jens
7fada2a227
- committed local changes git-svn-id: http://moon:8086/svn/software/trunk/projects/RBM@270 b431acfa-c32f-4a4a-93f1-934dc6c82436
541 lines
10 KiB
C++
541 lines
10 KiB
C++
/*
|
|
* Rbm.hpp
|
|
*
|
|
* Created on: 21.09.2014
|
|
* Author: jens
|
|
*/
|
|
|
|
#ifndef RBM_HPP_
|
|
#define RBM_HPP_
|
|
|
|
#include "VisibleLayer.hpp"
|
|
#include "HiddenLayer.hpp"
|
|
#include "Weights.hpp"
|
|
#include <cmath>
|
|
#include <Eigen/Dense>
|
|
|
|
using namespace Eigen;
|
|
|
|
void mylog(const char* format, ...);
|
|
#define printf mylog
|
|
|
|
#define EPSILON_SIGMA 0.05
|
|
|
|
class Rbm;
|
|
|
|
class RbmListener
|
|
{
|
|
public:
|
|
RbmListener() {}
|
|
virtual ~RbmListener()
|
|
{
|
|
}
|
|
|
|
virtual void onEpochTrained(const Rbm &obj) = 0;
|
|
};
|
|
|
|
class Rbm
|
|
{
|
|
public:
|
|
Rbm(Weights &weights, RbmListener *pListener = nullptr)
|
|
: m_w(weights)
|
|
, m_pListener(pListener)
|
|
, m_progress(0)
|
|
, m_sigma(1.0)
|
|
, m_sigmaDecay(1.0)
|
|
, m_weightDecay(0.0)
|
|
, m_lambda(1.0)
|
|
, m_sparsity(0)
|
|
, m_muWeights(0.01)
|
|
, m_muSparsity(0.01)
|
|
, m_momentum(0.5)
|
|
, m_doCancel(false)
|
|
, m_useVisibleGaussian(false)
|
|
, m_doRaoBlackwell(false)
|
|
, m_useProbsForHiddenReconstruction(false)
|
|
, m_doSparse(false)
|
|
, m_doNormalizeData(false)
|
|
, m_doLearnVariance(false)
|
|
, m_numGibbs(1)
|
|
{
|
|
Noise_Init(&m_noise, 0x32727155);
|
|
|
|
VectorXd a(4);
|
|
a << 1, 2, 3, 4;
|
|
VectorXd b(4);
|
|
|
|
b.array() = -a.array().exp();
|
|
|
|
cout << b << endl;
|
|
}
|
|
|
|
~Rbm()
|
|
{
|
|
cancel();
|
|
Noise_Free(&m_noise);
|
|
}
|
|
|
|
void sample(MatrixXd &src)
|
|
{
|
|
uint32_t i;
|
|
|
|
for (i=0; i < src.array().size(); i++)
|
|
{
|
|
src.array()(i) = (double)(src.array()(i) > Noise_Uniform(&m_noise));
|
|
}
|
|
}
|
|
|
|
void probsLogistic(MatrixXd &src)
|
|
{
|
|
src.array() = (-src.array()).exp();
|
|
src.array() += 1;
|
|
src.array() = 1.0/src.array();
|
|
}
|
|
|
|
void probsLogistic(RowVectorXd &src)
|
|
{
|
|
src.array() = (-src.array()).exp();
|
|
src.array() += 1;
|
|
src.array() = 1.0/src.array();
|
|
}
|
|
|
|
void probsLogistic(MatrixXd &src, const MatrixXd &sigma)
|
|
{
|
|
src.array() /= (sigma.array() + EPSILON_SIGMA);
|
|
src.array() = (-src.array()).exp();
|
|
src.array() += 1;
|
|
src.array() = 1.0/src.array();
|
|
}
|
|
|
|
void probsLogistic(RowVectorXd &src, const RowVectorXd &sigma)
|
|
{
|
|
src.array() /= (sigma.array() + EPSILON_SIGMA);
|
|
src.array() = (-src.array()).exp();
|
|
src.array() += 1;
|
|
src.array() = 1.0/src.array();
|
|
}
|
|
|
|
void probsGaussian(MatrixXd &src, const MatrixXd &sigma)
|
|
{
|
|
src.array() = 1 - src.array();
|
|
src.array() *= src.array();
|
|
src.array() *= -0.5;
|
|
|
|
MatrixXd var = sigma;
|
|
var.array() += EPSILON_SIGMA;
|
|
var.array() *= var.array();
|
|
|
|
src.array() /= var.array();
|
|
src.array() = src.array().exp();
|
|
|
|
MatrixXd k = var;
|
|
|
|
k.array() *= 2*3.14159265359;
