1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514
|
# /usr/bin/python
# Last Change: Thu Nov 16 02:00 PM 2006 J
# TODO:
# - which methods to avoid va shrinking to 0 ? There are several options,
# not sure which ones are appropriates
# - improve EM trainer
# - online EM
import numpy as N
import numpy.linalg as lin
from numpy.random import randn
#import _c_densities as densities
import densities
from kmean import kmean
from gauss_mix import GM
from misc import _DEF_ALPHA, _MIN_DBL_DELTA, _MIN_INV_COND
# Error classes
class GmmError(Exception):
"""Base class for exceptions in this module."""
pass
class GmmParamError:
"""Exception raised for errors in gmm params
Attributes:
expression -- input expression in which the error occurred
message -- explanation of the error
"""
def __init__(self, message):
self.message = message
def __str__(self):
return self.message
# Not sure yet about how to design different mixture models. Most of the code
# is different # (pdf, update part of EM, etc...) and I am not sure it makes
# sense to use inheritance for # interface specification in python, since its
# dynamic type systeme.
# Anyway, a mixture model class should encapsulates all details
# concerning getting sufficient statistics (SS), likelihood and bic.
class MixtureModel(object):
pass
class ExpMixtureModel(MixtureModel):
"""Class to model mixture of exponential pdf (eg Gaussian, exponential, Laplace,
etc..). This is a special case because some parts of EM are common for those
models..."""
pass
class GMM(ExpMixtureModel):
""" A class to model a Gaussian Mixture Model (GMM). An instance of
this class is created by giving weights, mean and variances in the ctor.
An instanciated object can be sampled, trained by EM.
The class method gen_model can be used without instanciation."""
def init_kmean(self, data, niter = 5):
""" Init the model with kmean."""
k = self.gm.k
d = self.gm.d
init = data[0:k, :]
(code, label) = kmean(data, init, niter)
w = N.ones(k) / k
mu = code.copy()
if self.gm.mode == 'diag':
va = N.zeros((k, d))
for i in range(k):
for j in range(d):
va[i,j] = N.cov(data[N.where(label==i), j], rowvar = 0)
elif self.gm.mode == 'full':
va = N.zeros((k*d, d))
for i in range(k):
va[i*d:i*d+d,:] = \
N.cov(data[N.where(label==i)], rowvar = 0)
else:
raise GmmParamError("mode " + str(mode) + " not recognized")
self.gm.set_param(w, mu, va)
self.isinit = True
def init_random(self, data):
""" Init the model at random."""
k = self.gm.k
d = self.gm.d
if self.gm.mode == 'diag':
w = N.ones(k) / k
mu = randn(k, d)
va = N.fabs(randn(k, d))
else:
raise GmmParamError("""init_random not implemented for
mode %s yet""", mode)
self.gm.set_param(w, mu, va)
self.isinit = True
# TODO:
# - format of parameters ? For variances, list of variances matrix,
# keep the current format, have 3d matrices ?
# - To handle the different modes, we could do something "fancy" such as
# replacing methods, to avoid checking cases everywhere and unconsistency.
def __init__(self, gm, init = 'kmean'):
""" Initialize a GMM with weight w, mean mu and variances va, and initialization
method for training init (kmean by default)"""
self.gm = gm
# Possible init methods
init_methods = {'kmean': self.init_kmean, 'random' : self.init_random}
if init not in init_methods:
raise GmmParamError('init method %s not recognized' + str(init))
self.init = init_methods[init]
self.isinit = False
self.initst = init
def sufficient_statistics(self, data):
""" Return normalized and non-normalized sufficient statistics
from the model.
Computes the latent variable distribution (a
posteriori probability) knowing the explicit data
for the Gaussian model (w, mu, var): gamma(t, i) =
P[state = i | observation = data(t); w, mu, va]
This is basically the E step of EM for GMM."""
n = data.shape[0]
# compute the gaussian pdf
tgd = multiple_gauss_den(data, self.gm.mu, self.gm.va)
# multiply by the weight
tgd *= self.gm.w
# Normalize to get a pdf
gd = tgd / N.sum(tgd, axis=1)[:, N.newaxis]
return gd, tgd
def update_em(self, data, gamma):
"""Computes update of the Gaussian Mixture Model (M step)
from the a posteriori pdf, computed by gmm_posterior
(E step).
