Logo ROOT   6.30.04
Reference Guide
 All Namespaces Files Pages
MethodCFMlpANN.h
Go to the documentation of this file.
1 // @(#)root/tmva $Id$
2 // Author: Andreas Hoecker, Joerg Stelzer, Helge Voss, Kai Voss
3 
4 /**********************************************************************************
5  * Project: TMVA - a Root-integrated toolkit for multivariate data analysis *
6  * Package: TMVA *
7  * Class : MethodCFMlpANN *
8  * Web : http://tmva.sourceforge.net *
9  * *
10  * Description: *
11  * Interface for Clermond-Ferrand artificial neural network. *
12  * The ANN code has been translated from FORTRAN77 (f2c); *
13  * see files: MethodCFMlpANN_f2c_mlpl3.cpp *
14  * MethodCFMlpANN_f2c_datacc.cpp *
15  * *
16  * -------------------------------------------------------------------- *
17  * Reference for the original FORTRAN version: *
18  * Authors : J. Proriol and contributions from ALEPH-Clermont-Fd *
19  * Team members. Contact : gaypas@afal11.cern.ch *
20  * *
21  * Copyright: Laboratoire Physique Corpusculaire *
22  * Universite de Blaise Pascal, IN2P3/CNRS *
23  * -------------------------------------------------------------------- *
24  * *
25  * Usage: options are given through Factory: *
26  * factory->BookMethod( "MethodCFMlpANN", OptionsString ); *
27  * *
28  * where: *
29  * TString OptionsString = "n_training_cycles:n_hidden_layers" *
30  * *
31  * default is: n_training_cycles = 5000, n_layers = 4 *
32  * note that the number of hidden layers in the NN is *
33  * *
34  * n_hidden_layers = n_layers - 2 *
35  * *
36  * since there is one input and one output layer. The number of *
37  * nodes (neurons) is predefined to be *
38  * *
39  * n_nodes[i] = nvars + 1 - i (where i=1..n_layers) *
40  * *
41  * with nvars being the number of variables used in the NN. *
42  * Hence, the default case is: n_neurons(layer 1 (input)) : nvars *
43  * n_neurons(layer 2 (hidden)): nvars-1 *
44  * n_neurons(layer 3 (hidden)): nvars-1 *
45  * n_neurons(layer 4 (out)) : 2 *
46  * *
47  * This artificial neural network usually needs a relatively large *
48  * number of cycles to converge (8000 and more). Overtraining can *
49  * be efficiently tested by comparing the signal and background *
50  * output of the NN for the events that were used for training and *
51  * an independent data sample (with equal properties). If the separation *
52  * performance is significantly better for the training sample, the *
53  * NN interprets statistical effects, and is hence overtrained. In *
54  * this case, the number of cycles should be reduced, or the size *
55  * of the training sample increased. *
56  * *
57  * Authors (alphabetical): *
58  * Andreas Hoecker <Andreas.Hocker@cern.ch> - CERN, Switzerland *
59  * Xavier Prudent <prudent@lapp.in2p3.fr> - LAPP, France *
60  * Helge Voss <Helge.Voss@cern.ch> - MPI-K Heidelberg, Germany *
61  * Kai Voss <Kai.Voss@cern.ch> - U. of Victoria, Canada *
62  * *
63  * Copyright (c) 2005: *
64  * CERN, Switzerland *
65  * U. of Victoria, Canada *
66  * MPI-K Heidelberg, Germany *
67  * LAPP, Annecy, France *
68  * *
69  * Redistribution and use in source and binary forms, with or without *
70  * modification, are permitted according to the terms listed in LICENSE *
71  * (http://tmva.sourceforge.net/LICENSE) *
72  * *
73  **********************************************************************************/
74 
75 #ifndef ROOT_TMVA_MethodCFMlpANN
76 #define ROOT_TMVA_MethodCFMlpANN
77 
78 //////////////////////////////////////////////////////////////////////////
79 // //
80 // MethodCFMlpANN //
81 // //
82 // Interface for Clermond-Ferrand artificial neural network //
83 // //
84 //////////////////////////////////////////////////////////////////////////
85 
86 #include <iosfwd>
87 
88 #include "TMVA/MethodBase.h"
89 #include "TMVA/MethodCFMlpANN_Utils.h"
90 #include "TMatrixF.h"
91 
92 namespace TMVA {
93 
94  class MethodCFMlpANN : public MethodBase, MethodCFMlpANN_Utils {
95 
96  public:
97 
98  MethodCFMlpANN( const TString& jobName,
99  const TString& methodTitle,
100  DataSetInfo& theData,
101  const TString& theOption = "3000:N-1:N-2");
102 
103  MethodCFMlpANN( DataSetInfo& theData,
104  const TString& theWeightFile);
105 
106  virtual ~MethodCFMlpANN( void );
107 
108  virtual Bool_t HasAnalysisType( Types::EAnalysisType type, UInt_t numberClasses, UInt_t /*numberTargets*/ );
109 
110  // training method
111  void Train( void );
112 
113  using MethodBase::ReadWeightsFromStream;
114 
115  // write weights to file
116  void AddWeightsXMLTo( void* parent ) const;
117 
118  // read weights from file
119  void ReadWeightsFromStream( std::istream& istr );
120  void ReadWeightsFromXML( void* wghtnode );
121  // calculate the MVA value
122  Double_t GetMvaValue( Double_t* err = 0, Double_t* errUpper = 0 );
123 
124  // data accessors for external functions
125  Double_t GetData ( Int_t isel, Int_t ivar ) const { return (*fData)(isel, ivar); }
126  Int_t GetClass( Int_t ivar ) const { return (*fClass)[ivar]; }
127 
128 
129  // ranking of input variables
130  const Ranking* CreateRanking() { return 0; }
131 
132  protected:
133 
134  // make ROOT-independent C++ class for classifier response (classifier-specific implementation)
135  void MakeClassSpecific( std::ostream&, const TString& ) const;
136 
137  // header and auxiliary classes
138  void MakeClassSpecificHeader( std::ostream&, const TString& = "" ) const;
139 
140  // get help message text
141  void GetHelpMessage() const;
142 
143  Int_t DataInterface( Double_t*, Double_t*, Int_t*, Int_t*, Int_t*, Int_t*,
144  Double_t*, Int_t*, Int_t* );
145 
146  private:
147 
148  void PrintWeights( std::ostream & o ) const;
149 
150  // the option handling methods
151  void DeclareOptions();
152  void ProcessOptions();
153 
154  // LUTs
155  TMatrixF *fData; // the (data,var) string
156  std::vector<Int_t> *fClass; // the event class (1=signal, 2=background)
157 
158  Int_t fNlayers; // number of layers (including input and output layers)
159  Int_t fNcycles; // number of training cycles
160  Int_t* fNodes; // number of nodes per layer
161 
162  // additional member variables for the independent NN::Evaluation phase
163  Double_t** fYNN; // weights
164  TString fLayerSpec; // the hidden layer specification string
165  Int_t MethodCFMlpANN_nsel;
166 
167  // auxiliary member functions
168  Double_t EvalANN( std::vector<Double_t>&, Bool_t& isOK );
169  void NN_ava ( Double_t* );
170  Double_t NN_fonc( Int_t, Double_t ) const;
171 
172  // default initialisation
173  void Init( void );
174 
175  ClassDef(MethodCFMlpANN,0); // Interface for Clermond-Ferrand artificial neural network
176  };
177 
178 } // namespace TMVA
179 
180 #endif
MethodCFMlpANN_Utils.h