def __call__( self, y_score, y_true = NotSet ):
"""
output -> The output space generated by the classifier.
target -> The targets which should be returned by the classifier.
"""
if y_true is NotSet:
self.sps = npCurrent.fp_array( y_score[0] )
self.pds = npCurrent.fp_array( y_score[1] )
self.pfs = npCurrent.fp_array( y_score[2] )
self.thresholds = npCurrent.fp_array( y_score[3] )
else:
# We have to determine what is signal and noise from the datasets using
# the targets:
try:
from sklearn.metrics import roc_curve
except ImportError:
# FIXME Can use previous function that we used here as an alternative
raise ImportError("sklearn is not available, please install it.")
self.pfs, self.pds, self.thresholds = roc_curve(y_true, y_score, pos_label=1, drop_intermediate=True)
pds = self.pds
bps = 1. - self.pfs
self.sps = np.sqrt( ( pds + bps )*.5 * np.sqrt( pds * bps ) )
def __discr_to_dict(self, model):
"""
Transform discriminators to dictionary
"""
if coreConf() is TuningToolCores.keras:
hw, hb = model.get_layer(name='dense_2').get_weights()
ow, ob = model.get_layer(name='dense_3').get_weights()
discrDict = {
'nodes':
npCurrent.int_array([hw.shape[0], hw.shape[1], ow.shape[1]]),
'weights':
np.concatenate(
[hw.reshape(-1, order='F'),
ow.reshape(-1, order='F')]),
'bias':
np.concatenate(
[hb.reshape(-1, order='F'),
ob.reshape(-1, order='F')]),
}
elif coreConf() is TuningToolCores.FastNet:
n = []
w = []
b = []
for l in range(model.getNumLayers()):
n.append(model.getNumNodes(l))
for l in range(len(n) - 1):
for j in range(n[l + 1]):
for k in range(n[l]):
w.append(model.getWeight(l, j, k))
b.append(model.getBias(l, j))
discrDict = {
'nodes': npCurrent.int_array(n),
'weights': npCurrent.fp_array(w),
'bias': npCurrent.fp_array(b)
}
self._debug('Extracted discriminator to raw dictionary.')
return discrDict
def revert(self, trnData, valData, tstData=None, **kw):
"""
Revert sort using the training, validation and testing datasets.
data = cross.revert( trnData, valData[, tstData=None], sort = sortValue)
"""
from math import floor
if not self.isRevertible():
self._logger.fatal(
"Attempted to revert CrossValidation method which reverse method was not implemented.",
NotImplementedError)
try:
sort = kw.pop('sort')
except:
TypeError('Needed argument "sort" not specified')
data = []
if not tstData:
tstData = [npCurrent.fp_array([]) for i in range(len(trnData))]
for cTrnData, cValData, cTstData in zip(trnData, valData, tstData):
# Retrieve total number of events:
evts = cTrnData.shape[npCurrent.odim] \
+ cValData.shape[npCurrent.odim] \
+ (cTstData.shape[npCurrent.odim] if cTstData.size else 0)
# Allocate the numpy array to hold
cData = npCurrent.fp_zeros(shape=npCurrent.shape(
npat=cTrnData.shape[npCurrent.pdim], nobs=evts))
# Calculate the remainder when we do equal splits in nBoxes:
remainder = evts % self._nBoxes
# The number of events in each splitted box:
evtsPerBox = floor(evts / self._nBoxes)
# Create a holder for the remainder events, which must be in the end of the
# data array
remainderData = npCurrent.fp_zeros(shape=npCurrent.shape(
npat=cTrnData.shape[npCurrent.pdim], nobs=remainder))
for boxIdx in range(self._nBoxes):
# Get the indexes where we will put our data in cData:
cStartPos = boxIdx * evtsPerBox
cEndPos = cStartPos + evtsPerBox
# And get the indexes and dataset where we will copy the values from:
startPos, endPos, ds = self.getBoxPosition(
sort,
boxIdx,
cTrnData,
cValData,
cTstData,
evtsPerBox=evtsPerBox,
remainder=remainder)
# Copy this box values to data:
cData[npCurrent.access(
pidx=':', oidx=slice(cStartPos,
cEndPos))] = ds[npCurrent.access(
pidx=':',
oidx=slice(startPos, endPos))]
# We also want to copy this box remainder if it exists to the remainder
# data:
if boxIdx < remainder:
# Take the row added to the end of dataset:
remainderData[npCurrent.access(
pidx=':',
oidx=boxIdx)] = ds[npCurrent.access(pidx=':',
oidx=endPos)]
# We finished looping over the boxes, now we copy the remainder data to
# the last positions of our original data np.array:
if remainder:
cData[npCurrent.access(pidx=':',
oidx=slice(evtsPerBox * self._nBoxes,
None))] = remainderData
# Finally, append the numpy array holding this class information to the
# data list:
data.append(cData)
return data
def __call__( self, fList, ringerOperation, **kw):
"""
Read ntuple and return patterns and efficiencies.
Arguments:
- fList: The file path or file list path. It can be an argument list of
two types:
o List: each element is a string path to the file;
o Comma separated string: each path is separated via a comma
o Folders: Expand folders recursively adding also files within them to analysis
- ringerOperation: Set Operation type. It can be both a string or the
RingerOperation
Optional arguments:
- filterType [None]: whether to filter. Use FilterType enumeration
- reference [Truth]: set reference for targets. Use Reference enumeration
- treePath [Set using operation]: set tree name on file, this may be set to
use different sources then the default.
Default for:
o Offline: Offline/Egamma/Ntuple/electron
o L2: Trigger/HLT/Egamma/TPNtuple/e24_medium_L1EM18VH
- l1EmClusCut [None]: Set L1 cluster energy cut if operating on the trigger
- l2EtCut [None]: Set L2 cluster energy cut value if operating on the trigger
- offEtCut [None]: Set Offline cluster energy cut value
- nClusters [None]: Read up to nClusters. Use None to run for all clusters.
- getRatesOnly [False]: Read up to nClusters. Use None to run for all clusters.
- etBins [None]: E_T bins (GeV) where the data should be segmented
- etaBins [None]: eta bins where the data should be segmented
- ringConfig [100]: A list containing the number of rings available in the data
for each eta bin.
- crossVal [None]: Whether to measure benchmark efficiency splitting it
by the crossVal-validation datasets
- extractDet [None]: Which detector to export (use Detector enumeration).
Defaults are:
o L2Calo: Calorimetry
o L2: Tracking
o Offline: Calorimetry
o Others: CaloAndTrack
- standardCaloVariables [False]: Whether to extract standard track variables.
- useTRT [False]: Whether to export TRT information when dumping track
variables.