"""
k = self.gm.k
d = self.gm.d
n = data.shape[0]
invn = 1.0/n
mGamma = N.sum(gamma, axis = 0)
if self.gm.mode == 'diag':
mu = N.zeros((k, d))
va = N.zeros((k, d))
gamma = gamma.T
for c in range(k):
x = N.dot(gamma[c:c+1,:], data)[0,:]
xx = N.dot(gamma[c:c+1,:], data ** 2)[0,:]
mu[c,:] = x / mGamma[c]
va[c,:] = xx / mGamma[c] - mu[c,:] ** 2
w = invn * mGamma
elif self.gm.mode == 'full':
# In full mode, this is the bottleneck: the triple loop
# kills performances. This is pretty straightforward
# algebra, so computing it in C should not be too difficult. The
# real problem is to have valid covariance matrices, and to keep
# them positive definite, maybe with special storage... Not sure
# it really worth the risk
mu = N.zeros((k, d))
va = N.zeros((k*d, d))
gamma = gamma.transpose()
for c in range(k):
#x = N.sum(N.outer(gamma[:, c],
# N.ones((1, d))) * data, axis = 0)
x = N.dot(gamma[c:c+1,:], data)[0,:]
xx = N.zeros((d, d))
# This should be much faster than recursing on n...
for i in range(d):
for j in range(d):
xx[i,j] = N.sum(data[:,i] * data[:,j] * gamma[c,:], axis = 0)
mu[c,:] = x / mGamma[c]
va[c*d:c*d+d,:] = xx / mGamma[c] - \
N.outer(mu[c,:], mu[c,:])
w = invn * mGamma
else:
raise GmmParamError("varmode not recognized")
self.gm.set_param(w, mu, va)
def likelihood(self, data):
""" Returns the current log likelihood of the model given
the data """
assert(self.isinit)
# compute the gaussian pdf
tgd = multiple_gauss_den(data, self.gm.mu, self.gm.va)
# multiply by the weight
tgd *= self.gm.w
return N.sum(N.log(N.sum(tgd, axis = 1)), axis = 0)
def bic(self, data):
""" Returns the BIC (Bayesian Information Criterion),
also called Schwarz information criterion. Can be used
to choose between different models which have different
number of clusters. The BIC is defined as:
BIC = 2 * ln(L) - k * ln(n)
where:
* ln(L) is the log-likelihood of the estimated model
* k is the number of degrees of freedom
* n is the number of frames
Not that depending on the literature, BIC may be defined as the opposite
of the definition given here. """
if self.gm.mode == 'diag':
""" for a diagonal model, we have
k - 1 (k weigths, but one constraint of normality)
+ k * d (means) + k * d (variances) """
free_deg = self.gm.k * (self.gm.d * 2 + 1) - 1
elif self.gm.mode == 'full':
""" for a full model, we have
k - 1 (k weigths, but one constraint of normality)
+ k * d (means) + k * d * d / 2 (each covariance matrice
has d **2 params, but with positivity constraint) """
if self.gm.d == 1:
free_deg = self.gm.k * 3 - 1
else:
free_deg = self.gm.k * (self.gm.d + 1 + self.gm.d ** 2 / 2) - 1
lk = self.likelihood(data)
n = N.shape(data)[0]
return bic(lk, free_deg, n)
# syntactic sugar
def __repr__(self):
repre = ""
repre += "Gaussian Mixture Model\n"
repre += " -> initialized by %s\n" % str(self.initst)
repre += self.gm.__repr__()
return repre
class EM:
"""An EM trainer. An EM trainer
trains from data, with a model
Not really useful yet"""
def __init__(self):
pass
def train(self, data, model, maxiter = 10, thresh = 1e-5):
"""
Train a model using data, and stops when the likelihood fails
behind a threshold, or when the number of iterations > niter,
whichever comes first
Args:
- data: contains the observed features, one row is one frame, ie one
observation of dimension d
- model: object of class Mixture
- maxiter: maximum number of iterations
The model is trained, and its parameters updated accordingly.