|
|
k.array() = k.array().sqrt();
|
|
k.array() = 1.0/k.array();
|
|
|
|
src.array() *= k.array();
|
|
}
|
|
|
|
void probsGaussian(RowVectorXd &src, const RowVectorXd &sigma)
|
|
{
|
|
src.array() = 1 - src.array();
|
|
src.array() *= src.array();
|
|
src.array() *= -0.5;
|
|
|
|
RowVectorXd var = sigma;
|
|
var.array() += EPSILON_SIGMA;
|
|
var.array() *= var.array();
|
|
|
|
src.array() /= var.array();
|
|
src.array() = src.array().exp();
|
|
|
|
RowVectorXd k = var;
|
|
|
|
k.array() *= 2*3.14159265359;
|
|
k.array() = k.array().sqrt();
|
|
k.array() = 1.0/k.array();
|
|
|
|
src.array() *= k.array();
|
|
}
|
|
|
|
void sampleGaussian(MatrixXd &src, const MatrixXd &sigma)
|
|
{
|
|
uint32_t i;
|
|
|
|
for (i=0; i < src.array().size(); i++)
|
|
{
|
|
src.array()(i) = sigma(i)*Noise_Gaussian(&m_noise) + src.array()(i);
|
|
}
|
|
}
|
|
|
|
RowVectorXd normalizeData(RowVectorXd const &src, RowVectorXd const &mu, RowVectorXd const &var)
|
|
{
|
|
// Remove mean
|
|
RowVectorXd res = src - mu;
|
|
// res.array() /= var.array() + EPSILON_SIGMA;
|
|
|
|
// cout << __PRETTY_FUNCTION__ << ": " << res << endl;
|
|
return res;
|
|
}
|
|
|
|
RowVectorXd calcMean(MatrixXd const &batch)
|
|
{
|
|
// Remove mean
|
|
RowVectorXd res = batch.colwise().mean();
|
|
|
|
// cout << __PRETTY_FUNCTION__ << ": " << res << endl;
|
|
return res;
|
|
|
|
}
|
|
|
|
RowVectorXd calcSigma(MatrixXd const &batch)
|
|
{
|
|
MatrixXd x = batch.rowwise() - batch.colwise().mean();
|
|
|
|
x.array() *= x.array();
|
|
|
|
RowVectorXd res = x.colwise().mean().array().sqrt();
|
|
|
|
// cout << __PRETTY_FUNCTION__ << ": " << res << endl;
|
|
return res;
|
|
}
|
|
|
|
MatrixXd calcZ(MatrixXd &v, MatrixXd &h)
|
|
{
|
|
|
|
MatrixXd t1(v.rows(), m_w.getNumVisible());
|
|
|
|
t1 = v - m_w.visibleBias().transpose().replicate(v.rows(), 1);
|
|
t1.array() *= t1.array();
|
|
t1.array() *= 0.5;
|
|
|
|
t1 -= (h * m_w.weights().transpose());
|
|
|
|
return t1;
|
|
}
|
|
|
|
void train(const LayerArray &vt, uint32_t numEpochs, uint32_t batchSize, double sigmaMin = 0.05)
|
|
{
|
|
uint32_t t, i;
|
|
uint32_t epoch;
|
|
uint32_t gibbs;
|
|
double dProgress = 1.0/numEpochs;
|
|
double kTrain = 1.0/vt.getSize();
|
|
|
|
batchSize = vt.getSize();
|
|
|
|
MatrixXd v(batchSize, m_w.getNumVisible());
|
|
MatrixXd h(batchSize, m_w.getNumHidden());
|
|
MatrixXd batch(batchSize, m_w.getNumVisible());
|
|
|
|
MatrixXd sumBiasV(1, m_w.getNumVisible());
|
|
MatrixXd sumBiasH(1, m_w.getNumHidden());
|
|
MatrixXd sumWeights(m_w.getNumVisible(), m_w.getNumHidden());
|
|
|
|
MatrixXd deltaVar(MatrixXd::Zero(1, m_w.getNumVisible()));
|
|
MatrixXd deltaBiasV(MatrixXd::Zero(1, m_w.getNumVisible()));
|
|
MatrixXd deltaBiasH(MatrixXd::Zero(1, m_w.getNumHidden()));
|
|
MatrixXd deltaWeights(MatrixXd::Zero(m_w.getNumVisible(), m_w.getNumHidden()));
|
|
|
|
MatrixXd diffErr(batchSize, m_w.getNumVisible());
|
|
|
|
m_progress = 0;
|
|
m_doCancel = false;
