BipolarLayer.h

Go to the documentation of this file.

/*!
 *
 * \brief Implements the bipolar (-1,1) state neuron layer
 *
 * \author Asja Fischer
 * \date 2013
 *
 * \par Copyright 1995-2017 Shark Development Team
 * 
 * <BR><HR>
 * This file is part of Shark.
 * <https://shark-ml.github.io/Shark/>
 * 
 * Shark is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published 
 * by the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 * 
 * Shark is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU Lesser General Public License for more details.
 * 
 * You should have received a copy of the GNU Lesser General Public License
 * along with Shark.  If not, see <http://www.gnu.org/licenses/>.
 *
 */
#ifndef SHARK_UNSUPERVISED_RBM_NEURONLAYERS_BIPOLARLAYER_H
#define SHARK_UNSUPERVISED_RBM_NEURONLAYERS_BIPOLARLAYER_H
 
#include <shark/Core/ISerializable.h>
#include <shark/Core/IParameterizable.h>
#include <shark/LinAlg/Base.h>
#include <shark/Data/BatchInterfaceAdaptStruct.h>
#include <shark/Core/Random.h>
#include <shark/Unsupervised/RBM/StateSpaces/TwoStateSpace.h>
#include <shark/Core/OpenMP.h>
namespace shark{
 
///\brief Layer of bipolar units taking values in {-1,1}. 
 
///A neuron in a BipolarLayer takes values in {-1,1} and the conditional probability to be 1 
///given the states of the neurons in the connected layer is determined by the sigmoid function
///and the input it gets from the connected layer.  
class BipolarLayer : public ISerializable, public IParameterizable<>{
private:
    ///\brief The bias terms associated with the neurons.
    RealVector m_bias; 
public:
    ///the state space of this neuron is binary
    typedef SymmetricBinarySpace StateSpace;
 
    ///\brief The sufficient statistics for the Binary Layer store the probability for a neuron to be on
    typedef RealVector SufficientStatistics;
    ///\brief Sufficient statistics of a batch of data.
    typedef Batch<SufficientStatistics>::type StatisticsBatch;
    
    /// \brief Returns the bias values of the units.
    const RealVector& bias()const{
        return m_bias;
    }
 
    /// \brief Returns the bias values of the units.
    RealVector& bias(){
        return m_bias;
    }
        
    ///\brief Resizes this neuron layer.
    ///
    ///@param newSize number of neurons in the layer
    void resize(std::size_t newSize){
        m_bias.resize(newSize);
    }
    
    ///\brief Returns the number of neurons of this layer.
    std::size_t size()const{
        return m_bias.size();
    }
    
    /// \brief Takes the input of the neuron and calculates the distribution of the neurons
    /// For binary neuronsthis is identical with the conditional probability for the neuron to be on given the state of the connected layer.
    ///
    /// @param input the batch of inputs of the neuron
    /// @param statistics sufficient statistics containing the probabilities of the neurons to be one
    /// @param beta the inverse Temperature of the RBM (typically 1) for the whole batch
    template<class Input, class BetaVector>
    void sufficientStatistics(Input const& input, StatisticsBatch& statistics,BetaVector const& beta)const{ // \todo: auch hier noch mal namen ueberdenken
        SIZE_CHECK(input.size2() == size());
        SIZE_CHECK(statistics.size2() == size());
        SIZE_CHECK(input.size1() == statistics.size1());
        
        for(std::size_t i = 0; i != input.size1(); ++i){
            noalias(row(statistics,i)) = sigmoid(2*(row(input,i)+m_bias)*beta(i));
        }
    }
    
