Trains neural networks to be used for jet flavour tagging. More...

#include <NeuralNetTrainer.h>

Inheritance diagram for NeuralNetTrainerProcessor:

Collaboration diagram for NeuralNetTrainerProcessor:

Public Member Functions
virtual Processor *	newProcessor ()

virtual void	init ()

virtual void	processRunHeader (LCRunHeader *pRun)

virtual void	processEvent (LCEvent *pEvent)

virtual void	end ()

Protected Member Functions
void	_displayCollectionNames (lcio::LCEvent *pEvent)

void	_trainNet (nnet::BackPropagationCGAlgorithm &pBackPropCGAlgo, nnet::NeuralNetDataSet &dataSet)

bool	_passesCuts (lcio::LCEvent *pEvent)

Protected Attributes
std::string	_JetCollectionName

std::string	_FlavourTagInputsCollectionName

std::string	_TrueJetFlavourCollectionName

int	_serialiseAsXML

nnet::NeuralNet::SerialisationMode	_outputFormat

std::map< std::string, std::string >	_filename

std::map< std::string, bool >	_trainThisNet

std::map< std::string, nnet::NeuralNetDataSet * >	_dataSet

std::map< std::string, int >	_numSignal

std::map< std::string, int >	_numBackground

std::vector< std::string >	_listOfSelectedNetNames

std::map< std::string, unsigned int >	_IndexOf

int	_nRun

int	_nEvent

int	_nAcceptedEvents

Static Protected Attributes
static const int	C_JET =4

static const int	B_JET =5

Detailed Description

Trains neural networks to be used for jet flavour tagging.

Trains flavour tagging networks using the BackPropagationCGAlgorithm (see the Neural Net Package page) with 500 epochs. The networks are trained on the following data:
If only 1 vertex is found (i.e. only the interaction point)

$tanh\left( \frac{D0Significance1}{100} \right)$

$tanh\left( \frac{D0Significance2}{100} \right)$

$tanh\left( \frac{Z0Significance1}{100} \right)$

$tanh\left( \frac{Z0Significance2}{100} \right)$

$JointProbRPhi$

$JointProbZ$

$tanh\left( \frac{3\times Momentum1}{E} \right)$

$tanh\left( \frac{3\times Momentum2}{E} \right)$

If 2 or more vertices are found

$tanh\left( \frac{DecayLengthSignificance}{6\times E} \right)$

$tanh\left( \frac{DecayLength}{10} \right)$

$tanh\left( \frac{PTCorrectedMass}{5} \right)$

$tanh\left( \frac{RawMomentum}{E} \right)$

$JointProbRPhi$

$JointProbZ$

$tanh\left( \frac{NumTracksInVertices}{10} \right)$

$SecondaryVertexProbability$

Where E is the jet energy, everything else is the data calculated by FlavourTagInputsProcessor.
Note that the processor applies it's own hard coded cuts. These are documented under the full description of NeuralNetTrainerProcessor::_passesCuts().

Input

A collection of ReconstructedParticles that represents the jets in the event (put in by your jet finder, say SatoruJetFinderProcessor).
A LCFloatVec collection that holds the true jet flavours, in the same order as the jets (put in by TrueAngularJetFlavourProcessor).
A LCFloatVec collection that holds the flavour tag inputs (put in by FlavourTagInputsProcessor)
Between 1 and 9 filenames for the generated networks. If an output filename is left blank then that network is not trained, but if none are supplied then a lcio::Exception is thrown.

Output

Trained neural networks to the filenames supplied, in the format requested. The LCIO file is not modified at all.

Parameters

JetCollectionName	Name of the ReconstructedParticle collection that represents jets.
FlavourTagInputsCollection	Name of the LCFloatVec collection that holds the flavour tag inputs.
TrueJetFlavourCollection	Name of the LCIntVec Collection that contains the true jet flavours.
Filename-b_net-1vtx	Output filename for the trained 1 vertex b-tag net.
Filename-c_net-1vtx	Output filename for the trained 1 vertex c-tag net.
Filename-bc_net-1vtx	Output filename for the trained 1 vertex c-tag (with only b background) net.
Filename-b_net-2vtx	Output filename for the trained 2 vertex b-tag net.
Filename-c_net-2vtx	Output filename for the trained 2 vertex c-tag net.
Filename-bc_net-2vtx	Output filename for the trained 2 vertex c-tag (with only b background) net.
Filename-b_net-3plusvtx	Output filename for the trained 3 or more vertices b-tag net.
Filename-c_net-3plusvtx	Output filename for the trained 3 or more vertices c-tag net.
Filename-bc_net-3plusvtx	Output filename for the trained 3 or more vertices c-tag (with only b background) net.