- supportTriggers [True]: Whether reading data comes from support triggers
"""
# Offline information branches:
__offlineBranches = ['el_et',
'el_eta',
#'el_loose',
#'el_medium',
#'el_tight',
'el_lhLoose',
'el_lhMedium',
'el_lhTight',
'mc_hasMC',
'mc_isElectron',
'mc_hasZMother',
'el_nPileupPrimaryVtx',
]
# Online information branches
__onlineBranches = []
__l2stdCaloBranches = ['trig_L2_calo_et',
'trig_L2_calo_eta',
'trig_L2_calo_phi',
'trig_L2_calo_e237', # rEta
'trig_L2_calo_e277', # rEta
'trig_L2_calo_fracs1', # F1: fraction sample 1
'trig_L2_calo_weta2', # weta2
'trig_L2_calo_ehad1', # energy on hadronic sample 1
'trig_L2_calo_emaxs1', # eratio
'trig_L2_calo_e2tsts1', # eratio
'trig_L2_calo_wstot',] # wstot
__l2trackBranches = [ # Do not add non patter variables on this branch list
#'trig_L2_el_pt',
#'trig_L2_el_eta',
#'trig_L2_el_phi',
#'trig_L2_el_caloEta',
#'trig_L2_el_charge',
#'trig_L2_el_nTRTHits',
#'trig_L2_el_nTRTHiThresholdHits',
'trig_L2_el_etOverPt',
'trig_L2_el_trkClusDeta',
'trig_L2_el_trkClusDphi',]
# Retrieve information from keyword arguments
filterType = retrieve_kw(kw, 'filterType', FilterType.DoNotFilter )
reference = retrieve_kw(kw, 'reference', Reference.Truth )
l1EmClusCut = retrieve_kw(kw, 'l1EmClusCut', None )
l2EtCut = retrieve_kw(kw, 'l2EtCut', None )
efEtCut = retrieve_kw(kw, 'efEtCut', None )
offEtCut = retrieve_kw(kw, 'offEtCut', None )
treePath = retrieve_kw(kw, 'treePath', None )
nClusters = retrieve_kw(kw, 'nClusters', None )
getRates = retrieve_kw(kw, 'getRates', True )
getRatesOnly = retrieve_kw(kw, 'getRatesOnly', False )
etBins = retrieve_kw(kw, 'etBins', None )
etaBins = retrieve_kw(kw, 'etaBins', None )
crossVal = retrieve_kw(kw, 'crossVal', None )
ringConfig = retrieve_kw(kw, 'ringConfig', 100 )
extractDet = retrieve_kw(kw, 'extractDet', None )
standardCaloVariables = retrieve_kw(kw, 'standardCaloVariables', False )
useTRT = retrieve_kw(kw, 'useTRT', False )
supportTriggers = retrieve_kw(kw, 'supportTriggers', True )
monitoring = retrieve_kw(kw, 'monitoring', None )
pileupRef = retrieve_kw(kw, 'pileupRef', NotSet )
import ROOT, pkgutil
#gROOT.ProcessLine (".x $ROOTCOREDIR/scripts/load_packages.C");
#ROOT.gROOT.Macro('$ROOTCOREDIR/scripts/load_packages.C')
if not( bool( pkgutil.find_loader( 'libTuningTools' ) ) and ROOT.gSystem.Load('libTuningTools') >= 0 ) and \
not( bool( pkgutil.find_loader( 'libTuningToolsLib' ) ) and ROOT.gSystem.Load('libTuningToolsLib') >= 0 ):
#ROOT.gSystem.Load('libTuningToolsPythonLib') < 0:
self._fatal("Could not load TuningTools library", ImportError)
if 'level' in kw: self.level = kw.pop('level')
# and delete it to avoid mistakes:
checkForUnusedVars( kw, self._warning )
del kw
### Parse arguments
# Mutual exclusive arguments:
if not getRates and getRatesOnly:
self._logger.error("Cannot run with getRates set to False and getRatesOnly set to True. Setting getRates to True.")
getRates = True
# Also parse operation, check if its type is string and if we can
# transform it to the known operation enum:
fList = csvStr2List ( fList )
fList = expandFolders( fList )
ringerOperation = RingerOperation.retrieve(ringerOperation)
reference = Reference.retrieve(reference)
if isinstance(l1EmClusCut, str):
l1EmClusCut = float(l1EmClusCut)
if l1EmClusCut:
l1EmClusCut = 1000.*l1EmClusCut # Put energy in MeV
__onlineBranches.append( 'trig_L1_emClus' )
if l2EtCut:
l2EtCut = 1000.*l2EtCut # Put energy in MeV
__onlineBranches.append( 'trig_L2_calo_et' )
if efEtCut:
efEtCut = 1000.*efEtCut # Put energy in MeV
__onlineBranches.append( 'trig_EF_calo_et' )
if offEtCut:
offEtCut = 1000.*offEtCut # Put energy in MeV
__offlineBranches.append( 'el_et' )
if not supportTriggers:
__onlineBranches.append( 'trig_L1_accept' )
# Check if treePath is None and try to set it automatically
if treePath is None:
treePath = 'Offline/Egamma/Ntuple/electron' if ringerOperation < 0 else \
'Trigger/HLT/Egamma/TPNtuple/e24_medium_L1EM18VH'
# Check whether using bins
useBins=False; useEtBins=False; useEtaBins=False
nEtaBins = 1; nEtBins = 1
# Set the detector which we should extract the information:
if extractDet is None:
if ringerOperation < 0:
extractDet = Detector.Calorimetry
elif ringerOperation is RingerOperation.L2Calo:
extractDet = Detector.Calorimetry
elif ringerOperation is RingerOperation.L2:
extractDet = Detector.Tracking
else:
extractDet = Detector.CaloAndTrack
else:
extractDet = Detector.retrieve( extractDet )
if etaBins is None: etaBins = npCurrent.fp_array([])
if type(etaBins) is list: etaBins=npCurrent.fp_array(etaBins)
if etBins is None: etBins = npCurrent.fp_array([])
if type(etBins) is list: etBins=npCurrent.fp_array(etBins)
if etBins.size:
etBins = etBins * 1000. # Put energy in MeV
nEtBins = len(etBins)-1
if nEtBins >= np.iinfo(npCurrent.scounter_dtype).max:
self._fatal(('Number of et bins (%d) is larger or equal than maximum '
'integer precision can hold (%d). Increase '
'TuningTools.coreDef.npCurrent scounter_dtype number of bytes.'), nEtBins,
np.iinfo(npCurrent.scounter_dtype).max)
# Flag that we are separating data through bins
useBins=True
useEtBins=True
self._debug('E_T bins enabled.')
if not type(ringConfig) is list and not type(ringConfig) is np.ndarray:
ringConfig = [ringConfig] * (len(etaBins) - 1) if etaBins.size else 1
if type(ringConfig) is list: ringConfig=npCurrent.int_array(ringConfig)
if not len(ringConfig):
self._fatal('Rings size must be specified.');
if etaBins.size:
nEtaBins = len(etaBins)-1
if nEtaBins >= np.iinfo(npCurrent.scounter_dtype).max:
self._fatal(('Number of eta bins (%d) is larger or equal than maximum '
'integer precision can hold (%d). Increase '
'TuningTools.coreDef.npCurrent scounter_dtype number of bytes.'), nEtaBins,
np.iinfo(npCurrent.scounter_dtype).max)
if len(ringConfig) != nEtaBins:
self._fatal(('The number of rings configurations (%r) must be equal than '
'eta bins (%r) region config'),ringConfig, etaBins)
useBins=True
useEtaBins=True
self._debug('eta bins enabled.')
else:
self._debug('eta/et bins disabled.')
### Prepare to loop:
# Open root file
t = ROOT.TChain(treePath)
for inputFile in progressbar(fList, len(fList),
logger = self._logger,
prefix = "Creating collection tree "):
# Check if file exists
f = ROOT.TFile.Open(inputFile, 'read')
if not f or f.IsZombie():
self._warning('Couldn''t open file: %s', inputFile)
continue
# Inform user whether TTree exists, and which options are available:
self._debug("Adding file: %s", inputFile)
obj = f.Get(treePath)
if not obj:
self._warning("Couldn't retrieve TTree (%s)!", treePath)
self._info("File available info:")
f.ReadAll()
f.ReadKeys()
f.ls()
continue
elif not isinstance(obj, ROOT.TTree):
self._fatal("%s is not an instance of TTree!", treePath, ValueError)
t.Add( inputFile )
# Turn all branches off.
t.SetBranchStatus("*", False)
# RingerPhysVal hold the address of required branches
event = ROOT.RingerPhysVal()
# Add offline branches, these are always needed
cPos = 0
for var in __offlineBranches:
self.__setBranchAddress(t,var,event)
# Add online branches if using Trigger
if ringerOperation > 0:
for var in __onlineBranches:
self.__setBranchAddress(t,var,event)
## Allocating memory for the number of entries
entries = t.GetEntries()
nobs = entries if (nClusters is None or nClusters > entries or nClusters < 1) \
else nClusters
## Retrieve the dependent operation variables:
if useEtBins:
etBranch = 'el_et' if ringerOperation < 0 else 'trig_L2_calo_et'
self.__setBranchAddress(t,etBranch,event)
self._debug("Added branch: %s", etBranch)
if not getRatesOnly:
npEt = npCurrent.scounter_zeros(shape=npCurrent.shape(npat = 1, nobs = nobs))
self._debug("Allocated npEt with size %r", npEt.shape)
if useEtaBins:
etaBranch = "el_eta" if ringerOperation < 0 else "trig_L2_calo_eta"
self.__setBranchAddress(t,etaBranch,event)
self._debug("Added branch: %s", etaBranch)
if not getRatesOnly:
npEta = npCurrent.scounter_zeros(shape=npCurrent.shape(npat = 1, nobs = nobs))
self._debug("Allocated npEta with size %r", npEta.shape)
# The base information holder, such as et, eta and pile-up
if pileupRef is NotSet:
if ringerOperation > 0:
pileupRef = PileupReference.avgmu
else:
pileupRef = PileupReference.nvtx
pileupRef = PileupReference.retrieve( pileupRef )
self._info("Using '%s' as pile-up reference.", PileupReference.tostring( pileupRef ) )
if pileupRef is PileupReference.nvtx:
pileupBranch = 'el_nPileupPrimaryVtx'
pileupDataType = np.uint16
elif pileupRef is PileupReference.avgmu:
pileupBranch = 'avgmu'
pileupDataType = np.float32
else:
raise NotImplementedError("Pile-up reference %r is not implemented." % pileupRef)
baseInfoBranch = BaseInfo((etBranch, etaBranch, pileupBranch, 'el_phi' if ringerOperation < 0 else 'trig_L2_el_phi',),
(npCurrent.fp_dtype, npCurrent.fp_dtype, npCurrent.fp_dtype, pileupDataType) )
baseInfo = [None, ] * baseInfoBranch.nInfo
# Make sure all baseInfoBranch information is available:
for idx in baseInfoBranch:
self.__setBranchAddress(t,baseInfoBranch.retrieveBranch(idx),event)
# Allocate numpy to hold as many entries as possible:
if not getRatesOnly:
# Retrieve the rings information depending on ringer operation
ringerBranch = "el_ringsE" if ringerOperation < 0 else \
"trig_L2_calo_rings"
self.__setBranchAddress(t,ringerBranch,event)
if ringerOperation > 0:
if ringerOperation is RingerOperation.L2:
for var in __l2trackBranches:
self.__setBranchAddress(t,var,event)
if standardCaloVariables:
if ringerOperation in (RingerOperation.L2, RingerOperation.L2Calo,):
for var in __l2stdCaloBranches:
self.__setBranchAddress(t, var, event)
else:
self._warning("Unknown standard calorimeters for Operation:%s. Setting operation back to use rings variables.",
RingerOperation.tostring(ringerOperation))
t.GetEntry(0)
npat = 0
if extractDet in (Detector.Calorimetry,
Detector.CaloAndTrack,
Detector.All):
if standardCaloVariables:
npat+= 6
else:
npat += ringConfig.max()
if extractDet in (Detector.Tracking,
Detector.CaloAndTrack,
Detector.All):
if ringerOperation is RingerOperation.L2:
if useTRT:
self._info("Using TRT information!")
