def get_batches(self,
batch_size,
num_batches=None,
shuffle=False,
cluster=False):
"""
:param batch_size:
:param num_batches:
:param shuffle:
:param cluster: cluster examples by their lengths; this might give performance boost (i.e. faster training).
:return:
"""
num_batches_per_epoch = int(math.ceil(self.num_examples / batch_size))
if num_batches is None:
num_batches = num_batches_per_epoch
num_epochs = int(math.ceil(num_batches / num_batches_per_epoch))
if shuffle:
random_idxs = random.sample(self.valid_idxs, len(self.valid_idxs))
if cluster:
sorted_idxs = sorted(random_idxs, key=self._sort_key)
sorted_grouped = lambda: list(grouper(sorted_idxs, batch_size))
grouped = lambda: random.sample(sorted_grouped(),
num_batches_per_epoch)
else:
random_grouped = lambda: list(grouper(random_idxs, batch_size))
grouped = random_grouped
else:
raw_grouped = lambda: list(grouper(self.valid_idxs, batch_size))
grouped = raw_grouped
batch_idx_tuples = itertools.chain.from_iterable(
grouped() for _ in range(num_epochs))
for _ in range(num_batches):
batch_idxs = tuple(i for i in next(batch_idx_tuples)
if i is not None)
batch_data = self.get_by_idxs(batch_idxs)
shared_batch_data = {}
for key, val in batch_data.items():
if key.startswith('*'):
assert self.shared is not None
shared_key = key[1:]
shared_batch_data[shared_key] = [
index(self.shared[shared_key], each) for each in val
]
batch_data.update(shared_batch_data)
batch_ds = DataSet(batch_data, self.data_type, shared=self.shared)
yield batch_idxs, batch_ds
def localGenerateInput(self,model,myInput):
"""
Function to select the next most informative point for refining the limit
surface search.
After this method is called, the self.inputInfo should be ready to be sent
to the model
@ In, model, model instance, an instance of a model
@ In, myInput, list, a list of the original needed inputs for the model (e.g. list of files, etc.)
@ Out, None
"""
if self.startAdaptive == True and self.adaptiveReady == True:
LimitSurfaceSearch.localGenerateInput(self,model,myInput)
#the adaptive sampler created the next point sampled vars
#find the closest branch
if self.hybridDETstrategy is not None: closestBranch, cdfValues, treer = self._checkClosestBranch()
else : closestBranch, cdfValues = self._checkClosestBranch()
if closestBranch is None: self.raiseADebug('An usable branch for next candidate has not been found => create a parallel branch!')
# add pbthresholds in the grid
investigatedPoint = {}
for key,value in cdfValues.items():
ind = utils.find_le_index(self.branchProbabilities[key],value)
if not ind: ind = 0
if value not in self.branchProbabilities[key]:
self.branchProbabilities[key].insert(ind,value)
self.branchValues[key].insert(ind,self.distDict[key].ppf(value))
investigatedPoint[key] = value
# collect investigated point
self.investigatedPoints.append(investigatedPoint)
if closestBranch:
info = self._retrieveBranchInfo(closestBranch)
self._constructEndInfoFromBranch(model, myInput, info, cdfValues)
else:
# create a new tree, since there are no branches that are close enough to the adaptive request
elm = ETS.HierarchicalNode(self.messageHandler,self.name + '_' + str(len(self.TreeInfo.keys())+1))
elm.add('name', self.name + '_'+ str(len(self.TreeInfo.keys())+1))
elm.add('startTime', 0.0)
# Initialize the endTime to be equal to the start one...
# It will modified at the end of each branch
elm.add('endTime', 0.0)
elm.add('runEnded',False)
elm.add('running',True)
elm.add('queue',False)
elm.add('completedHistory', False)
branchedLevel = {}
for key,value in cdfValues.items(): branchedLevel[key] = utils.index(self.branchProbabilities[key],value)
# The dictionary branchedLevel is stored in the xml tree too. That's because
# the advancement of the thresholds must follow the tree structure
elm.add('branchedLevel', branchedLevel)
if self.hybridDETstrategy is not None and not self.foundEpistemicTree:
# adaptive hybrid DET and not found a tree in the epistemic space
# take the first tree and modify the hybridsamplerCoordinate
hybridSampled = copy.deepcopy(self.TreeInfo.values()[0].getrootnode().get('hybridsamplerCoordinate'))
for hybridStrategy in hybridSampled:
for key in self.epistemicVariables.keys():
if key in hybridStrategy['SampledVars'].keys():
self.raiseADebug("epistemic var " + str(key)+" value = "+str(self.values[key]))
hybridStrategy['SampledVars'][key] = copy.copy(self.values[key])
hybridStrategy['SampledVarsPb'][key] = self.distDict[key].pdf(self.values[key])
hybridStrategy['prefix'] = len(self.TreeInfo.values())+1
# TODO: find a strategy to recompute the probability weight here (for now == PointProbability)
hybridStrategy['PointProbability'] = reduce(mul, self.inputInfo['SampledVarsPb'].values())
hybridStrategy['ProbabilityWeight'] = reduce(mul, self.inputInfo['SampledVarsPb'].values())
elm.add('hybridsamplerCoordinate', hybridSampled)
# Here it is stored all the info regarding the DET => we create the info for all the branchings and we store them
self.TreeInfo[self.name + '_' + str(len(self.TreeInfo.keys())+1)] = ETS.HierarchicalTree(self.messageHandler,elm)
self._createRunningQueueBeginOne(self.TreeInfo[self.name + '_' + str(len(self.TreeInfo.keys()))],branchedLevel, model,myInput)
return DynamicEventTree.localGenerateInput(self,model,myInput)
def localFinalizeActualSampling(self,jobObject,model,myInput,genRunQueue=True):
"""
General function (available to all samplers) that finalize the sampling
