/usr/share/pyshared/mvpa/measures/irelief.py is in python-mvpa 0.4.8-3.
This file is owned by root:root, with mode 0o644.
The actual contents of the file can be viewed below.
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 | # emacs: -*- mode: python; py-indent-offset: 4; indent-tabs-mode: nil -*-
# vi: set ft=python sts=4 ts=4 sw=4 et:
### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ##
#
# Copyright (c) 2008 Emanuele Olivetti <emanuele@relativita.com>
# See COPYING file distributed along with the PyMVPA package for the
# copyright and license terms.
#
### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ##
"""FeaturewiseDatasetMeasure performing multivariate Iterative RELIEF
(I-RELIEF) algorithm.
See : Y. Sun, Iterative RELIEF for Feature Weighting: Algorithms, Theories,
and Applications, IEEE Trans. on Pattern Analysis and Machine Intelligence
(TPAMI), vol. 29, no. 6, pp. 1035-1051, June 2007."""
__docformat__ = 'restructuredtext'
import numpy as N
from mvpa.measures.base import FeaturewiseDatasetMeasure
from mvpa.clfs.kernel import KernelSquaredExponential, KernelExponential, \
KernelMatern_3_2, KernelMatern_5_2
from mvpa.clfs.distance import pnorm_w
if __debug__:
from mvpa.base import debug
class IterativeRelief_Devel(FeaturewiseDatasetMeasure):
"""`FeaturewiseDatasetMeasure` that performs multivariate I-RELIEF
algorithm. Batch version allowing various kernels.
UNDER DEVELOPEMNT.
Batch I-RELIEF-2 feature weighting algorithm. Works for binary or
multiclass class-labels. Batch version with complexity O(T*N^2*I),
where T is the number of iterations, N the number of instances, I
the number of features.
See: Y. Sun, Iterative RELIEF for Feature Weighting: Algorithms,
Theories, and Applications, IEEE Trans. on Pattern Analysis and
Machine Intelligence (TPAMI), vol. 29, no. 6, pp. 1035-1051, June
2007. http://plaza.ufl.edu/sunyijun/Paper/PAMI_1.pdf
Note that current implementation allows to use only
exponential-like kernels. Support for linear kernel will be
added later.
"""
def __init__(self, threshold = 1.0e-2, kernel = None, kernel_width = 1.0,
w_guess = None, **kwargs):
"""Constructor of the IRELIEF class.
"""
# init base classes first
FeaturewiseDatasetMeasure.__init__(self, **kwargs)
# Threshold in W changes (stopping criterion for irelief)
self.threshold = threshold
if kernel == None:
self.kernel = KernelExponential
else:
self.kernel = kernel
self.w_guess = w_guess
self.w = None
self.kernel_width = kernel_width
def compute_M_H(self, label):
"""Compute hit/miss dictionaries.
For each instance compute the set of indices having the same
class label and different class label.
Note that this computation is independent of the number of
features.
"""
M = {}
H = {}
for i in range(label.size):
M[i] = N.where(label != label[i])[0]
tmp = (N.where(label == label[i])[0]).tolist()
tmp.remove(i)
# There must be at least two exampls for class label[i]
assert(tmp != [])
H[i] = N.array(tmp)
return M, H
def _call(self, dataset):
"""Computes featurewise I-RELIEF weights."""
samples = dataset.samples
NS, NF = samples.shape[:2]
if self.w_guess == None:
self.w = N.ones(NF, 'd')
# do normalization in all cases to be safe :)
self.w = self.w/(self.w**2).sum()
M, H = self.compute_M_H(dataset.labels)
while True:
self.k = self.kernel(length_scale = self.kernel_width/self.w)
d_w_k = self.k.compute(samples)
