpy.newaxis] |
| 464 |
|
| 465 |
p = -alpha + 0.5 * d |
| 466 |
b = 2 * beta |
| 467 |
a = 1e-5 + data.sum(-1) ** 2 |
| 468 |
|
| 469 |
s = csb.statistics.rand.gen_inv_gaussian(a, b, p) |
| 470 |
s = numpy.clip(s, 1e-300, 1e300) |
| 471 |
|
| 472 |
pdf.scales = s |
| 473 |
context.prior.estimate(s) |
| 474 |
pdf.prior = context.prior |
| 475 |
|
| 476 |
return pdf |
| 477 |
|
| 478 |
|
| 479 |
|
| 480 |
class InvGammaScaleMAP(AbstractEstimator): |
| 481 |
""" |
| 482 |
MAP estimator of the scales |
| 483 |
""" |
| 484 |
def estimate(self, context, data): |
| 485 |
pdf = ScaleMixture() |
| 486 |
alpha = context.prior.alpha |
| 487 |
beta = context.prior.beta |
| 488 |
d = context._d |
| 489 |
|
| 490 |
if len(data.shape) == 1: |
| 491 |
data = data[:, numpy.newaxis] |
| 492 |
|
| 493 |
p = -alpha + 0.5 * d |
| 494 |
b = 2 * beta |
| 495 |
a = 1e-5 + data.sum(-1) ** 2 |
| 496 |
|
| 497 |
s = numpy.clip((numpy.sqrt(b) * kve(p + 1, numpy.sqrt(a * b))) |
| 498 |
/ (numpy.sqrt(a) * kve(p, numpy.sqrt(a * b))), |
| 499 |
1e-10, 1e10) |
| 500 |
pdf.scales = s |
| 501 |
context.prior.estimate(s) |
| 502 |
pdf.prior = context.prior |
| 503 |
|
| 504 |
return pdf |
| 505 |
|
| 506 |
|
| 507 |
|
| 508 |
class GammaPrior(ScaleMixturePrior, Gamma): |
| 509 |
""" |
| 510 |
Gamma prior on mixture weights including a weak gamma prior on its parameter. |
| 511 |
""" |
| 512 |
|
| 513 |
def __init__(self, alpha=1., beta=1., hyper_alpha=(4., 1.), |
| 514 |
hyper_beta=(2., 1.)): |
| 515 |
|
| 516 |
super(GammaPrior, self).__init__(alpha, beta) |
| 517 |
|
| 518 |
self._hyper_alpha = Gamma(*hyper_alpha) |
| 519 |
self._hyper_beta = Gamma(*hyper_beta) |
| 520 |
self.estimator = GammaPosteriorSampler() |
| 521 |
|
| 522 |
@typedproperty(AbstractDensity) |
| 523 |
def hyper_beta(): |
| 524 |
pass |
| 525 |
|
| 526 |
@typedproperty(AbstractDensity) |
| 527 |
def hyper_alpha(): |
| 528 |
pass |
| 529 |
|
| 530 |
def log_prob(self, x): |
| 531 |
|
| 532 |
a, b = self['alpha'], self['beta'] |
| 533 |
|
| 534 |
l_a = self._hyper_alpha(a) |
| 535 |
l_b = self._hyper_beta(b) |
| 536 |
|
| 537 |
return super(GammaPrior, self).log_prob(x) + l_a + l_b |
| 538 |
|
| 539 |
|
| 540 |
|
| 541 |
class InvGammaPrior(ScaleMixturePrior, InverseGamma): |
| 542 |
""" |
| 543 |
Inverse Gamma prior on mixture weights including a weak gamma |
| 544 |
prior on its parameter. |
| 545 |
""" |
| 546 |
|
| 547 |
def __init__(self, alpha=1., beta=1., hyper_alpha=(4., 1.), |
| 548 |
hyper_beta=(2., 1.)): |
| 549 |
|
| 550 |
super(InvGammaPrior, self).__init__(alpha, beta) |
| 551 |
|
| 552 |
self._hyper_alpha = Gamma(*hyper_alpha) |
| 553 |
self._hyper_beta = Gamma(*hyper_beta) |
| 554 |
self.estimator = InvGammaPosteriorSampler() |
| 555 |
|
| 556 |
@typedproperty(AbstractDensity) |
| 557 |
def hyper_beta(): |
| 558 |
pass |
| 559 |
|
| 560 |
@typedproperty(AbstractDensity) |
| 561 |
def hyper_alpha(): |
| 562 |
pass |
| 563 |
|
| 564 |
def log_prob(self, x): |
| 565 |
|
| 566 |
a, b = self['alpha'], self['beta'] |
| 567 |
|
| 568 |
l_a = self._hyper_alpha(a) |
| 569 |
l_b = self._hyper_beta(b) |
| 570 |
|
| 571 |
return super(InvGammaPrior, self).log_prob(x) + l_a + l_b |
| 572 |
|
| 573 |
|
| 574 |
|
| 575 |
|