    /// \brief Samples from the distribution using either Gibbs- or flip-the-state sampling. 
    ///
    /// For alpha= 0 gibbs sampling is performed. That is the next state for neuron i is directly taken from the conditional distribution of the i-th neuron. 
    /// In the case of alpha=1, flip-the-state sampling is performed, which takes the last state into account and tries to do deterministically jump 
    /// into states with higher probability. This is counterbalanced by a higher chance to jump back into a lower probability state in later steps. 
    /// For alpha between 0 and 1 a mixture of both is performed.
    ///
    /// @param statistics sufficient statistics containing the probabilities of the neurons to be one
    /// @param state the state vector that shell hold the sampled states
    /// @param alpha factor changing from gibbs to flip-the state sampling. 0<=alpha<=1
    /// @param rng the random number generator used for sampling
    template<class Matrix, class Rng>
    void sample(StatisticsBatch const& statistics, Matrix& state, double alpha, Rng& rng) const{
        SIZE_CHECK(statistics.size2() == size());
        SIZE_CHECK(statistics.size1() == state.size1());
        SIZE_CHECK(statistics.size2() == state.size2());
        
        SHARK_CRITICAL_REGION{
            if(alpha == 0.0){//special case: normal gibbs sampling
                for(std::size_t s = 0; s != state.size1();++s){
                    for(std::size_t i = 0; i != state.size2();++i){
                        state(s,i) = random::coinToss(rng,statistics(s,i));
                        if(state(s,i)==0) state(s,i)=-1.;
                    }
                }
            }
            else{//flip-the state sampling
                for(size_t s = 0; s != state.size1(); ++s){
                    for (size_t i = 0; i != state.size2(); i++) {
                        double prob = statistics(s,i);
                        if (state(s,i) == -1) {
                            if (prob <= 0.5) {
                                prob = (1. - alpha) * prob + alpha * prob / (1. - prob);
                            } else {
                                prob = (1. - alpha) * prob  + alpha;
                            }
                        } else {
                            if (prob >= 0.5) {
                                prob = (1. - alpha) * prob + alpha * (1. - (1. - prob) / prob);
                            } else {
                                prob = (1. - alpha) * prob;
                            }
                        }
                        state(s,i) = random::coinToss(rng, prob);
                        if(state(s,i)==0) state(s,i)=-1.;
                    }
                }
            }
        }
    }
    
    /// \brief Computes the log of the probability of the given states in the conditional distribution
    ///
    /// Currently it is only possible to compute the case with alpha=0
    ///
    /// @param statistics the statistics of the conditional distribution
    /// @param state the state to check
    template<class Matrix>
    RealVector logProbability(StatisticsBatch const& statistics, Matrix const& state) const{
        SIZE_CHECK(statistics.size2() == size());
        SIZE_CHECK(statistics.size1() == state.size1());
        SIZE_CHECK(statistics.size2() == state.size2());
        
        RealVector logProbabilities(state.size1(),1.0);
        for(std::size_t s = 0; s != state.size1();++s){
            for(std::size_t i = 0; i != state.size2();++i){
                logProbabilities(s) += (state(s,i) > 0.0)? std::log(statistics(s,i)) : std::log(1-statistics(s,i)); 
            }
        }
        return logProbabilities;
    }
 
    /// \brief Transforms the current state of the neurons for the multiplication with the weight matrix of the RBM,
    /// i.e. calculates the value of the phi-function used in the interaction term.
    /// In the case of bipolar neurons the phi-function is just the identity.
    ///
    /// @param state the state matrix of the neuron layer
    /// @return the value of the phi-function
    template<class Matrix>
    Matrix const& phi(Matrix const& state)const{
        SIZE_CHECK(state.size2() == size());
        return state;   
    }
 
    
    /// \brief Returns the conditional expectation of the phi-function given the state of the connected layer,
    ///
    /// @param statistics the sufficient statistics of the layer
    RealMatrix expectedPhiValue(StatisticsBatch const& statistics)const{ 
        //calculation of the expectation: 1*P(h_i=1|v)- 1*(1-P(h_i=1|v))= 2*P(h_i=1|v)-1
        return 2*statistics - 1;    
    }
 
    /// \brief Returns the mean of the distribution
    /// 
    /// @param statistics the sufficient statistics of the layer for a whole batch
    RealMatrix mean(StatisticsBatch const& statistics)const{ 
        SIZE_CHECK(statistics.size2() == size());
        //calculation of the expectation: 1*P(h_i=1|v)- 1*(1-P(h_i=1|v))= 2*P(h_i=1|v)-1
        return 2*statistics - 1;    
    }
 