|
|
|
|
batch = vt.data();
|
|
|
|
m_w.mean() = calcMean(batch);
|
|
|
|
if (m_doLearnVariance)
|
|
{
|
|
m_w.sigma() = calcSigma(batch);
|
|
}
|
|
|
|
if (m_doNormalizeData)
|
|
{
|
|
for (i=0; i < batchSize; i++)
|
|
{
|
|
RowVectorXd x = batch.row(i);
|
|
batch.row(i) = normalizeData(x, m_w.mean(), m_w.sigma());
|
|
}
|
|
}
|
|
|
|
for (epoch=0; epoch < numEpochs; epoch++)
|
|
{
|
|
double err;
|
|
|
|
v = batch;
|
|
if (m_doCancel)
|
|
{
|
|
m_doCancel = false;
|
|
break;
|
|
}
|
|
|
|
// Create hidden layer base on training data
|
|
h = v * m_w.weights();
|
|
h += m_w.hiddenBias().replicate(batchSize, 1);
|
|
probsLogistic(h);
|
|
|
|
if (!m_doRaoBlackwell)
|
|
{
|
|
sample(h);
|
|
}
|
|
|
|
// Update weights (positive phase)
|
|
sumBiasV = v.colwise().sum();
|
|
if (!m_doSparse)
|
|
{
|
|
sumBiasH = h.colwise().sum();
|
|
}
|
|
sumWeights = v.transpose() * h;
|
|
diffErr = v;
|
|
|
|
for (gibbs=0; gibbs < m_numGibbs; gibbs++)
|
|
{
|
|
sample(h);
|
|
|
|
// Create visible reconstruction (a fantasy...)
|
|
v = h * m_w.weights().transpose();
|
|
v += m_w.visibleBias().replicate(batchSize, 1);
|
|
|
|
if (m_useVisibleGaussian)
|
|
{
|
|
if (!m_useProbsForHiddenReconstruction)
|
|
{
|
|
sampleGaussian(v, m_w.sigma().replicate(batchSize, 1));
|
|
}
|
|
}
|
|
else
|
|
{
|
|
probsLogistic(v, m_w.sigma().replicate(batchSize, 1));
|
|
if (!m_useProbsForHiddenReconstruction)
|
|
{
|
|
sample(v);
|
|
}
|
|
}
|
|
|
|
// Create hidden reconstruction
|
|
h = v * m_w.weights();
|
|
h += m_w.hiddenBias().replicate(batchSize, 1);
|
|
probsLogistic(h);
|
|
}
|
|
|
|
if (!m_doRaoBlackwell)
|
|
{
|
|
sample(h);
|
|
}
|
|
// Update weights (negative phase)
|
|
sumBiasV -= v.colwise().sum();
|
|
if (!m_doSparse)
|
|
{
|
|
sumBiasH -= h.colwise().sum();
|
|
}
|
|
sumWeights -= v.transpose() * h;
|
|
diffErr -= v;
|
|
|
|
deltaWeights = m_momentum*deltaWeights + m_muWeights*(kTrain*sumWeights - m_weightDecay*m_w.weights());
|
|
m_w.weights() += deltaWeights;
|
|
|
|
deltaBiasV = m_momentum*deltaBiasV + m_muWeights*kTrain*sumBiasV;
|
|
m_w.visibleBias() += deltaBiasV;
|
|
|
|
if (m_doSparse)
|
|
{
|
|
h = v * m_w.weights();
|
|
h += m_w.hiddenBias().replicate(batchSize, 1);
|
|
probsLogistic(h);
|
|
|
|
sumBiasH.fill(m_sparsity);
|
|
sumBiasH -= h.colwise().mean();
|
|
|
|
deltaBiasH = m_momentum*deltaBiasH + m_muSparsity*sumBiasH;
|
|
|
|
// cout << "Mean(" << m_sparsity << ") = " << (double)sumBiasH.array().mean() << endl;
|
|
// cout << sumBiasH << endl;
|
|
}
|
|
else
|
|
{
|
|
deltaBiasH = m_momentum*deltaBiasH + m_muWeights*kTrain*sumBiasH;
|
|
}
|
|
m_w.hiddenBias() += deltaBiasH;
|
|
|
|
if (m_w.sigma()[0] > sigmaMin)
|
|
{
|
|
m_w.sigma().array() *= m_sigmaDecay;
|
|
}
|
|
|
|
m_progress += dProgress;
|
|
if (m_pListener)
|
|
{
|
|
m_pListener->onEpochTrained(*this);
|
|
}
|
|
|
|
diffErr.array() *= diffErr.array();
|
|
err = diffErr.colwise().sum().sum();
|
|
|
|
cout << "err =" << endl;
|
|