Author: Mark Grimes (mark..nosp@m.grim.nosp@m.es@br.nosp@m.isto.nosp@m.l.ac..nosp@m.uk)

Definition at line 71 of file NeuralNetTrainer.h.

Member Function Documentation

void NeuralNetTrainerProcessor::_displayCollectionNames ( lcio::LCEvent * pEvent )

protected

Displays all of the available collections in the file.

Definition at line 533 of file NeuralNetTrainer.cc.

bool NeuralNetTrainerProcessor::_passesCuts ( lcio::LCEvent * pEvent )

protected

All the code for the cuts should be put in here; returns false if the event fails any of the cuts.

Currently selects jets for which the jet polar angle theta is -0.866<= cos(theta) <=0.866.

Definition at line 377 of file NeuralNetTrainer.cc.

void NeuralNetTrainerProcessor::_trainNet	(	nnet::BackPropagationCGAlgorithm &	pBackPropCGAlgo,
		nnet::NeuralNetDataSet &	dataSet
	)

protected

The training code, split off to make the code a bit more manageable

Definition at line 494 of file NeuralNetTrainer.cc.

Member Data Documentation

std::map<std::string,nnet::NeuralNetDataSet*> NeuralNetTrainerProcessor::_dataSet

protected

Map of the data sets for each of the selected nets

Definition at line 97 of file NeuralNetTrainer.h.

std::map<std::string,std::string> NeuralNetTrainerProcessor::_filename

protected

A map of the output filenames for the trained net. All of these maps use strings of the form "c_net-2vtx", "bc_net-3vtx" etcetera as the key

Definition at line 94 of file NeuralNetTrainer.h.

std::map<std::string,unsigned int> NeuralNetTrainerProcessor::_IndexOf

protected

Holds the positions of the inputs, so that, for example, you can get the position of D0Significance1 in the inputs LCFloatVec just with _IndexOf["D0Significance1"]

Definition at line 108 of file NeuralNetTrainer.h.

std::vector<std::string> NeuralNetTrainerProcessor::_listOfSelectedNetNames

protected

A list of the nets that have been selected for training, in the form of the strings used in the map keys above.

Definition at line 104 of file NeuralNetTrainer.h.

int NeuralNetTrainerProcessor::_nAcceptedEvents

protected

The number of events that have passed the cuts so far.

Definition at line 113 of file NeuralNetTrainer.h.

int NeuralNetTrainerProcessor::_nEvent

protected

The event number.

Definition at line 112 of file NeuralNetTrainer.h.

int NeuralNetTrainerProcessor::_nRun

protected

The run number.

Definition at line 111 of file NeuralNetTrainer.h.

std::map<std::string,int> NeuralNetTrainerProcessor::_numBackground

protected

Dito for the number of backgrounds.

Definition at line 100 of file NeuralNetTrainer.h.

std::map<std::string,int> NeuralNetTrainerProcessor::_numSignal

protected

Map of the number of signal events for each net. Not really used for anything, just printed as info when the net starts training.

Definition at line 98 of file NeuralNetTrainer.h.

std::map<std::string,bool> NeuralNetTrainerProcessor::_trainThisNet

protected

Map containing true or false depending on whether this net has been selected for training (if the user supplies an output filename in the steering file).

Definition at line 95 of file NeuralNetTrainer.h.

const int NeuralNetTrainerProcessor::B_JET =5

staticprotected

Ditto.

Definition at line 117 of file NeuralNetTrainer.h.

const int NeuralNetTrainerProcessor::C_JET =4

staticprotected

Just a useful constant for testing true jet flavour.

Definition at line 116 of file NeuralNetTrainer.h.

The documentation for this class was generated from the following files:

Public Member Functions

Protected Member Functions

Protected Attributes

Static Protected Attributes

Detailed Description

Input

Output

Member Function Documentation

Member Data Documentation