1 # -*- coding: utf-8 -*-
3 # Copyright (C) 2008-2019 EDF R&D
5 # This library is free software; you can redistribute it and/or
6 # modify it under the terms of the GNU Lesser General Public
7 # License as published by the Free Software Foundation; either
8 # version 2.1 of the License.
10 # This library is distributed in the hope that it will be useful,
11 # but WITHOUT ANY WARRANTY; without even the implied warranty of
12 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 # Lesser General Public License for more details.
15 # You should have received a copy of the GNU Lesser General Public
16 # License along with this library; if not, write to the Free Software
17 # Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 # See http://www.salome-platform.org/ or email : webmaster.salome@opencascade.com
21 # Author: Jean-Philippe Argaud, jean-philippe.argaud@edf.fr, EDF R&D
24 Définit les versions approximées des opérateurs tangents et adjoints.
26 __author__ = "Jean-Philippe ARGAUD"
28 import os, numpy, time, copy, types, sys, math, logging
29 from daCore.BasicObjects import Operator
30 from daCore.PlatformInfo import PlatformInfo
31 mpr = PlatformInfo().MachinePrecision()
32 mfp = PlatformInfo().MaximumPrecision()
33 # logging.getLogger().setLevel(logging.DEBUG)
35 # ==============================================================================
36 def ExecuteFunction( paire ):
37 assert len(paire) == 2, "Incorrect number of arguments"
39 __X = numpy.asmatrix(numpy.ravel( X )).T
40 __sys_path_tmp = sys.path ; sys.path.insert(0,funcrepr["__userFunction__path"])
41 __module = __import__(funcrepr["__userFunction__modl"], globals(), locals(), [])
42 __fonction = getattr(__module,funcrepr["__userFunction__name"])
43 sys.path = __sys_path_tmp ; del __sys_path_tmp
44 __HX = __fonction( __X )
45 return numpy.ravel( __HX )
47 # ==============================================================================
48 class FDApproximation(object):
50 Cette classe sert d'interface pour définir les opérateurs approximés. A la
51 création d'un objet, en fournissant une fonction "Function", on obtient un
52 objet qui dispose de 3 méthodes "DirectOperator", "TangentOperator" et
53 "AdjointOperator". On contrôle l'approximation DF avec l'incrément
54 multiplicatif "increment" valant par défaut 1%, ou avec l'incrément fixe
55 "dX" qui sera multiplié par "increment" (donc en %), et on effectue de DF
56 centrées si le booléen "centeredDF" est vrai.
63 avoidingRedundancy = True,
64 toleranceInRedundancy = 1.e-18,
65 lenghtOfRedundancy = -1,
72 import multiprocessing
73 self.__mpEnabled = True
75 self.__mpEnabled = False
77 self.__mpEnabled = False
78 self.__mpWorkers = mpWorkers
79 if self.__mpWorkers is not None and self.__mpWorkers < 1:
80 self.__mpWorkers = None
81 logging.debug("FDA Calculs en multiprocessing : %s (nombre de processus : %s)"%(self.__mpEnabled,self.__mpWorkers))
84 self.__mfEnabled = True
86 self.__mfEnabled = False
87 logging.debug("FDA Calculs en multifonctions : %s"%(self.__mfEnabled,))
89 if avoidingRedundancy:
91 self.__tolerBP = float(toleranceInRedundancy)
92 self.__lenghtRJ = int(lenghtOfRedundancy)
93 self.__listJPCP = [] # Jacobian Previous Calculated Points
94 self.__listJPCI = [] # Jacobian Previous Calculated Increment
95 self.__listJPCR = [] # Jacobian Previous Calculated Results
96 self.__listJPPN = [] # Jacobian Previous Calculated Point Norms