npat += 2
__l2trackBranches.append('trig_L2_el_nTRTHits')
__l2trackBranches.append('trig_L2_el_nTRTHiThresholdHits')
npat += 3
for var in __l2trackBranches:
self.__setBranchAddress(t,var,event)
self.__setBranchAddress(t,"trig_L2_el_pt",event)
elif ringerOperation < 0: # Offline
self._warning("Still need to implement tracking for the ringer offline.")
npPatterns = npCurrent.fp_zeros( shape=npCurrent.shape(npat=npat, #getattr(event, ringerBranch).size()
nobs=nobs)
)
self._debug("Allocated npPatterns with size %r", npPatterns.shape)
# Add E_T, eta and luminosity information
npBaseInfo = [npCurrent.zeros( shape=npCurrent.shape(npat=1, nobs=nobs ), dtype=baseInfoBranch.dtype(idx) )
for idx in baseInfoBranch]
else:
npPatterns = npCurrent.fp_array([])
npBaseInfo = [deepcopy(npCurrent.fp_array([])) for _ in baseInfoBranch]
## Allocate the branch efficiency collectors:
if getRates:
if ringerOperation < 0:
benchmarkDict = OrderedDict(
[( RingerOperation.Offline_CutBased_Loose , 'el_loose' ),
( RingerOperation.Offline_CutBased_Medium , 'el_medium' ),
( RingerOperation.Offline_CutBased_Tight , 'el_tight' ),
( RingerOperation.Offline_LH_Loose , 'el_lhLoose' ),
( RingerOperation.Offline_LH_Medium , 'el_lhMedium' ),
( RingerOperation.Offline_LH_Tight , 'el_lhTight' ),
])
else:
benchmarkDict = OrderedDict(
[( RingerOperation.L2Calo , 'trig_L2_calo_accept' ),
( RingerOperation.L2 , 'trig_L2_el_accept' ),
( RingerOperation.EFCalo , 'trig_EF_calo_accept' ),
( RingerOperation.HLT , 'trig_EF_el_accept' ),
])
from TuningTools.CreateData import BranchEffCollector, BranchCrossEffCollector
branchEffCollectors = OrderedDict()
branchCrossEffCollectors = OrderedDict()
for key, val in benchmarkDict.iteritems():
branchEffCollectors[key] = list()
branchCrossEffCollectors[key] = list()
# Add efficincy branch:
if getRates or getRatesOnly:
self.__setBranchAddress(t,val,event)
for etBin in range(nEtBins):
if useBins:
branchEffCollectors[key].append(list())
branchCrossEffCollectors[key].append(list())
for etaBin in range(nEtaBins):
etBinArg = etBin if useBins else -1
etaBinArg = etaBin if useBins else -1
argList = [ RingerOperation.tostring(key), val, etBinArg, etaBinArg ]
branchEffCollectors[key][etBin].append(BranchEffCollector( *argList ) )
if crossVal:
branchCrossEffCollectors[key][etBin].append(BranchCrossEffCollector( entries, crossVal, *argList ) )
# etBin
# etaBin
# benchmark dict
if self._logger.isEnabledFor( LoggingLevel.DEBUG ):
self._debug( 'Retrieved following branch efficiency collectors: %r',
[collector[0].printName for collector in traverse(branchEffCollectors.values())])
# end of (getRates)
etaBin = 0; etBin = 0
step = int(entries/100) if int(entries/100) > 0 else 1
## Start loop!
self._info("There is available a total of %d entries.", entries)
for entry in progressbar(range(entries), entries,
step = step, logger = self._logger,
prefix = "Looping over entries "):
#self._verbose('Processing eventNumber: %d/%d', entry, entries)
t.GetEntry(entry)
# Check if it is needed to remove energy regions (this means that if not
# within this range, it will be ignored as well for efficiency measuremnet)
if event.el_et < offEtCut:
self._verbose("Ignoring entry due to offline E_T cut.")
continue
# Add et distribution for all events
if not monitoring is None:
# Book all distribtions before the event selection
self.__fillHistograms(monitoring,filterType,event,False)
if ringerOperation > 0:
# Remove events which didn't pass L1_calo
if not supportTriggers and not event.trig_L1_accept:
#self._verbose("Ignoring entry due to L1Calo cut (trig_L1_accept = %r).", event.trig_L1_accept)
continue
if event.trig_L1_emClus < l1EmClusCut:
#self._verbose("Ignoring entry due to L1Calo E_T cut (%d < %r).", event.trig_L1_emClus, l1EmClusCut)
continue
if event.trig_L2_calo_et < l2EtCut:
#self._verbose("Ignoring entry due to L2Calo E_T cut.")
continue
if efEtCut is not None and event.trig_L2_calo_accept :
# EF calo is a container, search for electrons objects with et > cut
trig_EF_calo_et_list = stdvector_to_list(event.trig_EF_calo_et)
found=False
for v in trig_EF_calo_et_list:
if v < efEtCut: found=True
if found:
#self._verbose("Ignoring entry due to EFCalo E_T cut.")
continue
# Set discriminator target:
target = Target.Unknown
if reference is Reference.Truth:
if event.mc_isElectron and event.mc_hasZMother:
target = Target.Signal
elif not (event.mc_isElectron and (event.mc_hasZMother or event.mc_hasWMother) ):
target = Target.Background
elif reference is Reference.Off_Likelihood:
if event.el_lhTight: target = Target.Signal
elif not event.el_lhLoose: target = Target.Background
elif reference is Reference.AcceptAll:
target = Target.Signal if filterType is FilterType.Signal else Target.Background
else:
if event.el_tight: target = Target.Signal
elif not event.el_loose: target = Target.Background
# Run filter if it is defined
if filterType and \
( (filterType is FilterType.Signal and target != Target.Signal) or \
(filterType is FilterType.Background and target != Target.Background) or \
(target == Target.Unknown) ):
#self._verbose("Ignoring entry due to filter cut.")
