MVAComputer/src/ProcMLP.cc

0001 // -*- C++ -*-
0002 //
0003 // Package:     MVAComputer
0004 // Class  :     ProcMLP
0005 //
0006
0007 // Implementation:
0008 //     An evaluator for a feed-forward neural net (multi-layer perceptron).
0009 //     Each layer has (n + 1) x m weights for n input neurons, 1 bias
0010 //     and m neurons. Also each layer can select between linear and logistic
0011 //     activation function. The output from the last layer is returned.
0012 //
0013 // Author:      Christophe Saout
0014 // Created:     Sat Apr 24 15:18 CEST 2007
0015 //
0016
0017 #include <cstdlib>
0018 #include <algorithm>
0019 #include <iterator>
0020 #include <vector>
0021 #include <cmath>
0022
0023 #include "FWCore/Utilities/interface/Exception.h"
0024
0025 #include "PhysicsTools/MVAComputer/interface/VarProcessor.h"
0026 #include "PhysicsTools/MVAComputer/interface/Calibration.h"
0027
0028 using namespace PhysicsTools;
0029
0030 namespace {  // anonymous
0031
0032   class ProcMLP : public VarProcessor {
0033   public:
0034     typedef VarProcessor::Registry::Registry<ProcMLP, Calibration::ProcMLP> Registry;
0035
0036     ProcMLP(const char *name, const Calibration::ProcMLP *calib, const MVAComputer *computer);
0037     ~ProcMLP() override {}
0038
0039     void configure(ConfIterator iter, unsigned int n) override;
0040     void eval(ValueIterator iter, unsigned int n) const override;
0041     std::vector<double> deriv(ValueIterator iter, unsigned int n) const override;
0042
0043   private:
0044     struct Layer {
0045       Layer(const Calibration::ProcMLP::Layer &calib);
0046       Layer(const Layer &orig)
0047           : inputs(orig.inputs), neurons(orig.neurons), coeffs(orig.coeffs), sigmoid(orig.sigmoid) {}
0048
0049       unsigned int inputs;
0050       unsigned int neurons;
0051       std::vector<double> coeffs;
0052       bool sigmoid;
0053     };
0054
0055     std::vector<Layer> layers;
0056     unsigned int maxTmp;
0057   };
0058
0059   ProcMLP::Registry registry("ProcMLP");
0060
0061   ProcMLP::Layer::Layer(const Calibration::ProcMLP::Layer &calib)
0062       : inputs(calib.first.front().second.size()), neurons(calib.first.size()), sigmoid(calib.second) {
0063     typedef Calibration::ProcMLP::Neuron Neuron;
0064
0065     coeffs.resize(neurons * (inputs + 1));
0066     std::vector<double>::iterator inserter = coeffs.begin();
0067
0068     for (std::vector<Neuron>::const_iterator iter = calib.first.begin(); iter != calib.first.end(); iter++) {
0069       *inserter++ = iter->first;
0070
0071       if (iter->second.size() != inputs)
0072         throw cms::Exception("ProcMLPInput") << "ProcMLP neuron layer inconsistent." << std::endl;
0073
0074       inserter = std::copy(iter->second.begin(), iter->second.end(), inserter);
0075     }
0076   }
0077
0078   ProcMLP::ProcMLP(const char *name, const Calibration::ProcMLP *calib, const MVAComputer *computer)
0079       : VarProcessor(name, calib, computer), maxTmp(0) {
0080     std::copy(calib->layers.begin(), calib->layers.end(), std::back_inserter(layers));
0081
0082     for (unsigned int i = 0; i < layers.size(); i++) {
0083       maxTmp = std::max<unsigned int>(maxTmp, layers[i].neurons);
0084       if (i > 0 && layers[i - 1].neurons != layers[i].inputs)
0085         throw cms::Exception("ProcMLPInput") << "ProcMLP neuron layers do not connect "
0086                                                 "properly."
0087                                              << std::endl;
0088     }
0089   }
0090
0091   void ProcMLP::configure(ConfIterator iter, unsigned int n) {
0092     if (n != layers.front().inputs)
0093       return;
0094
0095     for (unsigned int i = 0; i < n; i++)
0096       iter++(Variable::FLAG_NONE);
0097
0098     for (unsigned int i = 0; i < layers.back().neurons; i++)
0099       iter << Variable::FLAG_NONE;
0100   }
0101
0102   void ProcMLP::eval(ValueIterator iter, unsigned int n) const {
0103     double *tmp = (double *)alloca(2 * maxTmp * sizeof(double));
0104     bool flip = false;
0105
0106     for (double *pos = tmp; iter; iter++, pos++)
0107       *pos = *iter;
0108
0109     double *output = nullptr;
0110     for (std::vector<Layer>::const_iterator layer = layers.begin(); layer != layers.end(); layer++, flip = !flip) {
0111       const double *input = &tmp[flip ? maxTmp : 0];
0112       output = &tmp[flip ? 0 : maxTmp];
0113       std::vector<double>::const_iterator coeff = layer->coeffs.begin();
0114       for (unsigned int i = 0; i < layer->neurons; i++) {
0115         double sum = *coeff++;
0116         for (unsigned int j = 0; j < layer->inputs; j++)
0117           sum += input[j] * *coeff++;
0118         if (layer->sigmoid)
0119           sum = 1.0 / (std::exp(-sum) + 1.0);
0120         *output++ = sum;
0121       }
0122     }
0123
0124     for (const double *pos = &tmp[flip ? maxTmp : 0]; pos < output; pos++)
0125       iter(*pos);
0126   }
0127
0128   std::vector<double> ProcMLP::deriv(ValueIterator iter, unsigned int n) const {
0129     std::vector<double> prevValues, nextValues;
0130     std::vector<double> prevMatrix, nextMatrix;
0131
0132     while (iter)
0133       nextValues.push_back(*iter++);
0134
0135     unsigned int size = nextValues.size();
0136     nextMatrix.resize(size * size);
0137     for (unsigned int i = 0; i < size; i++)
0138       nextMatrix[i * size + i] = 1.;
0139
0140     for (std::vector<Layer>::const_iterator layer = layers.begin(); layer != layers.end(); layer++) {
0141       prevValues.clear();
0142       std::swap(prevValues, nextValues);
0143       prevMatrix.clear();
0144       std::swap(prevMatrix, nextMatrix);
0145
0146       std::vector<double>::const_iterator coeff = layer->coeffs.begin();
0147       for (unsigned int i = 0; i < layer->neurons; i++) {
0148         double sum = *coeff++;
0149         for (unsigned int j = 0; j < layer->inputs; j++)
0150           sum += prevValues[j] * *coeff++;
0151
0152         double deriv;
0153         if (layer->sigmoid) {
0154           double e = std::exp(-sum);
0155           sum = 1.0 / (e + 1.0);
0156           deriv = 1.0 / (e + 1.0 / e + 2.0);
0157         } else
0158           deriv = 1.0;
0159
0160         nextValues.push_back(sum);
0161
0162         for (unsigned int k = 0; k < size; k++) {
0163           sum = 0.0;
0164           coeff -= layer->inputs;
0165           for (unsigned int j = 0; j < layer->inputs; j++)
0166             sum += prevMatrix[j * size + k] * *coeff++;
0167           nextMatrix.push_back(sum * deriv);
0168         }
0169       }
0170     }
0171
0172     return nextMatrix;
0173   }
0174
0175 }  // anonymous namespace