 
    /// \brief Returns the energy term this neuron adds to the energy function.
    ///
    /// @param state the state of the neuron layer 
    /// @param beta the inverse temperature of the i-th state
    /// @return the energy term of the neuron layer
    template<class Matrix, class BetaVector>
    RealVector energyTerm(Matrix const& state, BetaVector const& beta)const{
        SIZE_CHECK(state.size2() == size());
        
        RealVector energies = beta*prod(state,m_bias);
        return energies;
    }
    
 
    ///\brief Sums over all possible values of the terms of the energy function which depend on the this layer and returns the logarithmic result.
    ///
    ///This function is called by Energy when the unnormalized marginal probability of the connected layer is to be computed. 
    ///This function calculates the part which depends on the neurons which are to be marginalized out.
    ///(In the case of the bipolar hidden neuron, this is the term \f$ \sum_h e^{\vec h^T W \vec v+ \vec h^T \vec c} \f$). 
    ///The rest is calculated by the energy function.
    ///
    /// @param inputs the inputs of the neurons they get from the other layer
    /// @param beta the inverse temperature of the RBM
    /// @return the marginal distribution of the connected layer
    template<class Input>
    double logMarginalize(Input const& inputs, double beta) const{
        SIZE_CHECK(inputs.size() == size());
        long double logFactorization = 0;
        for(std::size_t i = 0; i != inputs.size(); ++i){
            long double arg = std::abs((inputs(i)+m_bias(i))*beta);
            logFactorization += softPlus(-2*arg)+arg;
        }
        return logFactorization;
    }
    
    ///\brief Calculates the expectation of the derivatives of the energy term of this neuron layer with respect to it's parameters - the bias weights.
    /// The expectation is taken with respect to the conditional probability distribution of the layer given the state of the connected layer.
    ///
    ///This function takes a batch of samples and weights the results
    ///@param derivative the derivative with respect to the parameters, the result is added on top of it to accumulate derivatives
    ///@param samples the samples from which the informations can be extracted
    ///@param weights The weights for alle samples
    template<class Vector, class SampleBatch, class WeightVector>
    void expectedParameterDerivative(Vector& derivative, SampleBatch const& samples, WeightVector const& weights )const{
        SIZE_CHECK(derivative.size() == size());
        noalias(derivative) += 2*prod(weights,samples.statistics) - sum(weights);
    }
    
    ///\brief Calculates the derivatives of the energy term of this neuron layer with respect to it's parameters - the bias weights. 
    ///
    ///This function takes a batch of samples and calculates a weighted derivative
    ///@param derivative the derivative with respect to the parameters, the result is added on top of it to accumulate derivatives
    ///@param samples the sample from which the informations can be extracted
    ///@param weights the weights for the single sample derivatives
    template<class Vector, class SampleBatch, class WeightVector>
    void parameterDerivative(Vector& derivative, SampleBatch const& samples, WeightVector const& weights)const{
        SIZE_CHECK(derivative.size() == size());
        noalias(derivative) += prod(weights,samples.state);
    }
    
    /// \brief Returns the vector with the parameters associated with the neurons in the layer, i.e. the bias vector.
    RealVector parameterVector()const{
        return m_bias;
    }
 
    /// \brief Sets the parameters associated with the neurons in the layer, i.e. the bias vector.
    void setParameterVector(RealVector const& newParameters){
        m_bias = newParameters;
    }
 
    /// \brief Returns the number of the parameters associated with the neurons in the layer.
    std::size_t numberOfParameters()const{
        return size();
    }
    
    /// \brief Reads the bias vector from an archive.
    ///
    /// @param archive the archive
    void read( InArchive & archive ){
        archive >> m_bias;
    }
 
    /// \brief Writes the bias vector to an archive.
    ///
    /// @param archive the archive
    void write( OutArchive & archive ) const{
        archive << m_bias;
    }
};
 
}
#endif