97 self.__listJPIN = [] # Jacobian Previous Calculated Increment Norms
99 self.__avoidRC = False
102 if isinstance(Function,types.FunctionType):
103 logging.debug("FDA Calculs en multiprocessing : FunctionType")
104 self.__userFunction__name = Function.__name__
106 mod = os.path.join(Function.__globals__['filepath'],Function.__globals__['filename'])
108 mod = os.path.abspath(Function.__globals__['__file__'])
109 if not os.path.isfile(mod):
110 raise ImportError("No user defined function or method found with the name %s"%(mod,))
111 self.__userFunction__modl = os.path.basename(mod).replace('.pyc','').replace('.pyo','').replace('.py','')
112 self.__userFunction__path = os.path.dirname(mod)
114 self.__userOperator = Operator( fromMethod = Function, avoidingRedundancy = self.__avoidRC, inputAsMultiFunction = self.__mfEnabled )
115 self.__userFunction = self.__userOperator.appliedTo # Pour le calcul Direct
116 elif isinstance(Function,types.MethodType):
117 logging.debug("FDA Calculs en multiprocessing : MethodType")
118 self.__userFunction__name = Function.__name__
120 mod = os.path.join(Function.__globals__['filepath'],Function.__globals__['filename'])
122 mod = os.path.abspath(Function.__func__.__globals__['__file__'])
123 if not os.path.isfile(mod):
124 raise ImportError("No user defined function or method found with the name %s"%(mod,))
125 self.__userFunction__modl = os.path.basename(mod).replace('.pyc','').replace('.pyo','').replace('.py','')
126 self.__userFunction__path = os.path.dirname(mod)
128 self.__userOperator = Operator( fromMethod = Function, avoidingRedundancy = self.__avoidRC, inputAsMultiFunction = self.__mfEnabled )
129 self.__userFunction = self.__userOperator.appliedTo # Pour le calcul Direct
131 raise TypeError("User defined function or method has to be provided for finite differences approximation.")
133 self.__userOperator = Operator( fromMethod = Function, avoidingRedundancy = self.__avoidRC, inputAsMultiFunction = self.__mfEnabled )
134 self.__userFunction = self.__userOperator.appliedTo
136 self.__centeredDF = bool(centeredDF)
137 if abs(float(increment)) > 1.e-15:
138 self.__increment = float(increment)
140 self.__increment = 0.01
144 self.__dX = numpy.asmatrix(numpy.ravel( dX )).T
145 logging.debug("FDA Reduction des doublons de calcul : %s"%self.__avoidRC)
147 logging.debug("FDA Tolerance de determination des doublons : %.2e"%self.__tolerBP)
149 # ---------------------------------------------------------
150 def __doublon__(self, e, l, n, v=None):
151 __ac, __iac = False, -1
152 for i in range(len(l)-1,-1,-1):
153 if numpy.linalg.norm(e - l[i]) < self.__tolerBP * n[i]:
154 __ac, __iac = True, i
155 if v is not None: logging.debug("FDA Cas%s déja calculé, récupération du doublon %i"%(v,__iac))
159 # ---------------------------------------------------------
160 def DirectOperator(self, X ):
162 Calcul du direct à l'aide de la fonction fournie.
164 logging.debug("FDA Calcul DirectOperator (explicite)")
166 _HX = self.__userFunction( X, argsAsSerie = True )
168 _X = numpy.asmatrix(numpy.ravel( X )).T
169 _HX = numpy.ravel(self.__userFunction( _X ))
173 # ---------------------------------------------------------
174 def TangentMatrix(self, X ):
176 Calcul de l'opérateur tangent comme la Jacobienne par différences finies,
177 c'est-à-dire le gradient de H en X. On utilise des différences finies
178 directionnelles autour du point X. X est un numpy.matrix.