continue
# Add et distribution for all events
if not monitoring is None:
# Book all distributions after the event selection
self.__fillHistograms(monitoring,filterType,event,True)
# Retrieve base information:
for idx in baseInfoBranch:
lInfo = getattr(event, baseInfoBranch.retrieveBranch(idx))
baseInfo[idx] = lInfo
if not getRatesOnly: npBaseInfo[idx][cPos] = lInfo
# Retrieve dependent operation region
if useEtBins:
etBin = self.__retrieveBinIdx( etBins, baseInfo[0] )
if useEtaBins:
etaBin = self.__retrieveBinIdx( etaBins, np.fabs( baseInfo[1]) )
# Check if bin is within range (when not using bins, this will always be true):
if (etBin < nEtBins and etaBin < nEtaBins):
# Retrieve patterns:
if not getRatesOnly:
if useEtBins: npEt[cPos] = etBin
if useEtaBins: npEta[cPos] = etaBin
## Retrieve calorimeter information:
cPat = 0
caloAvailable = True
if extractDet in (Detector.Calorimetry,
Detector.CaloAndTrack,
Detector.All):
if standardCaloVariables:
patterns = []
if ringerOperation is RingerOperation.L2Calo:
from math import cosh
cosh_eta = cosh( event.trig_L2_calo_eta )
# second layer ratio between 3x7 7x7
rEta = event.trig_L2_calo_e237 / event.trig_L2_calo_e277
base = event.trig_L2_calo_emaxs1 + event.trig_L2_calo_e2tsts1
# Ratio between first and second highest energy cells
eRatio = ( event.trig_L2_calo_emaxs1 - event.trig_L2_calo_e2tsts1 ) / base if base > 0 else 0
# ratio of energy in the first layer (hadronic particles should leave low energy)
F1 = event.trig_L2_calo_fracs1 / ( event.trig_L2_calo_et * cosh_eta )
# weta2 is calculated over the middle layer using 3 x 5
weta2 = event.trig_L2_calo_weta2
# wstot is calculated over the first layer using (typically) 20 strips
wstot = event.trig_L2_calo_wstot
# ratio between EM cluster and first hadronic layers:
Rhad1 = ( event.trig_L2_calo_ehad1 / cosh_eta ) / event.trig_L2_calo_et
# allocate patterns:
patterns = [rEta, eRatio, F1, weta2, wstot, Rhad1]
for pat in patterns:
npPatterns[npCurrent.access( pidx=cPat, oidx=cPos) ] = pat
cPat += 1
# end of ringerOperation
else:
# Remove events without rings
if getattr(event,ringerBranch).empty():
caloAvailable = False
# Retrieve rings:
if caloAvailable:
try:
patterns = stdvector_to_list( getattr(event,ringerBranch) )
lPat = len(patterns)
if lPat == ringConfig[etaBin]:
npPatterns[npCurrent.access(pidx=slice(cPat,ringConfig[etaBin]),oidx=cPos)] = patterns
else:
oldEtaBin = etaBin
if etaBin > 0 and ringConfig[etaBin - 1] == lPat:
etaBin -= 1
elif etaBin + 1 < len(ringConfig) and ringConfig[etaBin + 1] == lPat:
etaBin += 1
npPatterns[npCurrent.access(pidx=slice(cPat, ringConfig[etaBin]),oidx=cPos)] = patterns
self._warning(("Recovered event which should be within eta bin (%d: %r) "
"but was found to be within eta bin (%d: %r). "
"Its read eta value was of %f."),
oldEtaBin, etaBins[oldEtaBin:oldEtaBin+2],
etaBin, etaBins[etaBin:etaBin+2],
np.fabs( getattr(event,etaBranch)))
except ValueError:
self._logger.error(("Patterns size (%d) do not match expected "
"value (%d). This event eta value is: %f, and ringConfig is %r."),
lPat, ringConfig[etaBin], np.fabs( getattr(event,etaBranch)), ringConfig
)
continue
else:
if extractDet is Detector.Calorimetry:
# Also display warning when extracting only calorimetry!
self._warning("Rings not available")
continue
self._warning("Rings not available")
continue
cPat += ringConfig.max()
# which calo variables
# end of (extractDet needed calorimeter)
# And track information:
if extractDet in (Detector.Tracking,
Detector.CaloAndTrack,
Detector.All):
if caloAvailable or extractDet is Detector.Tracking:
if ringerOperation is RingerOperation.L2:
# Retrieve nearest deta/dphi only, so we need to find each one is the nearest:
if event.trig_L2_el_trkClusDeta.size():
clusDeta = npCurrent.fp_array( stdvector_to_list( event.trig_L2_el_trkClusDeta ) )
clusDphi = npCurrent.fp_array( stdvector_to_list( event.trig_L2_el_trkClusDphi ) )
bestTrackPos = int( np.argmin( clusDeta**2 + clusDphi**2 ) )
for var in __l2trackBranches:
npPatterns[npCurrent.access( pidx=cPat,oidx=cPos) ] = getattr(event, var)[bestTrackPos]
cPat += 1
else:
#self._verbose("Ignoring entry due to track information not available.")
continue
#for var in __l2trackBranches:
# npPatterns[npCurrent.access( pidx=cPat,oidx=cPos) ] = np.nan
# cPat += 1
elif ringerOperation < 0: # Offline
pass
# caloAvailable or only tracking
# end of (extractDet needs tracking)
# end of (getRatesOnly)
## Retrieve rates information:
if getRates:
for branch in branchEffCollectors.itervalues():
if not useBins:
branch.update(event)
else:
branch[etBin][etaBin].update(event)
if crossVal:
for branchCross in branchCrossEffCollectors.itervalues():
if not useBins:
branchCross.update(event)
else:
branchCross[etBin][etaBin].update(event)
# end of (getRates)
# We only increment if this cluster will be computed
cPos += 1
# end of (et/eta bins)
# Limit the number of entries to nClusters if desired and possible:
if not nClusters is None and cPos >= nClusters:
break
# for end
## Treat the rings information
if not getRatesOnly:
## Remove not filled reserved memory space:
if npPatterns.shape[npCurrent.odim] > cPos:
npPatterns = np.delete( npPatterns, slice(cPos,None), axis = npCurrent.odim)
## Segment data over bins regions:
# Also remove not filled reserved memory space:
if useEtBins:
npEt = npCurrent.delete( npEt, slice(cPos,None))
if useEtaBins:
npEta = npCurrent.delete( npEta, slice(cPos,None))
# Treat
npObject = self.treatNpInfo(cPos, npEt, npEta, useEtBins, useEtaBins,
nEtBins, nEtaBins, standardCaloVariables, ringConfig,
npPatterns, )
data = [self.treatNpInfo(cPos, npEt, npEta, useEtBins, useEtaBins,
nEtBins, nEtaBins, standardCaloVariables, ringConfig,
npData) for npData in npBaseInfo]
npBaseInfo = npCurrent.array( data, dtype=np.object )
else:
npObject = npCurrent.array([], dtype=npCurrent.dtype)
# not getRatesOnly
if getRates:
if crossVal:
for etBin in range(nEtBins):
for etaBin in range(nEtaBins):
for branchCross in branchCrossEffCollectors.itervalues():
if not useBins:
branchCross.finished()
else:
branchCross[etBin][etaBin].finished()
# Print efficiency for each one for the efficiency branches analysed:
for etBin in range(nEtBins) if useBins else range(1):
for etaBin in range(nEtaBins) if useBins else range(1):
for branch in branchEffCollectors.itervalues():
lBranch = branch if not useBins else branch[etBin][etaBin]
self._info('%s',lBranch)
if crossVal:
for branchCross in branchCrossEffCollectors.itervalues():
lBranchCross = branchCross if not useBins else branchCross[etBin][etaBin]
lBranchCross.dump(self._debug, printSort = True,
sortFcn = self._verbose)
# for branch
# for eta
# for et
# end of (getRates)
outputs = []
#if not getRatesOnly:
outputs.extend((npObject, npBaseInfo))
#if getRates:
outputs.extend((branchEffCollectors, branchCrossEffCollectors))
#outputs = tuple(outputs)
return outputs
def __retrieveBinIdx( self, bins, value ): return npCurrent.scounter_dtype.type(np.digitize(npCurrent.fp_array([value]), bins)[0]-1)
def __call__(self, fList, ringerOperation, **kw):
"""
Read ntuple and return patterns and efficiencies.
Arguments:
- fList: The file path or file list path. It can be an argument list of
two types:
o List: each element is a string path to the file;
o Comma separated string: each path is separated via a comma
o Folders: Expand folders recursively adding also files within them to analysis
- ringerOperation: Set Operation type. It can be both a string or the
RingerOperation
Optional arguments:
- filterType [None]: whether to filter. Use FilterType enumeration
- reference [Truth]: set reference for targets. Use Reference enumeration
- treePath [Set using operation]: set tree name on file, this may be set to
use different sources then the default.
Default for:
o Offline: Offline/Egamma/Ntuple/electron
o L2: Trigger/HLT/Egamma/TPNtuple/e24_medium_L1EM18VH
- l1EmClusCut [None]: Set L1 cluster energy cut if operating on the trigger
- l2EtCut [None]: Set L2 cluster energy cut value if operating on the trigger
- offEtCut [None]: Set Offline cluster energy cut value
- nClusters [None]: Read up to nClusters. Use None to run for all clusters.
- getRatesOnly [False]: Read up to nClusters. Use None to run for all clusters.