# set d_w_k to zero where distance=0 (i.e. kernel ==
# 1.0), otherwise I-RELIEF could not converge.
# XXX Note that kernel==1 for distance=0 only for
# exponential kernels!! IMPROVE
d_w_k[N.abs(d_w_k-1.0) < 1.0e-15] = 0.0
ni = N.zeros(NF, 'd')
for n in range(NS):
# d_w_k[n,n] could be omitted since == 0.0
gamma_n = 1.0 - N.nan_to_num(d_w_k[n, M[n]].sum() \
/ (d_w_k[n, :].sum()-d_w_k[n, n]))
alpha_n = N.nan_to_num(d_w_k[n, M[n]]/(d_w_k[n, M[n]].sum()))
beta_n = N.nan_to_num(d_w_k[n, H[n]]/(d_w_k[n, H[n]].sum()))
m_n = (N.abs(samples[n, :] - samples[M[n], :]) \
* alpha_n[:, None]).sum(0)
h_n = (N.abs(samples[n, :] - samples[H[n], :]) \
* beta_n[:, None]).sum(0)
ni += gamma_n*(m_n-h_n)
ni = ni/NS
ni_plus = N.clip(ni, 0.0, N.inf) # set all negative elements to zero
w_new = N.nan_to_num(ni_plus/(N.sqrt((ni_plus**2).sum())))
change = N.abs(w_new-self.w).sum()
if __debug__ and 'IRELIEF' in debug.active:
debug('IRELIEF',
"change=%.4f max=%f min=%.4f mean=%.4f std=%.4f #nan=%d"
% (change, w_new.max(), w_new.min(), w_new.mean(),
w_new.std(), N.isnan(w_new).sum()))
# update weights:
self.w = w_new
if change < self.threshold:
break
return self.w
class IterativeReliefOnline_Devel(IterativeRelief_Devel):
"""`FeaturewiseDatasetMeasure` that performs multivariate I-RELIEF
algorithm. Online version.
UNDER DEVELOPMENT
Online version with complexity O(T*N*I),
where N is the number of instances and I the number of features.
See: Y. Sun, Iterative RELIEF for Feature Weighting: Algorithms,
Theories, and Applications, IEEE Trans. on Pattern Analysis and
Machine Intelligence (TPAMI), vol. 29, no. 6, pp. 1035-1051, June
2007. http://plaza.ufl.edu/sunyijun/Paper/PAMI_1.pdf
Note that this implementation is not fully online, since hit and
miss dictionaries (H,M) are computed once at the beginning using
full access to all labels. This can be easily corrected to a full
online implementation. But this is not mandatory now since the
major goal of this current online implementation is reduction of
computational complexity.
"""
def __init__(self, a=5.0, permute=True, max_iter=3, **kwargs):
"""Constructor of the IRELIEF class.
"""
# init base classes first
IterativeRelief_Devel.__init__(self, **kwargs)
self.a = a # parameter of the learning rate
self.permute = permute # shuffle data when running I-RELIEF
self.max_iter = max_iter # maximum number of iterations
def _call(self, dataset):
"""Computes featurewise I-RELIEF-2 weights. Online version."""
NS = dataset.samples.shape[0]
NF = dataset.samples.shape[1]
if self.w_guess == None:
self.w = N.ones(NF, 'd')
# do normalization in all cases to be safe :)
self.w = self.w/(self.w**2).sum()
M, H = self.compute_M_H(dataset.labels)
ni = N.zeros(NF, 'd')
pi = N.zeros(NF, 'd')
if self.permute:
# indices to go through samples in random order
random_sequence = N.random.permutation(NS)
else:
random_sequence = N.arange(NS)
change = self.threshold + 1.0
iteration = 0
counter = 0.0
while change > self.threshold and iteration < self.max_iter:
if __debug__:
debug('IRELIEF', "Iteration %d" % iteration)
for t in range(NS):
counter += 1.0
n = random_sequence[t]
self.k = self.kernel(length_scale = self.kernel_width/self.w)