Returns:
likelihood (one value per iteration).
"""
if not isinstance(model, MixtureModel):
raise TypeError("expect a MixtureModel as a model")
# Initialize the data (may do nothing depending on the model)
model.init(data)
# Likelihood is kept
like = N.zeros(maxiter)
# Em computation, with computation of the likelihood
g, tgd = model.sufficient_statistics(data)
like[0] = N.sum(N.log(N.sum(tgd, 1)), axis = 0)
model.update_em(data, g)
for i in range(1, maxiter):
g, tgd = model.sufficient_statistics(data)
like[i] = N.sum(N.log(N.sum(tgd, 1)), axis = 0)
model.update_em(data, g)
if has_em_converged(like[i], like[i-1], thresh):
return like[0:i]
# # Em computation, with computation of the likelihood
# g, tgd = model.sufficient_statistics(data)
# like[0] = N.sum(N.log(N.sum(tgd, 1)), axis = 0)
# model.update_em(data, g)
# for i in range(1, maxiter):
# print "=== Iteration %d ===" % i
# isreg = False
# for j in range(model.gm.k):
# va = model.gm.va[j]
# if va.any() < _MIN_INV_COND:
# isreg = True
# print "\tregularization detected"
# print "\t" + str(va)
# model.gm.va[j] = regularize_diag(va)
# print "\t" + str(va) + ", " + str(model.gm.va[j])
# print "\t" + str(gauss_den(data, model.gm.mu[j], model.gm.va[j]))
# print "\tend regularization detected"
# var = va
#
# g, tgd = model.sufficient_statistics(data)
# try:
# assert not( (N.isnan(tgd)).any() )
# if isreg:
# print var
# except AssertionError:
# print "tgd is nan..."
# print model.gm.va[13,:]
# print 1/model.gm.va[13,:]
# print densities.gauss_den(data, model.gm.mu[13], model.gm.va[13])
# print N.isnan((multiple_gauss_den(data, model.gm.mu, model.gm.va))).any()
# print "Exciting"
# import sys
# sys.exit(-1)
# like[i] = N.sum(N.log(N.sum(tgd, 1)), axis = 0)
# model.update_em(data, g)
# assert not( model.gm.va.any() < 1e-6)
# if has_em_converged(like[i], like[i-1], thresh):
# return like[0:i]
return like
def regularize_diag(variance, alpha = _DEF_ALPHA):
delta = N.sum(variance) / variance.size
if delta > _MIN_DBL_DELTA:
return variance + alpha * delta
else:
return variance + alpha * _MIN_DBL_DELTA
def regularize_full(variance):
# Trace of a positive definite matrix is always > 0
delta = N.trace(variance) / variance.shape[0]
if delta > _MIN_DBL_DELTA:
return variance + alpha * delta
else:
return variance + alpha * _MIN_DBL_DELTA
# Misc functions
def bic(lk, deg, n):
""" Expects lk to be log likelihood """
return 2 * lk - deg * N.log(n)
def has_em_converged(like, plike, thresh):
""" given likelihood of current iteration like and previous
iteration plike, returns true is converged: based on comparison
of the slope of the likehood with thresh"""
diff = N.abs(like - plike)
avg = 0.5 * (N.abs(like) + N.abs(plike))
if diff / avg < thresh:
return True
else:
return False
def multiple_gauss_den(data, mu, va):
"""Helper function to generate several Gaussian
pdf (different parameters) from the same data"""
mu = N.atleast_2d(mu)
va = N.atleast_2d(va)
K = mu.shape[0]
n = data.shape[0]
d = mu.shape[1]
y = N.zeros((K, n))
if mu.size == va.size:
for i in range(K):
y[i] = densities.gauss_den(data, mu[i, :], va[i, :])
return y.T
else:
for i in range(K):
y[i] = densities.gauss_den(data, mu[i, :],
va[d*i:d*i+d, :])
return y.T
if __name__ == "__main__":
