def run_many_merge_moves(
hmodel, Data, SS, evBound=None, nMergeTrials=1, compList=list(), randstate=np.random, mPairIDs=None, **mergeKwArgs
):
""" Run (potentially many) merge move on hmodel
Args
-------
hmodel
Data
SS
nMergeTrials : number of merges to try
compList : list of components to include in attempted merges
randstate : numpy random number generator
Returns
-------
hmodel
SS
evBound
MTracker
"""
nMergeTrials = np.maximum(nMergeTrials, len(compList))
MTracker = MergeTracker(SS.K)
MSelector = MergePairSelector()
# Exclude all pairs for which we did not compute the combined entropy Hz
# Hz is always stored in KxK matrix. Pairs that were skipped have zeros.
aList = list()
bList = list()
if SS.hasMergeTerm("ElogqZ"):
Hz = SS.getMergeTerm("ElogqZ")
for kA in xrange(SS.K):
for kB in xrange(kA + 1, SS.K):
if Hz[kA, kB] == 0:
aList.append(kA)
bList.append(kB)
if len(aList) > 0:
MTracker.addPairsToExclude(aList, bList)
if evBound is None:
newEv = hmodel.calc_evidence(SS=SS)
else:
newEv = evBound
trialID = 0
shift = np.zeros(SS.K, dtype=np.int32)
while trialID < nMergeTrials and MTracker.hasAvailablePairs():
oldEv = newEv
if mPairIDs is not None:
if len(mPairIDs) == 0:
break
kA, kB = mPairIDs.pop(0)
try:
MTracker.verifyPair(kA, kB)
except AssertionError:
print " AssertionError skipped with mPairIDs!", kA, kB
continue
elif len(compList) > 0:
kA = compList.pop()
if kA not in MTracker.getAvailableComps():
continue
kB = None
else:
kA = None
kB = None
hmodel, SS, newEv, MoveInfo = run_merge_move(
hmodel,
Data,
SS,
oldEv,
kA=kA,
kB=kB,
randstate=randstate,
MSelector=MSelector,
MTracker=MTracker,
**mergeKwArgs
)
if MoveInfo["didAccept"]:
assert newEv >= oldEv
if mPairIDs is not None:
mPairIDs = _reindexCandidatePairsAfterAcceptedMerge(mPairIDs, kA, kB)
trialID += 1
MTracker.recordResult(**MoveInfo)
return hmodel, SS, newEv, MTracker
def run_many_merge_moves(hmodel,
Data,
SS,
evBound=None,
nMergeTrials=1,
compList=list(),
randstate=np.random,
mPairIDs=None,
**mergeKwArgs):
''' Run (potentially many) merge move on hmodel
Args
-------
hmodel
Data
SS
nMergeTrials : number of merges to try
compList : list of components to include in attempted merges
randstate : numpy random number generator
Returns
-------
hmodel
SS
evBound
MTracker
'''
nMergeTrials = np.maximum(nMergeTrials, len(compList))
MTracker = MergeTracker(SS.K)
MSelector = MergePairSelector()
# Exclude all pairs for which we did not compute the combined entropy Hz
# Hz is always stored in KxK matrix. Pairs that were skipped have zeros.
aList = list()
bList = list()
if SS.hasMergeTerm('ElogqZ'):
Hz = SS.getMergeTerm('ElogqZ')
for kA in xrange(SS.K):
for kB in xrange(kA + 1, SS.K):
if Hz[kA, kB] == 0:
aList.append(kA)
bList.append(kB)
if len(aList) > 0:
MTracker.addPairsToExclude(aList, bList)
if evBound is None:
newEv = hmodel.calc_evidence(SS=SS)
else:
newEv = evBound
trialID = 0
shift = np.zeros(SS.K, dtype=np.int32)
while trialID < nMergeTrials and MTracker.hasAvailablePairs():
oldEv = newEv
if mPairIDs is not None:
if len(mPairIDs) == 0:
break
kA, kB = mPairIDs.pop(0)
try:
MTracker.verifyPair(kA, kB)
except AssertionError:
print ' AssertionError skipped with mPairIDs!', kA, kB
continue
elif len(compList) > 0:
kA = compList.pop()
if kA not in MTracker.getAvailableComps():
continue
kB = None
else:
kA = None
kB = None
hmodel, SS, newEv, MoveInfo = run_merge_move(hmodel,
Data,
SS,
oldEv,
kA=kA,
kB=kB,
randstate=randstate,
MSelector=MSelector,
MTracker=MTracker,
**mergeKwArgs)
if MoveInfo['didAccept']:
assert newEv >= oldEv
if mPairIDs is not None:
mPairIDs = _reindexCandidatePairsAfterAcceptedMerge(
mPairIDs, kA, kB)
trialID += 1
MTracker.recordResult(**MoveInfo)
return hmodel, SS, newEv, MTracker
def run_merge_move(
curModel,
Data,
SS=None,
curEv=None,
doVizMerge=False,
kA=None,
kB=None,
MTracker=None,
MSelector=None,
mergename="marglik",
randstate=np.random.RandomState(),
doUpdateAllComps=0,
savedir=None,
doVerbose=False,
doWriteLog=False,
**kwargs
):
""" Creates candidate model with two components merged,
and returns either candidate or current model,
whichever has higher log probability (ELBO).