180 Différences finies centrées (approximation d'ordre 2):
181 1/ Pour chaque composante i de X, on ajoute et on enlève la perturbation
182 dX[i] à la composante X[i], pour composer X_plus_dXi et X_moins_dXi, et
183 on calcule les réponses HX_plus_dXi = H( X_plus_dXi ) et HX_moins_dXi =
185 2/ On effectue les différences (HX_plus_dXi-HX_moins_dXi) et on divise par
187 3/ Chaque résultat, par composante, devient une colonne de la Jacobienne
189 Différences finies non centrées (approximation d'ordre 1):
190 1/ Pour chaque composante i de X, on ajoute la perturbation dX[i] à la
191 composante X[i] pour composer X_plus_dXi, et on calcule la réponse
192 HX_plus_dXi = H( X_plus_dXi )
193 2/ On calcule la valeur centrale HX = H(X)
194 3/ On effectue les différences (HX_plus_dXi-HX) et on divise par
196 4/ Chaque résultat, par composante, devient une colonne de la Jacobienne
199 logging.debug("FDA Début du calcul de la Jacobienne")
200 logging.debug("FDA Incrément de............: %s*X"%float(self.__increment))
201 logging.debug("FDA Approximation centrée...: %s"%(self.__centeredDF))
203 if X is None or len(X)==0:
204 raise ValueError("Nominal point X for approximate derivatives can not be None or void (X=%s)."%(str(X),))
206 _X = numpy.asmatrix(numpy.ravel( X )).T
208 if self.__dX is None:
209 _dX = self.__increment * _X
211 _dX = numpy.asmatrix(numpy.ravel( self.__dX )).T
213 if (_dX == 0.).any():
216 _dX = numpy.where( _dX == 0., float(self.__increment), _dX )
218 _dX = numpy.where( _dX == 0., moyenne, _dX )
220 __alreadyCalculated = False
222 __bidon, __alreadyCalculatedP = self.__doublon__(_X, self.__listJPCP, self.__listJPPN, None)
223 __bidon, __alreadyCalculatedI = self.__doublon__(_dX, self.__listJPCI, self.__listJPIN, None)
224 if __alreadyCalculatedP == __alreadyCalculatedI > -1:
225 __alreadyCalculated, __i = True, __alreadyCalculatedP
226 logging.debug("FDA Cas J déja calculé, récupération du doublon %i"%__i)
228 if __alreadyCalculated:
229 logging.debug("FDA Calcul Jacobienne (par récupération du doublon %i)"%__i)
230 _Jacobienne = self.__listJPCR[__i]
232 logging.debug("FDA Calcul Jacobienne (explicite)")
233 if self.__centeredDF:
235 if self.__mpEnabled and not self.__mfEnabled:
237 "__userFunction__path" : self.__userFunction__path,
238 "__userFunction__modl" : self.__userFunction__modl,
239 "__userFunction__name" : self.__userFunction__name,
242 for i in range( len(_dX) ):
244 _X_plus_dXi = numpy.array( _X.A1, dtype=float )
245 _X_plus_dXi[i] = _X[i] + _dXi
246 _X_moins_dXi = numpy.array( _X.A1, dtype=float )
247 _X_moins_dXi[i] = _X[i] - _dXi
249 _jobs.append( (_X_plus_dXi, funcrepr) )
250 _jobs.append( (_X_moins_dXi, funcrepr) )
252 import multiprocessing
253 self.__pool = multiprocessing.Pool(self.__mpWorkers)
254 _HX_plusmoins_dX = self.__pool.map( ExecuteFunction, _jobs )
259 for i in range( len(_dX) ):
260 _Jacobienne.append( numpy.ravel( _HX_plusmoins_dX[2*i] - _HX_plusmoins_dX[2*i+1] ) / (2.*_dX[i]) )
262 elif self.__mfEnabled:
264 for i in range( len(_dX) ):
266 _X_plus_dXi = numpy.array( _X.A1, dtype=float )
267 _X_plus_dXi[i] = _X[i] + _dXi
268 _X_moins_dXi = numpy.array( _X.A1, dtype=float )
269 _X_moins_dXi[i] = _X[i] - _dXi
271 _xserie.append( _X_plus_dXi )
272 _xserie.append( _X_moins_dXi )
274 _HX_plusmoins_dX = self.DirectOperator( _xserie )
277 for i in range( len(_dX) ):