d_w_k_xn_Mn = self.k.compute(dataset.samples[None, n, :],
dataset.samples[M[n], :]).squeeze()
d_w_k_xn_Mn_sum = d_w_k_xn_Mn.sum()
d_w_k_xn_x = self.k.compute(dataset.samples[None, n, :],
dataset.samples).squeeze()
gamma_n = 1.0 - d_w_k_xn_Mn_sum / d_w_k_xn_x.sum()
alpha_n = d_w_k_xn_Mn / d_w_k_xn_Mn_sum
d_w_k_xn_Hn = self.k.compute(dataset.samples[None, n, :],
dataset.samples[H[n], :]).squeeze()
beta_n = d_w_k_xn_Hn / d_w_k_xn_Hn.sum()
m_n = (N.abs(dataset.samples[n, :] - dataset.samples[M[n], :]) \
* alpha_n[:, N.newaxis]).sum(0)
h_n = (N.abs(dataset.samples[n, :] - dataset.samples[H[n], :]) \
* beta_n[:, N.newaxis]).sum(0)
pi = gamma_n * (m_n-h_n)
learning_rate = 1.0 / (counter * self.a + 1.0)
ni_new = ni + learning_rate * (pi - ni)
ni = ni_new
# set all negative elements to zero
ni_plus = N.clip(ni, 0.0, N.inf)
w_new = N.nan_to_num(ni_plus / (N.sqrt((ni_plus ** 2).sum())))
change = N.abs(w_new - self.w).sum()
if t % 10 == 0 and __debug__ and 'IRELIEF' in debug.active:
debug('IRELIEF',
"t=%d change=%.4f max=%f min=%.4f mean=%.4f std=%.4f"
" #nan=%d" %
(t, change, w_new.max(), w_new.min(), w_new.mean(),
w_new.std(), N.isnan(w_new).sum()))
self.w = w_new
if change < self.threshold and iteration > 0:
break
iteration += 1
return self.w
class IterativeRelief(FeaturewiseDatasetMeasure):
"""`FeaturewiseDatasetMeasure` that performs multivariate I-RELIEF
algorithm. Batch version.
Batch I-RELIEF-2 feature weighting algorithm. Works for binary or
multiclass class-labels. Batch version with complexity O(T*N^2*I),
where T is the number of iterations, N the number of instances, I
the number of features.
See: Y. Sun, Iterative RELIEF for Feature Weighting: Algorithms,
Theories, and Applications, IEEE Trans. on Pattern Analysis and
Machine Intelligence (TPAMI), vol. 29, no. 6, pp. 1035-1051, June
2007. http://plaza.ufl.edu/sunyijun/Paper/PAMI_1.pdf
Note that current implementation allows to use only
exponential-like kernels. Support for linear kernel will be
added later.
"""
def __init__(self, threshold=1.0e-2, kernel_width=1.0,
w_guess=None, **kwargs):
"""Constructor of the IRELIEF class.
"""
# init base classes first
FeaturewiseDatasetMeasure.__init__(self, **kwargs)
# Threshold in W changes (stopping criterion for irelief).
self.threshold = threshold
self.w_guess = w_guess
self.w = None
self.kernel_width = kernel_width
def compute_M_H(self, label):
"""Compute hit/miss dictionaries.
For each instance compute the set of indices having the same
class label and different class label.
Note that this computation is independent of the number of
features.
XXX should it be some generic function since it doesn't use self
"""
M = {}
H = {}
for i in range(label.size):
M[i] = N.where(label != label[i])[0]
tmp = (N.where(label == label[i])[0]).tolist()
tmp.remove(i)
# There must be least two exampls for class label[i]
assert(tmp != [])
H[i] = N.array(tmp)
return M, H
def k(self, distances):
"""Exponential kernel."""
kd = N.exp(-distances/self.kernel_width)
# set kd to zero where distance=0 otherwise I-RELIEF could not converge.
kd[N.abs(distances) < 1.0e-15] = 0.0
return kd
def _call(self, dataset):
"""Computes featurewise I-RELIEF weights."""