import copy
#=============================
# Simple GMM with 5 components
#=============================
#+++++++++++++++++++++++++++++
# Meta parameters of the model
# - k: Number of components
# - d: dimension of each Gaussian
# - mode: Mode of covariance matrix: full or diag
# - nframes: number of frames (frame = one data point = one
# row of d elements
k = 2
d = 1
mode = 'full'
nframes = 1e3
#+++++++++++++++++++++++++++++++++++++++++++
# Create an artificial GMM model, samples it
#+++++++++++++++++++++++++++++++++++++++++++
print "Generating the mixture"
# Generate a model with k components, d dimensions
w, mu, va = GM.gen_param(d, k, mode, spread = 3)
gm = GM(d, k, mode)
gm.set_param(w, mu, va)
# Sample nframes frames from the model
data = gm.sample(nframes)
#++++++++++++++++++++++++
# Learn the model with EM
#++++++++++++++++++++++++
# Init the model
print "Init a model for learning, with kmean for initialization"
lgm = GM(d, k, mode)
gmm = GMM(lgm, 'kmean')
gmm.init(data)
# Keep the initialized model for drawing
gm0 = copy.copy(lgm)
# The actual EM, with likelihood computation
niter = 10
like = N.zeros(niter)
print "computing..."
for i in range(niter):
g, tgd = gmm.sufficient_statistics(data)
like[i] = N.sum(N.log(N.sum(tgd, 1)), axis = 0)
gmm.update_em(data, g)
# # Alternative form, by using EM class: as the EM class
# # is quite rudimentary now, it is not very useful, just save
# # a few lines
# em = EM()
# like = em.train(data, gmm, niter)
#+++++++++++++++
# Draw the model
#+++++++++++++++
print "drawing..."
import pylab as P
P.subplot(2, 1, 1)
if not d == 1:
# Draw what is happening
P.plot(data[:, 0], data[:, 1], '.', label = '_nolegend_')
# Real confidence ellipses
Xre, Yre = gm.conf_ellipses()
P.plot(Xre[0], Yre[0], 'g', label = 'true confidence ellipsoides')
for i in range(1,k):
P.plot(Xre[i], Yre[i], 'g', label = '_nolegend_')
# Initial confidence ellipses as found by kmean
X0e, Y0e = gm0.conf_ellipses()
P.plot(X0e[0], Y0e[0], 'k', label = 'initial confidence ellipsoides')
for i in range(1,k):
P.plot(X0e[i], Y0e[i], 'k', label = '_nolegend_')
# Values found by EM
Xe, Ye = lgm.conf_ellipses()
P.plot(Xe[0], Ye[0], 'r', label = 'confidence ellipsoides found by EM')
for i in range(1,k):
P.plot(Xe[i], Ye[i], 'r', label = '_nolegend_')
P.legend(loc = 0)
else:
# Real confidence ellipses
h = gm.plot1d()
[i.set_color('g') for i in h['pdf']]
h['pdf'][0].set_label('true pdf')
# Initial confidence ellipses as found by kmean
h0 = gm0.plot1d()
[i.set_color('k') for i in h0['pdf']]
h0['pdf'][0].set_label('initial pdf')
# Values found by EM
hl = lgm.plot1d(fill = 1, level = 0.66)
[i.set_color('r') for i in hl['pdf']]
hl['pdf'][0].set_label('pdf found by EM')
P.legend(loc = 0)
P.subplot(2, 1, 2)
P.plot(like)
P.title('log likelihood')
# #++++++++++++++++++
# # Export the figure
# #++++++++++++++++++
# F = P.gcf()
# DPI = F.get_dpi()
# DefaultSize = F.get_size_inches()
# # the default is 100dpi for savefig:
# F.savefig("example1.png")
# # Now make the image twice as big, while keeping the fonts and all the
# # same size
# F.set_figsize_inches( (DefaultSize[0]*2, DefaultSize[1]*2) )
# Size = F.get_size_inches()
# print "Size in Inches", Size
# F.savefig("example2.png")
P.show()
|