cout << err << endl;
|
|
|
|
} // Number of epochs
|
|
}
|
|
|
|
double getProgress() const
|
|
{
|
|
return m_progress;
|
|
}
|
|
|
|
double getEnergy(const VectorXd& visible, const VectorXd& hidden)
|
|
{
|
|
double energy;
|
|
|
|
energy = m_w.visibleBias() * visible;
|
|
energy += m_w.hiddenBias() * hidden;
|
|
energy += visible.transpose() * m_w.weights() * hidden;
|
|
|
|
return -energy/(m_sigma*m_sigma);
|
|
}
|
|
|
|
RowVectorXd toHidden(const RowVectorXd& v)
|
|
{
|
|
RowVectorXd h(m_w.getNumHidden());
|
|
RowVectorXd vn(m_w.getNumVisible());
|
|
|
|
h = v * m_w.weights();
|
|
h += m_w.hiddenBias();
|
|
probsLogistic(h);
|
|
|
|
return h;
|
|
}
|
|
|
|
RowVectorXd toVisible(const RowVectorXd& h)
|
|
{
|
|
RowVectorXd v(m_w.getNumVisible());
|
|
v = h * m_w.weights().transpose();
|
|
v += m_w.visibleBias();
|
|
|
|
if (m_useVisibleGaussian)
|
|
{
|
|
// probsGaussian(v, m_w.sigma());
|
|
}
|
|
else
|
|
{
|
|
probsLogistic(v, m_w.sigma());
|
|
}
|
|
return v;
|
|
}
|
|
|
|
void setSigma(double value)
|
|
{
|
|
m_sigma = value;
|
|
}
|
|
|
|
void setSigmaDecay(double value)
|
|
{
|
|
m_sigmaDecay = value;
|
|
}
|
|
|
|
void setWeightDecay(double value)
|
|
{
|
|
m_weightDecay = value;
|
|
}
|
|
|
|
void setLambda(double value)
|
|
{
|
|
m_lambda = value;
|
|
}
|
|
|
|
void setSparsity(double value)
|
|
{
|
|
m_sparsity = value;
|
|
}
|
|
|
|
void setUseVisibleGaussian(bool flag)
|
|
{
|
|
m_useVisibleGaussian = flag;
|
|
}
|
|
|
|
void setDoRaoBlackwell(bool flag)
|
|
{
|
|
m_doRaoBlackwell = flag;
|
|
}
|
|
|
|
void setUseProbsForHiddenReconstruction(bool flag)
|
|
{
|
|
m_useProbsForHiddenReconstruction = flag;
|
|
}
|
|
|
|
void setDoSparse(bool flag)
|
|
{
|
|
m_doSparse = flag;
|
|
}
|
|
|
|
void setNormalizeData(bool flag)
|
|
{
|
|
m_doNormalizeData = flag;
|
|
}
|
|
|
|
void setDoLearnVariance(bool flag)
|
|
{
|
|
m_doLearnVariance = flag;
|
|
if (!flag)
|
|
{
|
|
m_w.sigma().fill(m_sigma);
|
|
}
|
|
}
|
|
|
|
void setNumGibbs(uint32_t value)
|
|
{
|
|
m_numGibbs = value;
|
|
}
|
|
|
|
void setMuWeights(double value)
|
|
{
|
|
m_muWeights = value;
|
|
}
|
|
|
|
void setMuSparsity(double value)
|
|
{
|
|
m_muSparsity = value;
|
|
}
|
|
|
|
void setMomentum(double value)
|
|
{
|
|
m_momentum = value;
|
|
}
|
|
|
|
void cancel()
|
|
{
|
|
m_doCancel = true;
|
|
// while(m_doCancel);
|
|
}
|
|
|
|
private:
|
|
Weights &m_w;
|
|
RbmListener *m_pListener;
|
|
noise_gen_t m_noise;
|
|
double m_progress;
|
|
double m_sigma;
|
|
double m_sigmaDecay;
|
|
double m_weightDecay;
|
|
double m_lambda;
|
|
double m_sparsity;
|
|
double m_muWeights;
|
|
double m_muSparsity;
|
|
double m_momentum;
|
|
bool m_useVisibleGaussian;
|
|
bool m_doRaoBlackwell;
|
|
bool m_useProbsForHiddenReconstruction;
|
|
bool m_doSparse;
|
|
bool m_doNormalizeData;
|
|
bool m_doLearnVariance;
|
|
volatile bool m_doCancel;
|
|
uint32_t m_numGibbs;
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
#endif /* RBM_HPP_ */
|