Args
--------
curModel : bnpy model whose components will be merged
Data : bnpy Data object
SS : bnpy SuffStatDict object for Data under curModel
must contain precomputed merge entropy in order to try a merge.
curEv : current evidence bound, provided to save re-computation.
curEv = curModel.calc_evidence(SS=SS)
kA, kB : (optional) integer ids for which specific components to merge
excludeList : (optional) list of integer ids excluded when selecting
which components to merge. useful when doing multiple
rounds of merges, since precomputed merge terms are
valid for one merge only.
Returns
--------
hmodel, SS, evBound, MoveInfo
hmodel := candidate or current model (bnpy HModel object)
SS := suff stats for Data under hmodel
evBound := log evidence (ELBO) of Data under hmodel
MoveInfo := dict of info about this merge move, with fields
didAccept := boolean flag, true if candidate accepted
msg := human-readable string about this move
kA, kB := indices of the components to be merged.
"""
if SS is None:
LP = curModel.calc_local_params(Data)
SS = curModel.get_global_suff_stats(Data, LP, doPrecompEntropy=True, doPrecompMerge=True)
if curEv is None:
curEv = curModel.calc_evidence(SS=SS)
if MTracker is None:
MTracker = MergeTracker(SS.K)
if MSelector is None:
MSelector = MergePairSelector()
# Need at least two components to merge!
if curModel.allocModel.K == 1:
MoveInfo = dict(didAccept=0, msg="need >= 2 comps to merge")
return curModel, SS, curEv, MoveInfo
if not SS.hasMergeTerms() and curModel.allocModel.requireMergeTerms():
MoveInfo = dict(didAccept=0, msg="suff stats did not have merge terms")
return curModel, SS, curEv, MoveInfo
if kA is not None and kA not in MTracker.getAvailableComps():
MoveInfo = dict(didAccept=0, msg="target comp kA must be excluded.")
return curModel, SS, curEv, MoveInfo
# Select which 2 components kA, kB in {1, 2, ... K} to merge
if kA is None or kB is None:
kA, kB = select_merge_components(
curModel, Data, SS, kA=kA, MTracker=MTracker, MSelector=MSelector, mergename=mergename, randstate=randstate
)
# Create candidate merged model
propModel, propSS = propose_merge_candidate(curModel, SS, kA, kB, doUpdateAllComps=doUpdateAllComps)
# Decide whether to accept the merge
propEv = propModel.calc_evidence(SS=propSS)
if np.isnan(propEv) or np.isinf(propEv):
raise ValueError("propEv should never be nan/inf")
if doVizMerge:
viz_merge_proposal(curModel, propModel, kA, kB, curEv, propEv)
evDiff = propEv - curEv
if hasattr(SS, "nDoc") and np.abs(propEv - curEv) > 0.05 * np.abs(curEv):
print "CRAP! ---------------------------------------!!!!$$$$$$$$"
print " propEv % .5e" % (propEv)
print " curEv % .5e" % (curEv)
MoveInfo = dict(didAccept=0, kA=kA, kB=kB, msg="CRAP. bad proposed evidence.")
return curModel, SS, curEv, MoveInfo
if hasattr(SS, "nDoc") and (propEv > 0 and curEv < 0):
print "CRAP! ---------------------------------------!!!!@@@@@@@@"
print " propEv % .5e" % (propEv)
print " curEv % .5e" % (curEv)
MoveInfo = dict(didAccept=0, kA=kA, kB=kB, msg="CRAP. bad proposed evidence.")
return curModel, SS, curEv, MoveInfo
if propEv >= curEv:
MSelector.reindexAfterMerge(kA, kB)
msg = "merge %3d & %3d | ev +%.3e ****" % (kA, kB, propEv - curEv)
MoveInfo = dict(didAccept=1, kA=kA, kB=kB, msg=msg, evDiff=evDiff)
if doWriteLog:
log_merge_move(MoveInfo, MSelector, curModel, SS, savedir)
return propModel, propSS, propEv, MoveInfo
else:
msg = "merge %3d & %3d | ev -%.3e" % (kA, kB, curEv - propEv)
MoveInfo = dict(didAccept=0, kA=kA, kB=kB, msg=msg, evDiff=evDiff)
if doWriteLog:
log_merge_move(MoveInfo, MSelector, curModel, SS, savedir)
return curModel, SS, curEv, MoveInfo
def run_merge_move(curModel,
Data,
SS=None,
curEv=None,
doVizMerge=False,
kA=None,
kB=None,
MTracker=None,
MSelector=None,
mergename='marglik',
randstate=np.random.RandomState(),
doUpdateAllComps=0,
savedir=None,
doVerbose=False,
doWriteLog=False,
**kwargs):
''' Creates candidate model with two components merged,
and returns either candidate or current model,
whichever has higher log probability (ELBO).