278 _Jacobienne.append( numpy.ravel( _HX_plusmoins_dX[2*i] - _HX_plusmoins_dX[2*i+1] ) / (2.*_dX[i]) )
282 for i in range( _dX.size ):
284 _X_plus_dXi = numpy.array( _X.A1, dtype=float )
285 _X_plus_dXi[i] = _X[i] + _dXi
286 _X_moins_dXi = numpy.array( _X.A1, dtype=float )
287 _X_moins_dXi[i] = _X[i] - _dXi
289 _HX_plus_dXi = self.DirectOperator( _X_plus_dXi )
290 _HX_moins_dXi = self.DirectOperator( _X_moins_dXi )
292 _Jacobienne.append( numpy.ravel( _HX_plus_dXi - _HX_moins_dXi ) / (2.*_dXi) )
296 if self.__mpEnabled and not self.__mfEnabled:
298 "__userFunction__path" : self.__userFunction__path,
299 "__userFunction__modl" : self.__userFunction__modl,
300 "__userFunction__name" : self.__userFunction__name,
303 _jobs.append( (_X.A1, funcrepr) )
304 for i in range( len(_dX) ):
305 _X_plus_dXi = numpy.array( _X.A1, dtype=float )
306 _X_plus_dXi[i] = _X[i] + _dX[i]
308 _jobs.append( (_X_plus_dXi, funcrepr) )
310 import multiprocessing
311 self.__pool = multiprocessing.Pool(self.__mpWorkers)
312 _HX_plus_dX = self.__pool.map( ExecuteFunction, _jobs )
316 _HX = _HX_plus_dX.pop(0)
319 for i in range( len(_dX) ):
320 _Jacobienne.append( numpy.ravel(( _HX_plus_dX[i] - _HX ) / _dX[i]) )
322 elif self.__mfEnabled:
324 _xserie.append( _X.A1 )
325 for i in range( len(_dX) ):
326 _X_plus_dXi = numpy.array( _X.A1, dtype=float )
327 _X_plus_dXi[i] = _X[i] + _dX[i]
329 _xserie.append( _X_plus_dXi )
331 _HX_plus_dX = self.DirectOperator( _xserie )
333 _HX = _HX_plus_dX.pop(0)
336 for i in range( len(_dX) ):
337 _Jacobienne.append( numpy.ravel(( _HX_plus_dX[i] - _HX ) / _dX[i]) )
341 _HX = self.DirectOperator( _X )
342 for i in range( _dX.size ):
344 _X_plus_dXi = numpy.array( _X.A1, dtype=float )
345 _X_plus_dXi[i] = _X[i] + _dXi
347 _HX_plus_dXi = self.DirectOperator( _X_plus_dXi )
349 _Jacobienne.append( numpy.ravel(( _HX_plus_dXi - _HX ) / _dXi) )
352 _Jacobienne = numpy.asmatrix( numpy.vstack( _Jacobienne ) ).T
354 if self.__lenghtRJ < 0: self.__lenghtRJ = 2 * _X.size
355 while len(self.__listJPCP) > self.__lenghtRJ:
356 self.__listJPCP.pop(0)
357 self.__listJPCI.pop(0)
358 self.__listJPCR.pop(0)
359 self.__listJPPN.pop(0)
360 self.__listJPIN.pop(0)
361 self.__listJPCP.append( copy.copy(_X) )
362 self.__listJPCI.append( copy.copy(_dX) )
363 self.__listJPCR.append( copy.copy(_Jacobienne) )
364 self.__listJPPN.append( numpy.linalg.norm(_X) )
365 self.__listJPIN.append( numpy.linalg.norm(_Jacobienne) )
367 logging.debug("FDA Fin du calcul de la Jacobienne")
371 # ---------------------------------------------------------
372 def TangentOperator(self, paire ):
374 Calcul du tangent à l'aide de la Jacobienne.
377 assert len(paire) == 1, "Incorrect lenght of arguments"
379 assert len(_paire) == 2, "Incorrect number of arguments"
381 assert len(paire) == 2, "Incorrect number of arguments"
384 _Jacobienne = self.TangentMatrix( X )
385 if dX is None or len(dX) == 0:
387 # Calcul de la forme matricielle si le second argument est None
388 # -------------------------------------------------------------
389 if self.__mfEnabled: return [_Jacobienne,]
390 else: return _Jacobienne
393 # Calcul de la valeur linéarisée de H en X appliqué à dX
394 # ------------------------------------------------------
395 _dX = numpy.asmatrix(numpy.ravel( dX )).T
396 _HtX = numpy.dot(_Jacobienne, _dX)
397 if self.__mfEnabled: return [_HtX.A1,]
400 # ---------------------------------------------------------
401 def AdjointOperator(self, paire ):