- etBins [None]: E_T bins (GeV) where the data should be segmented
- etaBins [None]: eta bins where the data should be segmented
- ringConfig [100]: A list containing the number of rings available in the data
for each eta bin.
- crossVal [None]: Whether to measure benchmark efficiency splitting it
by the crossVal-validation datasets
- extractDet [None]: Which detector to export (use Detector enumeration).
Defaults are:
o L2Calo: Calorimetry
o L2: Tracking
o Offline: Calorimetry
o Others: CaloAndTrack
- standardCaloVariables [False]: Whether to extract standard track variables.
- useTRT [False]: Whether to export TRT information when dumping track
variables.
- supportTriggers [True]: Whether reading data comes from support triggers
"""
__eventBranches = [
'EventNumber', 'RunNumber', 'RandomRunNumber', 'MCChannelNumber',
'RandomLumiBlockNumber', 'MCPileupWeight', 'VertexZPosition',
'Zcand_M', 'Zcand_pt', 'Zcand_eta', 'Zcand_phi', 'Zcand_y',
'isTagTag'
]
__trackBranches = [
'elCand2_deltaeta1', 'elCand2_DeltaPOverP',
'elCand2_deltaphiRescaled', 'elCand2_d0significance',
'elCand2_trackd0pvunbiased', 'elCand2_eProbabilityHT'
]
__monteCarloBranches = [
'type',
'origin',
'originbkg',
'typebkg',
'isTruthElectronFromZ',
'TruthParticlePdgId',
'firstEgMotherPdgId',
'TruthParticleBarcode',
'firstEgMotherBarcode',
'MotherPdgId',
'MotherBarcode',
'FirstEgMotherTyp',
'FirstEgMotherOrigin',
'dRPdgId',
]
__onlineBranches = ['match', 'ringerMatch', 'ringer_rings']
__offlineBranches = ['et', 'eta']
# The current pid map used as offline reference
pidConfigs = {
key: value
for key, value in RingerOperation.efficiencyBranches().iteritems()
if key in (RingerOperation.Offline_LH_Tight,
RingerOperation.Offline_LH_Medium,
RingerOperation.Offline_LH_Loose,
RingerOperation.Offline_LH_VeryLoose)
}
# Retrieve information from keyword arguments
filterType = retrieve_kw(kw, 'filterType', FilterType.DoNotFilter)
reference = retrieve_kw(kw, 'reference', Reference.AcceptAll)
offEtCut = retrieve_kw(kw, 'offEtCut', None)
l2EtCut = retrieve_kw(kw, 'l2EtCut', None)
treePath = retrieve_kw(kw, 'treePath', 'ZeeCandidate')
nClusters = retrieve_kw(kw, 'nClusters', None)
etBins = retrieve_kw(kw, 'etBins', None)
etaBins = retrieve_kw(kw, 'etaBins', None)
crossVal = retrieve_kw(kw, 'crossVal', None)
ringConfig = retrieve_kw(kw, 'ringConfig', 100)
monitoring = retrieve_kw(kw, 'monitoring', None)
pileupRef = retrieve_kw(kw, 'pileupRef', NotSet)
getRates = retrieve_kw(kw, 'getRates', True)
getRatesOnly = retrieve_kw(kw, 'getRatesOnly', False)
getTagsOnly = retrieve_kw(kw, 'getTagsOnly', False)
extractDet = retrieve_kw(kw, 'extractDet', None)
import ROOT
#gROOT.ProcessLine (".x $ROOTCOREDIR/scripts/load_packages.C");
#ROOT.gROOT.Macro('$ROOTCOREDIR/scripts/load_packages.C')
if ROOT.gSystem.Load('libTuningTools') < 0:
self._fatal("Could not load TuningTools library", ImportError)
if 'level' in kw: self.level = kw.pop('level')
# and delete it to avoid mistakes:
checkForUnusedVars(kw, self._warning)
del kw
### Parse arguments
# Also parse operation, check if its type is string and if we can
# transform it to the known operation enum:
fList = csvStr2List(fList)
fList = expandFolders(fList)
ringerOperation = RingerOperation.retrieve(ringerOperation)
reference = Reference.retrieve(reference)
# Offline E_T cut
if offEtCut:
offEtCut = 1000. * offEtCut # Put energy in MeV
# Check whether using bins
useBins = False
useEtBins = False
useEtaBins = False
nEtaBins = 1
nEtBins = 1
if etaBins is None: etaBins = npCurrent.fp_array([])
if type(etaBins) is list: etaBins = npCurrent.fp_array(etaBins)
if etBins is None: etBins = npCurrent.fp_array([])
if type(etBins) is list: etBins = npCurrent.fp_array(etBins)
if etBins.size:
etBins = etBins * 1000. # Put energy in MeV
nEtBins = len(etBins) - 1
if nEtBins >= np.iinfo(npCurrent.scounter_dtype).max:
self._fatal((
'Number of et bins (%d) is larger or equal than maximum '
'integer precision can hold (%d). Increase '
'TuningTools.coreDef.npCurrent scounter_dtype number of bytes.'
), nEtBins,
np.iinfo(npCurrent.scounter_dtype).max)
# Flag that we are separating data through bins
useBins = True
useEtBins = True
self._debug('E_T bins enabled.')
if not type(ringConfig) is list and not type(ringConfig) is np.ndarray:
ringConfig = [ringConfig] * (len(etaBins) -
1) if etaBins.size else 1
if type(ringConfig) is list:
ringConfig = npCurrent.int_array(ringConfig)
if not len(ringConfig):
self._fatal('Rings size must be specified.')
if etaBins.size:
nEtaBins = len(etaBins) - 1
if nEtaBins >= np.iinfo(npCurrent.scounter_dtype).max:
self._fatal((
'Number of eta bins (%d) is larger or equal than maximum '
'integer precision can hold (%d). Increase '
'TuningTools.coreDef.npCurrent scounter_dtype number of bytes.'
), nEtaBins,
np.iinfo(npCurrent.scounter_dtype).max)
if len(ringConfig) != nEtaBins:
self._fatal((
'The number of rings configurations (%r) must be equal than '
'eta bins (%r) region config'), ringConfig, etaBins)
useBins = True
useEtaBins = True
self._debug('eta bins enabled.')
else:
self._debug('eta/et bins disabled.')
# The base information holder, such as et, eta and pile-up
if pileupRef is NotSet:
if ringerOperation > 0:
pileupRef = PileupReference.avgmu
else:
pileupRef = PileupReference.nvtx
pileupRef = PileupReference.retrieve(pileupRef)
self._info("Using '%s' as pile-up reference.",
PileupReference.tostring(pileupRef))
# Candidates: (1) is tags and (2) is probes. Default is probes
self._candIdx = 1 if getTagsOnly else 2
# Mutual exclusive arguments:
if not getRates and getRatesOnly:
self._logger.error(
"Cannot run with getRates set to False and getRatesOnly set to True. Setting getRates to True."
)
getRates = True
### Prepare to loop:
t = ROOT.TChain(treePath)
for inputFile in progressbar(fList,
len(fList),
logger=self._logger,
prefix="Creating collection tree "):
# Check if file exists
f = ROOT.TFile.Open(inputFile, 'read')
if not f or f.IsZombie():
self._warning('Couldn' 't open file: %s', inputFile)
continue
# Inform user whether TTree exists, and which options are available:
self._debug("Adding file: %s", inputFile)
obj = f.Get(treePath)
if not obj:
self._warning("Couldn't retrieve TTree (%s)!", treePath)
self._info("File available info:")
f.ReadAll()
f.ReadKeys()
f.ls()
continue
elif not isinstance(obj, ROOT.TTree):
self._fatal("%s is not an instance of TTree!", treePath,
ValueError)
t.Add(inputFile)
