Belle II Software  release-08-01-10
HopfieldNetwork.cc
1 /**************************************************************************
2  * basf2 (Belle II Analysis Software Framework) *
3  * Author: The Belle II Collaboration *
4  * *
5  * See git log for contributors and copyright holders. *
6  * This file is licensed under LGPL-3.0, see LICENSE.md. *
7  **************************************************************************/
8 
9 #include <tracking/trackFindingVXD/trackSetEvaluator/HopfieldNetwork.h>
10 
11 #include <framework/logging/Logger.h>
12 #include <framework/utilities/TRandomWrapper.h>
13 
14 #include <Eigen/Dense>
15 
16 #include <numeric>
17 
18 using namespace Belle2;
19 
21  std::vector<OverlapResolverNodeInfo>& overlapResolverNodeInfos, unsigned short nIterations)
22 {
23  //Start value for neurons if they are compatible.
24  //Each compatible connection activates a node with this value.
25  //As the sum of all the activations shall be less than one, we divide the
26  //activiation by the total number of Nodes.
27  //Incompatible Nodes get later minus one, which counteracts all activation,
28  //if the incompatible Node is active.
29  if (overlapResolverNodeInfos.size() < 2) {
30  B2DEBUG(20, "No reason to doHopfield with less than 2 nodes!");
31  return 0;
32  }
33 
34  const float compatibilityValue = (1.0 - m_omega) / static_cast<float>(overlapResolverNodeInfos.size() - 1);
35 
36  const size_t overlapSize = overlapResolverNodeInfos.size();
37 
38  //Weight matrix; knows compatibility between each possible pair of Nodes
39  Eigen::MatrixXd W(overlapSize, overlapSize);
40  //A): Set all elements to compatible:
41  W.fill(compatibilityValue);
42 
43  //B): Inform weight matrix elements of incompatible neurons:
44  for (const auto& aTC : overlapResolverNodeInfos) {
45  for (unsigned int overlapIndex : aTC.overlaps) {
46  W(aTC.trackIndex, overlapIndex) = -1.0;
47  }
48  }
49 
50 
51  // Neuron values
52  Eigen::VectorXd x(overlapSize);
53  // randomize neuron values for first iteration:
54  for (unsigned int i = 0; i < overlapSize; i++) {
55  x(i) = gRandom->Uniform(1.0); // WARNING: original does Un(0;0.1) not Un(0;1)!
56  }
57 
58  //Store for results from last round:
59  Eigen::VectorXd xOld(overlapSize);
60 
61  //Order of execution for neuron values:
62  std::vector<unsigned short> sequenceVector(overlapSize);
63  //iota fills the vector with 0, 1, 2, ... , (size-1)
64  std::iota(sequenceVector.begin(), sequenceVector.end(), 0);
65 
66  //The following block will be evaluated to empty, if LOG_NO_B2DEBUG is defined:
67  B2DEBUG(29, "sequenceVector with length " << sequenceVector.size());
68  B2DEBUG(29, "Entries are from begin to end:");
69  for (auto&& entry : sequenceVector) {
70  B2DEBUG(29, std::to_string(entry) + ", ");
71  }
72 
73  //Store all values of c for protocolling:
74  std::vector<float> cValues(nIterations);
75  //Store for maximum change of weights between iterations.
76  float c = 1.0;
77 
78  //Iterate until change in weights is small:
79  unsigned iIterations = 0;
80 
81  float T = m_T;
82 
83 
84  while (c > m_cmax) {
85  std::shuffle(sequenceVector.begin(), sequenceVector.end(), TRandomWrapper());
86 
87  xOld = x;
88 
89  for (unsigned int i : sequenceVector) {
90  float aTempVal = W.row(i).dot(x);
91  float act = aTempVal + m_omega * overlapResolverNodeInfos[i].qualityIndicator;
92  x(i) = 0.5 * (1. + tanh(act / T));
93  }
94 
95  T = 0.5 * (T + m_Tmin);
96 
97  //Determine maximum change in weights:
98  c = (x - xOld).cwiseAbs().maxCoeff();
99  B2DEBUG(21, "c value is " << c << " at iteration " << iIterations);
100  cValues.at(iIterations) = c;
101 
102  if (iIterations + 1 == nIterations) {
103  B2INFO("Hopfield reached maximum number of iterations without convergence. cValues are:");
104  for (auto&& entry : cValues) {
105  B2INFO(std::to_string(entry));
106  }
107  break;
108  }
109  iIterations++;
110  }
111 
112  //Copy Node values into the activity state of the OverlapResolverNodeInfo objects:
113  for (unsigned int i = 0; i < overlapSize; i++) {
114  overlapResolverNodeInfos[i].activityState = x(i);
115  }
116 
117  return iIterations;
118 }
float m_omega
tuning parameter of the hopfield network
float m_T
start temperature of annealing
float m_cmax
maximal change of weights between iterations
unsigned short doHopfield(std::vector< OverlapResolverNodeInfo > &overlapResolverNodeInfos, unsigned short nIterations=20)
Performance of the actual algorithm.
float m_Tmin
minimal temperature allowed
Abstract base class for different kinds of events.
Wrap TRandom to be useable as a uniform random number generator with STL algorithms like std::shuffle...