Using TMVA

Ostap hosts couple of classes, that simplifies the training and using of TMVA.

Training TMVA

tSignal  = ... ## signal     TTree/TChain
tBkg     = ... ## background TTree/TChain
## book TMVA trainer     
from ostap.tools.tmva import Trainer 
trainer = Trainer (
   name    = 'TestTMVA' ,   
   methods = [
   # type                   name   configuration
   ( ROOT.TMVA.Types.kMLP        , 'MLP'        , 'H:!V:EstimatorType=CE:VarTransform=N:NCycles=200:HiddenLayers=N+3:TestRate=5:!UseRegulator' ) ,
   ( ROOT.TMVA.Types.kBDT        , 'BDTG'       , 'H:!V:NTrees=100:MinNodeSize=2.5%:BoostType=Grad:Shrinkage=0.10:UseBaggedBoost:BaggedSampleFraction=0.5:nCuts=20:MaxDepth=2' ) , 
   ( ROOT.TMVA.Types.kCuts       , 'Cuts'       , 'H:!V:FitMethod=MC:EffSel:SampleSize=200000:VarProp=FSmart' ) ,
   ( ROOT.TMVA.Types.kFisher     , 'Fisher'     , 'H:!V:Fisher:VarTransform=None:CreateMVAPdfs:PDFInterpolMVAPdf=Spline2:NbinsMVAPdf=50:NsmoothMVAPdf=10' ),
   ( ROOT.TMVA.Types.kLikelihood , 'Likelihood' , 'H:!V:TransformOutput:PDFInterpol=Spline2:NSmoothSig[0]=20:NSmoothBkg[0]=20:NSmoothBkg[1]=10:NSmooth=1:NAvEvtPerBin=50' )
   ] ,
   variables  = [ 'var1' , 'var2' ,  'var3' ] , ## Variables to be used for training 
   signal     = tSignal                       , ## ``Signal'' sample 
   background = tBkg                          , ## ``Background'' sample  
   verbose    = False )

Optionally one can specify also signal_cuts and background_cuts.

Traing TMVA itself is trivial, one needs to invoke the method train:

weights_files = trainer.train ()

It returs the list/tuple of weight-XML-files, the output of TMVA trainer. Optionally one can retrieve also the list of _C++-class-files, using the proeprty class_files or everything together in a form of tar-file using the property tar_file:

weight_files = trainer.weight_files ## XML weights  
class_files  = trainer.class_files  ## C++ classes 
tar_file     = trainer.tar_file     ## everything together

Using TMVA

To use trained TMVA one exploits TMVA reader:

from ostap.tools.tmva import Reader
reader = Reader( 
   'MyMLP' ,
   variables     = [ ('var1' , lambda s : s.var1 )   ,
                     ('var2' , lambda s : s.var2 )   ,
                     ('var3' , lambda s : s.var3 ) ] ,
  weights_files = tar_file  )

variables     = [ ('var1' , lambda s : s.rapidity     )   , ## use another name 
                  ('var2' , lambda s : s.pt/1000      )   , ## make some rescaling
                  ('var3' , lambda s : atan2(s.y,s.x) ) ] , ## make more complicated calculations

variables     = [ ('var1' , lambda s : s[0] )   , ## use another name 
                  ('var2' , lambda s : s[1] )   , ## make some rescaling
                  ('var3' , lambda s : a[2] ) ] , ## make more complicated calculations

variables     = [ 'var1' , 'var2' , 'var3' ] 

variables     = [ ('var1' , lambda s : getattr( s , 'var1') ) ,
                  ('var2' , lambda s : getattr( s , 'var2') ) ,
                  ('var3' , lambda s : getattr( s , 'var3') ) ]

As weight_files arguments one can use either the list of weights-files from the trainer, or, much easier, use the single 'tar'-file from the trainer. The methods, available from the weight files can be checked as

print reader.methods 
`

And the usage of the reader is rather trivial, e.g. one can explicitly request the responce for certain set of arguments:

v1,  v2,  v3 = .... 
mlp  = reader['MLP']                     ## get  one method 
print 'MLP value is %s'  % mlp ( v1 , v2 ,  v3 )

In practice, one practially always uses it with TTree/TChain/RooAbsData/RooArgSet, in this case one use 1-argument call, assuming then proper accessor functions are supplied:

tree = ... ##  the tree 
mlp  = reader['MLP']                     ## get  one method
for i in tree :                          ## loop over the entries 
    print 'MLP value is %s'  % mlp ( i ) ## get the  value