# Turn all branches off.
t.SetBranchStatus("*", False)
# RingerPhysVal hold the address of required branches
event = ROOT.SkimmedNtuple()
# Ready to retrieve the total number of events
t.GetEntry(0)
## Allocating memory for the number of entries
entries = t.GetEntries()
nobs = entries if (nClusters is None or nClusters > entries or nClusters < 1) \
else nClusters
## Retrieve the dependent operation variables:
if useEtBins:
etBranch = ('elCand%d_et') % (
self._candIdx) if ringerOperation < 0 else ('fcCand%d_et') % (
self._candIdx)
self.__setBranchAddress(t, etBranch, event)
self._debug("Added branch: %s", etBranch)
npEt = npCurrent.scounter_zeros(
shape=npCurrent.shape(npat=1, nobs=nobs))
self._debug("Allocated npEt with size %r", npEt.shape)
if useEtaBins:
etaBranch = ('elCand%d_eta') % (
self._candIdx) if ringerOperation < 0 else ('fcCand%d_eta') % (
self._candIdx)
self.__setBranchAddress(t, etaBranch, event)
self._debug("Added branch: %s", etaBranch)
npEta = npCurrent.scounter_zeros(
shape=npCurrent.shape(npat=1, nobs=nobs))
self._debug("Allocated npEta with size %r", npEta.shape)
if reference is Reference.Truth:
self.__setBranchAddress(t, ('elCand%d_isTruthElectronFromZ') %
(self._candIdx), event)
for var in __offlineBranches:
self.__setBranchAddress(t, ('elCand%d_%s') % (self._candIdx, var),
event)
#for var in pidConfigs.values():
# self.__setBranchAddress(t,var,event)
for var in __trackBranches:
self.__setBranchAddress(t, var, event)
# Add online branches if using Trigger
if ringerOperation > 0:
for var in __onlineBranches:
self.__setBranchAddress(t,
('fcCand%d_%s') % (self._candIdx, var),
event)
else:
self.__setBranchAddress(t, ('elCand%d_%s') %
(self._candIdx, 'ringer_rings'), event)
if pileupRef is PileupReference.nvtx:
pileupBranch = 'Nvtx'
pileupDataType = np.uint16
elif pileupRef is PileupReference.avgmu:
pileupBranch = 'averageIntPerXing'
pileupDataType = np.float32
else:
raise NotImplementedError(
"Pile-up reference %r is not implemented." % pileupRef)
#for var in __eventBranches +
for var in [pileupBranch]:
self.__setBranchAddress(t, var, event)
### Allocate memory
if extractDet == (Detector.Calorimetry):
npat = ringConfig.max()
elif extractDet == (Detector.Tracking):
npat = len(__trackBranches)
# NOTE: Check if pat is correct for both Calo and track data
elif extractDet in (Detector.CaloAndTrack, Detector.All):
npat = ringConfig.max() + len(__trackBranches)
npPatterns = npCurrent.fp_zeros(shape=npCurrent.shape(
npat=npat, #getattr(event, ringerBranch).size()
nobs=nobs))
self._debug("Allocated npPatterns with size %r", npPatterns.shape)
baseInfoBranch = BaseInfo(
(etBranch, etaBranch, pileupBranch),
(npCurrent.fp_dtype, npCurrent.fp_dtype, pileupDataType))
baseInfo = [
None,
] * baseInfoBranch.nInfo
# Add E_T, eta and luminosity information
npBaseInfo = [
npCurrent.zeros(shape=npCurrent.shape(npat=1, nobs=nobs),
dtype=baseInfoBranch.dtype(idx))
for idx in baseInfoBranch
]
from TuningTools.CreateData import BranchEffCollector, BranchCrossEffCollector
branchEffCollectors = OrderedDict()
branchCrossEffCollectors = OrderedDict()
if ringerOperation < 0:
from operator import itemgetter
benchmarkDict = OrderedDict(
sorted([(key, value) for key, value in
RingerOperation.efficiencyBranches().iteritems()
if key < 0 and not (isinstance(value, (list, tuple)))],
key=itemgetter(0)))
else:
benchmarkDict = OrderedDict()
for key, val in benchmarkDict.iteritems():
branchEffCollectors[key] = list()
branchCrossEffCollectors[key] = list()
# Add efficincy branch:
if ringerOperation < 0:
self.__setBranchAddress(t, val, event)
for etBin in range(nEtBins):
if useBins:
branchEffCollectors[key].append(list())
branchCrossEffCollectors[key].append(list())
for etaBin in range(nEtaBins):
etBinArg = etBin if useBins else -1
etaBinArg = etaBin if useBins else -1
argList = [
RingerOperation.tostring(key), val, etBinArg, etaBinArg
]
branchEffCollectors[key][etBin].append(
BranchEffCollector(*argList))
if crossVal:
branchCrossEffCollectors[key][etBin].append(
BranchCrossEffCollector(entries, crossVal,
*argList))
# etBin
# etaBin
# benchmark dict
if self._logger.isEnabledFor(LoggingLevel.DEBUG):
self._debug(
'Retrieved following branch efficiency collectors: %r', [
collector[0].printName
for collector in traverse(branchEffCollectors.values())
])
etaBin = 0
etBin = 0
step = int(entries / 100) if int(entries / 100) > 0 else 1
## Start loop!
self._info("There is available a total of %d entries.", entries)
cPos = 0
### Loop over entries
for entry in progressbar(range(entries),
entries,
step=step,
logger=self._logger,
prefix="Looping over entries "):
self._verbose('Processing eventNumber: %d/%d', entry, entries)
t.GetEntry(entry)
#print self.__getEt(event)
if event.elCand2_et < offEtCut:
self._debug(
"Ignoring entry due to offline E_T cut. E_T = %1.3f < %1.3f MeV",
event.elCand2_et, offEtCut)
continue
# Add et distribution for all events
if ringerOperation > 0:
if event.fcCand2_et < l2EtCut:
self._debug("Ignoring entry due Fast Calo E_T cut.")
continue
# Add et distribution for all events
# Set discriminator target:
target = Target.Unknown
# Monte Carlo cuts
if reference is Reference.Truth:
if getattr(event, ('elCand%d_isTruthElectronFromZ') %
(self._candIdx)):
target = Target.Signal
elif not getattr(event, ('elCand%d_isTruthElectronFromZ') %
(self._candIdx)):
target = Target.Background
# Offline Likelihood cuts
elif reference is Reference.Off_Likelihood:
if getattr(event,
pidConfigs[RingerOperation.Offline_LH_Tight]):
target = Target.Signal
elif not getattr(
event,
pidConfigs[RingerOperation.Offline_LH_VeryLoose]):
target = Target.Background
# By pass everything (Default)
elif reference is Reference.AcceptAll:
target = Target.Signal if filterType is FilterType.Signal else Target.Background
# Run filter if it is defined
if filterType and \
( (filterType is FilterType.Signal and target != Target.Signal) or \
(filterType is FilterType.Background and target != Target.Background) or \
(target == Target.Unknown) ):
#self._verbose("Ignoring entry due to filter cut.")
continue
## Retrieve base information and rings:
for idx in baseInfoBranch:
lInfo = getattr(event, baseInfoBranch.retrieveBranch(idx))
baseInfo[idx] = lInfo
# Retrieve dependent operation region
if useEtBins:
etBin = self.__retrieveBinIdx(etBins, baseInfo[0])
if useEtaBins:
etaBin = self.__retrieveBinIdx(etaBins, np.fabs(baseInfo[1]))
# Check if bin is within range (when not using bins, this will always be true):
if (etBin < nEtBins and etaBin < nEtaBins):
if useEtBins: npEt[cPos] = etBin
if useEtaBins: npEta[cPos] = etaBin
# Online operation
cPat = 0
caloAvailable = True
if ringerOperation > 0 and self.__get_ringer_onMatch(
event) < 1:
continue
# TODO Treat case where we don't use rings energy
# Check if the rings empty
if self.__get_rings_energy(event, ringerOperation).empty():
self._debug(
'No rings available in this event. Skipping...')
caloAvailable = False
# Retrieve rings:
if extractDet in (Detector.Calorimetry, Detector.CaloAndTrack,
Detector.All):
if caloAvailable:
try:
pass
patterns = stdvector_to_list(
self.__get_rings_energy(
event, ringerOperation))
lPat = len(patterns)
if lPat == ringConfig[etaBin]:
npPatterns[npCurrent.access(
pidx=slice(cPat, ringConfig[etaBin]),
oidx=cPos)] = patterns
else:
oldEtaBin = etaBin
if etaBin > 0 and ringConfig[etaBin -
1] == lPat:
etaBin -= 1
elif etaBin + 1 < len(
ringConfig) and ringConfig[etaBin +
1] == lPat:
etaBin += 1
npPatterns[npCurrent.access(
pidx=slice(cPat, ringConfig[etaBin]),
oidx=cPos)] = patterns
self._warning((
"Recovered event which should be within eta bin (%d: %r) "
"but was found to be within eta bin (%d: %r). "
"Its read eta value was of %f."),
oldEtaBin,
etaBins[oldEtaBin:oldEtaBin + 2],
etaBin,
etaBins[etaBin:etaBin + 2],
np.fabs(getattr(
event, etaBranch)))
except ValueError:
self._logger.error((
"Patterns size (%d) do not match expected "
"value (%d). This event eta value is: %f, and ringConfig is %r."