Args
--------
curModel : bnpy model whose components will be merged
Data : bnpy Data object
SS : bnpy SuffStatDict object for Data under curModel
must contain precomputed merge entropy in order to try a merge.
curEv : current evidence bound, provided to save re-computation.
curEv = curModel.calc_evidence(SS=SS)
kA, kB : (optional) integer ids for which specific components to merge
excludeList : (optional) list of integer ids excluded when selecting
which components to merge. useful when doing multiple
rounds of merges, since precomputed merge terms are
valid for one merge only.
Returns
--------
hmodel, SS, evBound, MoveInfo
hmodel := candidate or current model (bnpy HModel object)
SS := suff stats for Data under hmodel
evBound := log evidence (ELBO) of Data under hmodel
MoveInfo := dict of info about this merge move, with fields
didAccept := boolean flag, true if candidate accepted
msg := human-readable string about this move
kA, kB := indices of the components to be merged.
'''
if SS is None:
LP = curModel.calc_local_params(Data)
SS = curModel.get_global_suff_stats(Data,
LP,
doPrecompEntropy=True,
doPrecompMerge=True)
if curEv is None:
curEv = curModel.calc_evidence(SS=SS)
if MTracker is None:
MTracker = MergeTracker(SS.K)
if MSelector is None:
MSelector = MergePairSelector()
# Need at least two components to merge!
if curModel.allocModel.K == 1:
MoveInfo = dict(didAccept=0, msg="need >= 2 comps to merge")
return curModel, SS, curEv, MoveInfo
if not SS.hasMergeTerms() and curModel.allocModel.requireMergeTerms():
MoveInfo = dict(didAccept=0, msg="suff stats did not have merge terms")
return curModel, SS, curEv, MoveInfo
if kA is not None and kA not in MTracker.getAvailableComps():
MoveInfo = dict(didAccept=0, msg="target comp kA must be excluded.")
return curModel, SS, curEv, MoveInfo
# Select which 2 components kA, kB in {1, 2, ... K} to merge
if kA is None or kB is None:
kA, kB = select_merge_components(curModel,
Data,
SS,
kA=kA,
MTracker=MTracker,
MSelector=MSelector,
mergename=mergename,
randstate=randstate)
# Create candidate merged model
propModel, propSS = propose_merge_candidate(
curModel, SS, kA, kB, doUpdateAllComps=doUpdateAllComps)
# Decide whether to accept the merge
propEv = propModel.calc_evidence(SS=propSS)
if np.isnan(propEv) or np.isinf(propEv):
raise ValueError('propEv should never be nan/inf')
if doVizMerge:
viz_merge_proposal(curModel, propModel, kA, kB, curEv, propEv)
evDiff = propEv - curEv
if hasattr(SS, 'nDoc') and np.abs(propEv - curEv) > 0.05 * np.abs(curEv):
print 'CRAP! ---------------------------------------!!!!$$$$$$$$'
print ' propEv % .5e' % (propEv)
print ' curEv % .5e' % (curEv)
MoveInfo = dict(didAccept=0,
kA=kA,
kB=kB,
msg="CRAP. bad proposed evidence.")
return curModel, SS, curEv, MoveInfo
if hasattr(SS, 'nDoc') and (propEv > 0 and curEv < 0):
print 'CRAP! ---------------------------------------!!!!@@@@@@@@'
print ' propEv % .5e' % (propEv)
print ' curEv % .5e' % (curEv)
MoveInfo = dict(didAccept=0,
kA=kA,
kB=kB,
msg="CRAP. bad proposed evidence.")
return curModel, SS, curEv, MoveInfo
if propEv >= curEv:
MSelector.reindexAfterMerge(kA, kB)
msg = "merge %3d & %3d | ev +%.3e ****" % (kA, kB, propEv - curEv)
MoveInfo = dict(didAccept=1, kA=kA, kB=kB, msg=msg, evDiff=evDiff)
if doWriteLog:
log_merge_move(MoveInfo, MSelector, curModel, SS, savedir)
return propModel, propSS, propEv, MoveInfo
else:
msg = "merge %3d & %3d | ev -%.3e" % (kA, kB, curEv - propEv)
MoveInfo = dict(didAccept=0, kA=kA, kB=kB, msg=msg, evDiff=evDiff)
if doWriteLog:
log_merge_move(MoveInfo, MSelector, curModel, SS, savedir)
return curModel, SS, curEv, MoveInfo