samples = dataset.samples
NS, NF = samples.shape[:2]
if self.w_guess == None:
w = N.ones(NF, 'd')
w /= (w ** 2).sum() # do normalization in all cases to be safe :)
M, H = self.compute_M_H(dataset.labels)
while True:
d_w_k = self.k(pnorm_w(data1=samples, weight=w, p=1))
ni = N.zeros(NF, 'd')
for n in range(NS):
# d_w_k[n, n] could be omitted since == 0.0
gamma_n = 1.0 - N.nan_to_num(d_w_k[n, M[n]].sum() \
/ (d_w_k[n, :].sum() - d_w_k[n, n]))
alpha_n = N.nan_to_num(d_w_k[n, M[n]] / (d_w_k[n, M[n]].sum()))
beta_n = N.nan_to_num(d_w_k[n, H[n]] / (d_w_k[n, H[n]].sum()))
m_n = (N.abs(samples[n, :] - samples[M[n], :]) \
* alpha_n[:, None]).sum(0)
h_n = (N.abs(samples[n, :] - samples[H[n], :]) \
* beta_n[:, None]).sum(0)
ni += gamma_n*(m_n - h_n)
ni = ni / NS
ni_plus = N.clip(ni, 0.0, N.inf) # set all negative elements to zero
w_new = N.nan_to_num(ni_plus / (N.sqrt((ni_plus**2).sum())))
change = N.abs(w_new - w).sum()
if __debug__ and 'IRELIEF' in debug.active:
debug('IRELIEF',
"change=%.4f max=%f min=%.4f mean=%.4f std=%.4f #nan=%d" \
% (change, w_new.max(), w_new.min(), w_new.mean(),
w_new.std(), N.isnan(w_new).sum()))
# update weights:
w = w_new
if change < self.threshold:
break
self.w = w
return w
class IterativeReliefOnline(IterativeRelief):
"""`FeaturewiseDatasetMeasure` that performs multivariate I-RELIEF
algorithm. Online version.
This algorithm is exactly the one in the referenced paper
(I-RELIEF-2 online), using weighted 1-norm and Exponential
Kernel.
"""
def __init__(self, a=10.0, permute=True, max_iter=3, **kwargs):
"""Constructor of the IRELIEF class.
"""
# init base classes first
IterativeRelief.__init__(self, **kwargs)
self.a = a # parameter of the learning rate
self.permute = permute # shuffle data when running I-RELIEF
self.max_iter = max_iter # maximum number of iterations
def _call(self, dataset):
"""Computes featurewise I-RELIEF-2 weights. Online version."""
# local bindings
samples = dataset.samples
NS, NF = samples.shape[:2]
threshold = self.threshold
a = self.a
if self.w_guess == None:
w = N.ones(NF, 'd')
# do normalization in all cases to be safe :)
w /= (w ** 2).sum()
M, H = self.compute_M_H(dataset.labels)
ni = N.zeros(NF, 'd')
pi = N.zeros(NF, 'd')
if self.permute:
# indices to go through x in random order
random_sequence = N.random.permutation(NS)
else:
random_sequence = N.arange(NS)
change = threshold + 1.0
iteration = 0
counter = 0.0
while change > threshold and iteration < self.max_iter:
if __debug__:
debug('IRELIEF', "Iteration %d" % iteration)
for t in range(NS):
counter += 1.0
n = random_sequence[t]
d_xn_x = N.abs(samples[n, :] - samples)
d_w_k_xn_x = self.k((d_xn_x * w).sum(1))
d_w_k_xn_Mn = d_w_k_xn_x[M[n]]
d_w_k_xn_Mn_sum = d_w_k_xn_Mn.sum()
gamma_n = 1.0 - d_w_k_xn_Mn_sum / d_w_k_xn_x.sum()
alpha_n = d_w_k_xn_Mn / d_w_k_xn_Mn_sum
d_w_k_xn_Hn = d_w_k_xn_x[H[n]]
beta_n = d_w_k_xn_Hn / d_w_k_xn_Hn.sum()
m_n = (d_xn_x[M[n], :] * alpha_n[:, None]).sum(0)
h_n = (d_xn_x[H[n], :] * beta_n[:, None]).sum(0)
pi = gamma_n * (m_n - h_n)
learning_rate = 1.0 / (counter * a + 1.0)
ni_new = ni + learning_rate * (pi - ni)
ni = ni_new
# set all negative elements to zero
ni_plus = N.clip(ni, 0.0, N.inf)
w_new = N.nan_to_num(ni_plus / (N.sqrt((ni_plus ** 2).sum())))
change = N.abs(w_new - w).sum()
if t % 10 == 0 and __debug__ and 'IRELIEF' in debug.active:
debug('IRELIEF',
"t=%d change=%.4f max=%f min=%.4f mean=%.4f std=%.4f"
" #nan=%d" %
(t, change, w_new.max(), w_new.min(), w_new.mean(),
w_new.std(), N.isnan(w_new).sum()))
w = w_new
if change < threshold and iteration > 0:
break
iteration += 1
self.w = w
return w
|