), lPat, ringConfig[etaBin],
np.fabs(
getattr(event, etaBranch)),
ringConfig)
continue
cPat += ringConfig[etaBin]
else:
# Also display warning when extracting only calorimetry!
self._warning("Rings not available")
continue
if extractDet in (Detector.Tracking, Detector.CaloAndTrack,
Detector.All):
for var in __trackBranches:
npPatterns[npCurrent.access(pidx=cPat,
oidx=cPos)] = getattr(
event, var)
if var == 'elCand2_eProbabilityHT':
from math import log
TRT_PID = npPatterns[npCurrent.access(pidx=cPat,
oidx=cPos)]
epsilon = 1e-99
if TRT_PID >= 1.0: TRT_PID = 1.0 - 1.e-15
elif TRT_PID <= 0.0: TRT_PID = epsilon
tau = 15.0
TRT_PID = -(1 / tau) * log((1.0 / TRT_PID) - 1.0)
npPatterns[npCurrent.access(pidx=cPat,
oidx=cPos)] = TRT_PID
cPat += 1
## Retrieve rates information:
if getRates and ringerOperation < 0:
#event.elCand2_isEMVerLoose2015 = not( event.elCand2_isEMVeryLoose2015 & 34896 )
event.elCand2_isEMLoose2015 = not (
event.elCand2_isEMLoose2015 & 34896)
event.elCand2_isEMMedium2015 = not (
event.elCand2_isEMMedium2015 & 276858960)
event.elCand2_isEMTight2015 = not (
event.elCand2_isEMTight2015 & 281053264)
for branch in branchEffCollectors.itervalues():
if not useBins:
branch.update(event)
else:
branch[etBin][etaBin].update(event)
if crossVal:
for branchCross in branchCrossEffCollectors.itervalues(
):
if not useBins:
branchCross.update(event)
else:
branchCross[etBin][etaBin].update(event)
# end of (getRates)
if not monitoring is None:
self.__fillHistograms(monitoring, filterType, pileupRef,
pidConfigs, event)
# We only increment if this cluster will be computed
cPos += 1
# end of (et/eta bins)
# Limit the number of entries to nClusters if desired and possible:
if not nClusters is None and cPos >= nClusters:
break
# for end
## Treat the rings information
## Remove not filled reserved memory space:
if npPatterns.shape[npCurrent.odim] > cPos:
npPatterns = np.delete(npPatterns,
slice(cPos, None),
axis=npCurrent.odim)
## Segment data over bins regions:
# Also remove not filled reserved memory space:
if useEtBins:
npEt = npCurrent.delete(npEt, slice(cPos, None))
if useEtaBins:
npEta = npCurrent.delete(npEta, slice(cPos, None))
# Treat
standardCaloVariables = False
npObject = self.treatNpInfo(
cPos,
npEt,
npEta,
useEtBins,
useEtaBins,
nEtBins,
nEtaBins,
standardCaloVariables,
ringConfig,
npPatterns,
)
data = [
self.treatNpInfo(cPos, npEt, npEta, useEtBins, useEtaBins, nEtBins,
nEtaBins, standardCaloVariables, ringConfig,
npData) for npData in npBaseInfo
]
npBaseInfo = npCurrent.array(data, dtype=np.object)
if getRates:
if crossVal:
for etBin in range(nEtBins):
for etaBin in range(nEtaBins):
for branchCross in branchCrossEffCollectors.itervalues(
):
if not useBins:
branchCross.finished()
else:
branchCross[etBin][etaBin].finished()
# Print efficiency for each one for the efficiency branches analysed:
for etBin in range(nEtBins) if useBins else range(1):
for etaBin in range(nEtaBins) if useBins else range(1):
for branch in branchEffCollectors.itervalues():
lBranch = branch if not useBins else branch[etBin][
etaBin]
self._info('%s', lBranch)
if crossVal:
for branchCross in branchCrossEffCollectors.itervalues(
):
lBranchCross = branchCross if not useBins else branchCross[
etBin][etaBin]
lBranchCross.dump(self._debug,
printSort=True,
sortFcn=self._verbose)
# for branch
# for eta
# for et
else:
branchEffCollectors = None
branchCrossEffCollectors = None
# end of (getRates)
outputs = []
outputs.extend((npObject, npBaseInfo))
if getRates:
outputs.extend((branchEffCollectors, branchCrossEffCollectors))
return outputs
def trainC_Exp(self):
"""
Train expert feedforward neural network
"""
if coreConf() is TuningToolCores.ExMachina:
self._fatal("Expert Neural Networks not implemented for ExMachina")
elif coreConf() is TuningToolCores.FastNet:
self._fatal("Expert Neural Networks not implemented for FastNet")
elif coreConf() is TuningToolCores.keras:
from copy import deepcopy
# Set batch size:
if self.batchMethod is BatchSizeMethod.MinClassSize:
self.__batchSize(self._bkgSize if self._sgnSize > self._bkgSize
else self._sgnSize)
elif self.batchMethod is BatchSizeMethod.HalfSizeSignalClass:
self.__batchSize(self._sgnSize // 2)
elif self.batchMethod is BatchSizeMethod.OneSample:
self.__batchSize(1)
references = ['SP', 'Pd', 'Pf']
# Holder of the discriminators:
tunedDiscrList = []
tuningInfo = {}
for idx, ref in enumerate(references):
rawDictTempl = {'discriminator': None, 'benchmark': None}
history = self._model[ref].fit(
self._trnData,
self._trnTarget,
epochs=self.trainOptions['nEpochs'],
batch_size=self.batchSize,
callbacks=[self._historyCallback, self._earlyStopping]
#, callbacks = [self._earlyStopping]
,
verbose=0,
validation_data=(self._valData, self._valTarget),
shuffle=self.trainOptions['shuffle'])
# Retrieve raw network
rawDictTempl['discriminator'] = self.__expDiscr_to_dict(
self._model[ref])
rawDictTempl['benchmark'] = self.references[idx]
tunedDiscrList.append(deepcopy(rawDictTempl))
tuningInfo[ref] = DataTrainEvolution(history).toRawObj()
try:
from sklearn.metrics import roc_curve
except ImportError:
# FIXME Can use previous function that we used here as an alternative
raise ImportError(
"sklearn is not available, please install it.")
# Retrieve performance:
opRoc, tstRoc = Roc(), Roc()
for idx, tunedDiscrDict in enumerate(tunedDiscrList):
discr = tunedDiscrDict['discriminator']
if self.doPerf:
self._debug('Retrieving performance for %s networks.' %
(ref))
# propagate inputs:
trnOutput = self._model[ref].predict(self._trnData)
valOutput = self._model[ref].predict(self._valData)
tstOutput = self._model[ref].predict(
self._tstData
) if self._tstData else npCurrent.fp_array([])
try:
allOutput = np.concatenate(
[trnOutput, valOutput, tstOutput])
allTarget = np.concatenate([
self._trnTarget, self._valTarget,
self._tstTarget
])
except ValueError:
allOutput = np.concatenate([trnOutput, valOutput])
allTarget = np.concatenate(
[self._trnTarget, self._valTarget])
# Retrieve Rocs:
opRoc(allOutput, allTarget)
if self._tstData: tstRoc(tstOutput, self._tstTarget)
else: tstRoc(valOutput, self._valTarget)
# Add rocs to output information
# TODO Change this to raw object
tunedDiscrDict['summaryInfo'] = {
'roc_operation': opRoc.toRawObj(),
'roc_test': tstRoc.toRawObj()
}
for ref2 in self.references:
opPoint = opRoc.retrieve(ref2)
tstPoint = tstRoc.retrieve(ref2)
# Print information:
self._info(
'%s NETWORKS Operation (%s): sp = %f, pd = %f, pf = %f, thres = %f',
ref, ref2.name, opPoint.sp_value,
opPoint.pd_value, opPoint.pf_value,
opPoint.thres_value)
self._info(
'%s NETWORKS Test (%s): sp = %f, pd = %f, pf = %f, thres = %f',
ref, ref2.name, tstPoint.sp_value,
tstPoint.pd_value, tstPoint.pf_value,
tstPoint.thres_value)
self._info("Finished trainC_Exp for %s networks." % (ref))
self._debug("Finished trainC_Exp on python side.")