403 Calcul de l'adjoint à l'aide de la Jacobienne.
406 assert len(paire) == 1, "Incorrect lenght of arguments"
408 assert len(_paire) == 2, "Incorrect number of arguments"
410 assert len(paire) == 2, "Incorrect number of arguments"
413 _JacobienneT = self.TangentMatrix( X ).T
414 if Y is None or len(Y) == 0:
416 # Calcul de la forme matricielle si le second argument est None
417 # -------------------------------------------------------------
418 if self.__mfEnabled: return [_JacobienneT,]
419 else: return _JacobienneT
422 # Calcul de la valeur de l'adjoint en X appliqué à Y
423 # --------------------------------------------------
424 _Y = numpy.asmatrix(numpy.ravel( Y )).T
425 _HaY = numpy.dot(_JacobienneT, _Y)
426 if self.__mfEnabled: return [_HaY.A1,]
429 # ==============================================================================
441 Implémentation informatique de l'algorithme MMQR, basée sur la publication :
442 David R. Hunter, Kenneth Lange, "Quantile Regression via an MM Algorithm",
443 Journal of Computational and Graphical Statistics, 9, 1, pp.60-77, 2000.
446 # Recuperation des donnees et informations initiales
447 # --------------------------------------------------
448 variables = numpy.ravel( x0 )
449 mesures = numpy.ravel( y )
450 increment = sys.float_info[0]
453 quantile = float(quantile)
455 # Calcul des parametres du MM
456 # ---------------------------
457 tn = float(toler) / n
458 e0 = -tn / math.log(tn)
459 epsilon = (e0-tn)/(1+math.log(e0))
461 # Calculs d'initialisation
462 # ------------------------
463 residus = mesures - numpy.ravel( func( variables ) )
464 poids = 1./(epsilon+numpy.abs(residus))
465 veps = 1. - 2. * quantile - residus * poids
466 lastsurrogate = -numpy.sum(residus*veps) - (1.-2.*quantile)*numpy.sum(residus)
469 # Recherche iterative
470 # -------------------
471 while (increment > toler) and (iteration < maxfun) :
474 Derivees = numpy.array(fprime(variables))
475 Derivees = Derivees.reshape(n,p) # Necessaire pour remettre en place la matrice si elle passe par des tuyaux YACS
476 DeriveesT = Derivees.transpose()
477 M = numpy.dot( DeriveesT , (numpy.array(numpy.matrix(p*[poids,]).T)*Derivees) )
478 SM = numpy.transpose(numpy.dot( DeriveesT , veps ))
479 step = - numpy.linalg.lstsq( M, SM, rcond=-1 )[0]
481 variables = variables + step
482 if bounds is not None:
483 while( (variables < numpy.ravel(numpy.asmatrix(bounds)[:,0])).any() or (variables > numpy.ravel(numpy.asmatrix(bounds)[:,1])).any() ):
485 variables = variables - step
486 residus = mesures - numpy.ravel( func(variables) )
487 surrogate = numpy.sum(residus**2 * poids) + (4.*quantile-2.) * numpy.sum(residus)
489 while ( (surrogate > lastsurrogate) and ( max(list(numpy.abs(step))) > 1.e-16 ) ) :
491 variables = variables - step
492 residus = mesures - numpy.ravel( func(variables) )
493 surrogate = numpy.sum(residus**2 * poids) + (4.*quantile-2.) * numpy.sum(residus)
495 increment = lastsurrogate-surrogate
496 poids = 1./(epsilon+numpy.abs(residus))
497 veps = 1. - 2. * quantile - residus * poids
498 lastsurrogate = -numpy.sum(residus * veps) - (1.-2.*quantile)*numpy.sum(residus)
502 Ecart = quantile * numpy.sum(residus) - numpy.sum( residus[residus<0] )
504 return variables, Ecart, [n,p,iteration,increment,0]
506 # ==============================================================================
507 if __name__ == "__main__":
508 print('\n AUTODIAGNOSTIC\n')