calculation just ended. In this case (DET), The function reads the
information from the ended calculation, updates the working variables, and
creates the new inputs for the next branches
@ In, jobObject, instance, an instance of a JobHandler
@ In, model, model instance, it is the instance of a RAVEN model
@ In, myInput, list, the generating input
@ In, genRunQueue, bool, optional, True if the RunQueue needs to be updated
@ Out, None
"""
self.workingDir = model.workingDir
# returnBranchInfo = self.__readBranchInfo(jobObject.output)
# Get the parent element tree (xml object) to retrieve the information needed to create the new inputs
parentNode = self._retrieveParentNode(jobObject.identifier)
# set runEnded and running to true and false respectively
parentNode.add('runEnded',True)
parentNode.add('running',False)
parentNode.add('endTime',self.actualEndTime)
# Read the branch info from the parent calculation (just ended calculation)
# This function stores the information in the dictionary 'self.actualBranchInfo'
# If no branch info, this history is concluded => return
if not self.__readBranchInfo(jobObject.output, jobObject.getWorkingDir()):
parentNode.add('completedHistory', True)
return False
# Collect the branch info in a multi-level dictionary
endInfo = {'endTime':self.actualEndTime,'endTimeStep':self.actualEndTs,'branchDist':list(self.actualBranchInfo.keys())[0]}
endInfo['branchChangedParams'] = self.actualBranchInfo[endInfo['branchDist']]
# check if RELAP7 mode is activated, in case prepend the "<distribution>" string
if any("<distribution>" in s for s in self.branchProbabilities.keys()):
endInfo['branchDist'] = self.toBeSampled.keys()[self.toBeSampled.values().index(endInfo['branchDist'])]
#endInfo['branchDist'] = "<distribution>"+endInfo['branchDist']
parentNode.add('actualEndTimeStep',self.actualEndTs)
# # Get the parent element tree (xml object) to retrieve the information needed to create the new inputs
# if(jobObject.identifier == self.TreeInfo[self.rootToJob[jobObject.identifier]].getrootnode().name): endInfo['parentNode'] = self.TreeInfo[self.rootToJob[jobObject.identifier]].getrootnode()
# else: endInfo['parentNode'] = list(self.TreeInfo[self.rootToJob[jobObject.identifier]].getrootnode().iter(jobObject.identifier))[0]
endInfo['parentNode'] = parentNode
# get the branchedLevel dictionary
branchedLevel = {}
for distk, distpb in zip(endInfo['parentNode'].get('SampledVarsPb').keys(),endInfo['parentNode'].get('SampledVarsPb').values()):
#for distk, distpb in zip(endInfo['parentNode'].get('initiatorDistribution'),endInfo['parentNode'].get('PbThreshold')):
if distk not in self.epistemicVariables.keys(): branchedLevel[distk] = utils.index(self.branchProbabilities[distk],distpb)
if not branchedLevel: self.raiseAnError(RuntimeError,'branchedLevel of node '+jobObject.identifier+'not found!')
# Loop of the parameters that have been changed after a trigger gets activated
for key in endInfo['branchChangedParams']:
endInfo['n_branches'] = 1 + int(len(endInfo['branchChangedParams'][key]['actualValue']))
if(len(endInfo['branchChangedParams'][key]['actualValue']) > 1):
# Multi-Branch mode => the resulting branches from this parent calculation (just ended)
# will be more then 2
# unchangedPb = probability (not conditional probability yet) that the event does not occur
unchangedPb = 0.0
try:
# changed_pb = probability (not conditional probability yet) that the event A occurs and the final state is 'alpha' """
for pb in xrange(len(endInfo['branchChangedParams'][key]['associatedProbability'])): unchangedPb = unchangedPb + endInfo['branchChangedParams'][key]['associatedProbability'][pb]
except KeyError: self.raiseAWarning("KeyError:"+str(key))
if(unchangedPb <= 1): endInfo['branchChangedParams'][key]['unchangedPb'] = 1.0-unchangedPb
else: self.raiseAWarning("unchangedPb > 1:"+str(unchangedPb))
else:
# Two-Way mode => the resulting branches from this parent calculation (just ended) = 2
if branchedLevel[endInfo['branchDist']] > len(self.branchProbabilities[endInfo['branchDist']])-1: pb = 1.0
else: pb = self.branchProbabilities[endInfo['branchDist']][branchedLevel[endInfo['branchDist']]]
endInfo['branchChangedParams'][key]['unchangedPb'] = 1.0 - pb
endInfo['branchChangedParams'][key]['associatedProbability'] = [pb]
self.branchCountOnLevel = 0
# # set runEnded and running to true and false respectively
# endInfo['parentNode'].add('runEnded',True)
# endInfo['parentNode'].add('running',False)
# endInfo['parentNode'].add('endTime',self.actualEndTime)
# The branchedLevel counter is updated
if branchedLevel[endInfo['branchDist']] < len(self.branchProbabilities[endInfo['branchDist']]): branchedLevel[endInfo['branchDist']] += 1
# Append the parent branchedLevel (updated for the new branch/es) in the list tha contains them
# (it is needed in order to avoid overlapping among info coming from different parent calculations)
# When this info is used, they are popped out
self.branchedLevel.append(branchedLevel)
# Append the parent end info in the list tha contains them
# (it is needed in order to avoid overlapping among info coming from different parent calculations)
# When this info is used, they are popped out
self.endInfo.append(endInfo)
# Compute conditional probability
self.computeConditionalProbability()
# Create the inputs and put them in the runQueue dictionary (if genRunQueue is true)
if genRunQueue: self._createRunningQueue(model,myInput)
return True