return tunedDiscrList, tuningInfo
def train_c(self):
"""
Train feedforward neural network
"""
from copy import deepcopy
# Holder of the discriminators:
tunedDiscrList = []
tuningInfo = {}
# Set batch size:
if self.batchMethod is BatchSizeMethod.MinClassSize:
self.__batchSize(self._bkgSize if self._sgnSize > self._bkgSize
else self._sgnSize)
elif self.batchMethod is BatchSizeMethod.HalfSizeSignalClass:
self.__batchSize(self._sgnSize // 2)
elif self.batchMethod is BatchSizeMethod.OneSample:
self.__batchSize(1)
rawDictTempl = {'discriminator': None, 'benchmark': None}
if coreConf() is TuningToolCores.keras:
history = self._model.fit(
self._trnData,
self._trnTarget
#, epochs = self.trainOptions['nEpochs']
,
batch_size=self.batchSize
#, callbacks = [self._historyCallback, self._earlyStopping]
,
callbacks=[self._earlyStopping],
verbose=0,
validation_data=(self._valData, self._valTarget),
shuffle=self.trainOptions['shuffle'])
# Retrieve raw network
rawDictTempl['discriminator'] = self.__discr_to_dict(self._model)
rawDictTempl['benchmark'] = self.references[0]
tunedDiscrList.append(deepcopy(rawDictTempl))
tuningInfo = DataTrainEvolution(history).toRawObj()
try:
from sklearn.metrics import roc_curve
except ImportError:
# FIXME Can use previous function that we used here as an alternative
raise ImportError(
"sklearn is not available, please install it.")
elif coreConf() is TuningToolCores.FastNet:
self._debug('executing train_c')
[discriminatorPyWrapperList,
trainDataPyWrapperList] = self._core.train_c()
self._debug('finished train_c')
# Transform model tolist of dict
if self.doMultiStop:
for idx, discr in enumerate(discriminatorPyWrapperList):
rawDictTempl['discriminator'] = self.__discr_to_dict(discr)
rawDictTempl['benchmark'] = self.references[idx]
# FIXME This will need to be improved if set to tune for multiple
# Pd and Pf values.
tunedDiscrList.append(deepcopy(rawDictTempl))
else:
rawDictTempl['discriminator'] = self.__discr_to_dict(
discriminatorPyWrapperList[0])
rawDictTempl['benchmark'] = self.references[0]
if self.useTstEfficiencyAsRef and self.sortIdx is not None:
rawDictTempl['sortIdx'] = self.sortIdx
tunedDiscrList.append(deepcopy(rawDictTempl))
tuningInfo = DataTrainEvolution(trainDataPyWrapperList).toRawObj()
# TODO
# cores
# Retrieve performance:
opRoc, tstRoc = Roc(), Roc()
for idx, tunedDiscrDict in enumerate(tunedDiscrList):
discr = tunedDiscrDict['discriminator']
if self.doPerf:
self._debug('Retrieving performance.')
if coreConf() is TuningToolCores.keras:
# propagate inputs:
trnOutput = self._model.predict(self._trnData)
valOutput = self._model.predict(self._valData)
tstOutput = self._model.predict(
self._tstData
) if self._tstData else npCurrent.fp_array([])
try:
allOutput = np.concatenate(
[trnOutput, valOutput, tstOutput])
allTarget = np.concatenate([
self._trnTarget, self._valTarget, self._tstTarget
])
except ValueError:
allOutput = np.concatenate([trnOutput, valOutput])
allTarget = np.concatenate(
[self._trnTarget, self._valTarget])
# Retrieve Rocs:
opRoc(allOutput, allTarget)
if self._tstData: tstRoc(tstOutput, self._tstTarget)
else: tstRoc(valOutput, self._valTarget)
elif coreConf() is TuningToolCores.FastNet:
perfList = self._core.valid_c(
discriminatorPyWrapperList[idx])
opRoc(perfList[1])
tstRoc(perfList[0])
# Add rocs to output information
# TODO Change this to raw object
tunedDiscrDict['summaryInfo'] = {
'roc_operation': opRoc.toRawObj(),
'roc_test': tstRoc.toRawObj()
}
for ref in self.references:
if coreConf() is TuningToolCores.FastNet:
# FastNet won't loop on this, this is just looping for keras right now
ref = tunedDiscrDict['benchmark']
opPoint = opRoc.retrieve(ref)
tstPoint = tstRoc.retrieve(ref)
# Print information:
self._info(
'Operation (%s): sp = %f, pd = %f, pf = %f, thres = %f',
ref.name, opPoint.sp_value, opPoint.pd_value,
opPoint.pf_value, opPoint.thres_value)
self._info(
'Test (%s): sp = %f, pd = %f, pf = %f, thres = %f',
ref.name, tstPoint.sp_value, tstPoint.pd_value,
tstPoint.pf_value, tstPoint.thres_value)
if coreConf() is TuningToolCores.FastNet:
break
self._debug("Finished train_c on python side.")
return tunedDiscrList, tuningInfo
def __init__(self, **kw):
Logger.__init__(self, kw)
self.references = ReferenceBenchmarkCollection([])
coreframe = coreConf.core_framework()
self.doPerf = retrieve_kw(kw, 'doPerf', True)
self.batchMethod = BatchSizeMethod.retrieve(
retrieve_kw( kw, 'batchMethod', BatchSizeMethod.MinClassSize \
if not 'batchSize' in kw or kw['batchSize'] is NotSet else BatchSizeMethod.Manual )
)
self.batchSize = retrieve_kw(kw, 'batchSize', NotSet)
epochs = retrieve_kw(kw, 'epochs', 10000)
maxFail = retrieve_kw(kw, 'maxFail', 50)
self.useTstEfficiencyAsRef = retrieve_kw(kw, 'useTstEfficiencyAsRef',
False)
self._merged = retrieve_kw(kw, 'merged', False)
self.networks = retrieve_kw(kw, 'networks', NotSet)
self.sortIdx = None
if coreConf() is TuningToolCores.FastNet:
seed = retrieve_kw(kw, 'seed', None)
self._core = coreframe(level=LoggingLevel.toC(self.level),
seed=seed)
self._core.trainFcn = retrieve_kw(kw, 'algorithmName', 'trainrp')
self._core.showEvo = retrieve_kw(kw, 'showEvo', 50)
self._core.multiStop = retrieve_kw(kw, 'doMultiStop', True)
self._core.epochs = epochs
self._core.maxFail = maxFail
# TODO Add properties
elif coreConf() is TuningToolCores.keras:
self._core = coreframe
from keras import callbacks
from keras.optimizers import RMSprop, SGD
from TuningTools.keras_util.callbacks import PerformanceHistory
self.trainOptions = dict()
self.trainOptions['optmin_alg'] = retrieve_kw(
kw, 'optmin_alg', RMSprop(lr=0.001, rho=0.9, epsilon=1e-08))
#self.trainOptions['optmin_alg'] = retrieve_kw( kw, 'optmin_alg', SGD(lr=0.1, decay=1e-6, momentum=0.7) )
self.trainOptions['costFunction'] = retrieve_kw(
kw, 'binary_crossentropy', 'mean_squared_error'
) # 'binary_crossentropy' #'mean_squared_error' #
self.trainOptions['metrics'] = retrieve_kw(kw, 'metrics', [
'accuracy',
])
self.trainOptions['shuffle'] = retrieve_kw(kw, 'shuffle', True)
self._multiStop = retrieve_kw(kw, 'doMultiStop', True)
self.trainOptions['nEpochs'] = epochs
self.trainOptions['nFails'] = maxFail
#self._earlyStopping = callbacks.EarlyStopping( monitor='val_Tuning_L2Calo_SP_sp_value' # val_loss, self.trainOptions['metrics'][0] FIXME This must change
# , patience=self.trainOptions['nFails']
# , verbose=0
# , mode='max')
self._earlyStopping = callbacks.EarlyStopping(
monitor='val_loss' # val_acc
,
patience=self.trainOptions['nFails'],
verbose=0,
mode='auto')
self._historyCallback = PerformanceHistory(
display=retrieve_kw(kw, 'showEvo', 50))
else:
self._fatal("TuningWrapper not implemented for %s", coreConf)
checkForUnusedVars(kw, self._debug)
del kw
# Set default empty values:
if coreConf() is TuningToolCores.keras:
self._emptyData = npCurrent.fp_array([])
elif coreConf() is TuningToolCores.FastNet:
self._emptyData = list()
self._emptyHandler = None
if coreConf() is TuningToolCores.keras:
self._emptyTarget = npCurrent.fp_array([[]]).reshape(
npCurrent.access(pidx=1, oidx=0))
elif coreConf() is TuningToolCores.FastNet:
self._emptyTarget = None
# Set holders:
self._trnData = self._emptyData
self._valData = self._emptyData
self._tstData = self._emptyData
self._trnHandler = self._emptyHandler
self._valHandler = self._emptyHandler
self._tstHandler = self._emptyHandler
self._trnTarget = self._emptyTarget
self._valTarget = self._emptyTarget
self._tstTarget = self._emptyTarget