jens
57cdfd7065
- GUI: added gaussian hidden, added mini batch size - Rbm: added mini batch training revised sample functions, reverted to old weight decay git-svn-id: http://moon:8086/svn/software/trunk/projects/RBM@305 b431acfa-c32f-4a4a-93f1-934dc6c82436
567 lines
13 KiB
C++
567 lines
13 KiB
C++
/*
|
|
* To change this license header, choose License Headers in Project Properties.
|
|
* To change this template file, choose Tools | Templates
|
|
* and open the template in the editor.
|
|
*/
|
|
|
|
#include "Rbm.hpp"
|
|
|
|
void mylog(const char* format, ...);
|
|
#define printf mylog
|
|
|
|
#define EPSILON_SIGMA 0.001
|
|
|
|
Rbm::Rbm(Weights &weights, const MatrixXd &batch)
|
|
: m_w(weights)
|
|
, m_batch(batch)
|
|
, m_variableSigma(weights.getNumVisible())
|
|
, m_progress(0)
|
|
{
|
|
setMiniBatchSize(batch.rows());
|
|
Noise_Init(&m_noise, 0x32727155);
|
|
m_variableSigma.fill(m_params.m_constantSigma);
|
|
updateHiddenBatch();
|
|
}
|
|
|
|
Rbm::~Rbm()
|
|
{
|
|
Noise_Free(&m_noise);
|
|
}
|
|
|
|
void Rbm::noiseGaussian(MatrixXd &dst)
|
|
{
|
|
for (size_t i=0; i < dst.rows(); i++)
|
|
{
|
|
for (size_t j=0; j < dst.cols(); j++)
|
|
{
|
|
dst(i, j) = Noise_Gaussian(&m_noise);
|
|
}
|
|
}
|
|
}
|
|
|
|
void Rbm::noiseUniform(MatrixXd &dst)
|
|
{
|
|
for (size_t i=0; i < dst.rows(); i++)
|
|
{
|
|
for (size_t j=0; j < dst.cols(); j++)
|
|
{
|
|
dst(i, j) = Noise_Uniform(&m_noise);
|
|
}
|
|
}
|
|
}
|
|
|
|
void Rbm::sampleGaussian(MatrixXd &dst, MatrixXd const &src, const MatrixXd &sigma)
|
|
{
|
|
MatrixXd n(src.rows(), src.cols());
|
|
|
|
noiseGaussian(n);
|
|
dst = sigma.array()*n.array() + src.array();
|
|
}
|
|
|
|
void Rbm::sampleGaussian(MatrixXd &srcDst, const MatrixXd &sigma)
|
|
{
|
|
MatrixXd n(srcDst.rows(), srcDst.cols());
|
|
|
|
noiseGaussian(n);
|
|
srcDst.array() += sigma.array()*n.array();
|
|
}
|
|
|
|
void Rbm::sampleGaussian(MatrixXd &srcDst, const double &sigma)
|
|
{
|
|
sampleGaussian(srcDst, sigma*MatrixXd::Ones(srcDst.rows(), srcDst.cols()));
|
|
}
|
|
|
|
void Rbm::sample(MatrixXd &srcDst)
|
|
{
|
|
sample(srcDst, srcDst);
|
|
}
|
|
|
|
void Rbm::sample(MatrixXd &dst, MatrixXd const &src)
|
|
{
|
|
MatrixXd n(src.rows(), src.cols());
|
|
|
|
noiseUniform(n);
|
|
|
|
dst = (src.array() > n.array()).cast<double>();
|
|
|
|
}
|
|
|
|
void Rbm::probsLogistic(MatrixXd &src)
|
|
{
|
|
src = (1 + (-src.array()).exp()).array().cwiseInverse();
|
|
}
|
|
|
|
void Rbm::probsLogistic(RowVectorXd &src)
|
|
{
|
|
src = (1 + (-src.array()).exp()).array().cwiseInverse();
|
|
}
|
|
|
|
void Rbm::probsLogistic(MatrixXd &src, const MatrixXd &sigma)
|
|
{
|
|
src.array() /= (sigma.array() + EPSILON_SIGMA);
|
|
probsLogistic(src);
|
|
}
|
|
|
|
void Rbm::probsLogistic(RowVectorXd &src, const RowVectorXd &sigma)
|
|
{
|
|
src.array() /= (sigma.array() + EPSILON_SIGMA);
|
|
probsLogistic(src);
|
|
}
|
|
|
|
void Rbm::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 Rbm::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();
|
|
}
|
|
|
|
RowVectorXd Rbm::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 Rbm::calcMean(MatrixXd const &batch)
|
|
{
|
|
// Remove mean
|
|
RowVectorXd res = batch.colwise().mean();
|
|
|
|
// cout << __PRETTY_FUNCTION__ << ": " << res << endl;
|
|
return res;
|
|
|
|
}
|
|
|
|
RowVectorXd Rbm::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 Rbm::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 Rbm::train(uint32_t numEpochs, double sigmaMin)
|
|
{
|
|
uint32_t i;
|
|
uint32_t epoch;
|
|
uint32_t gibbs;
|
|
|
|
size_t trainingSize = m_batch.rows();
|
|
size_t trainingSizeRemain = trainingSize;
|
|
size_t batchRowIndex = 0;
|
|
|
|
double dProgress = 1.0/(numEpochs*(double)trainingSize/std::min(m_params.m_miniBatchSize, trainingSize));
|
|
|
|
MatrixXd dBiasV_curr(MatrixXd::Zero(1, m_w.getNumVisible()));
|
|
MatrixXd dBiasH_curr(MatrixXd::Zero(1, m_w.getNumHidden()));
|
|
MatrixXd dW_curr(MatrixXd::Zero(m_w.getNumVisible(), m_w.getNumHidden()));
|
|
MatrixXd dBiasV(MatrixXd::Zero(1, m_w.getNumVisible()));
|
|
MatrixXd dBiasH(MatrixXd::Zero(1, m_w.getNumHidden()));
|
|
MatrixXd dW(MatrixXd::Zero(m_w.getNumVisible(), m_w.getNumHidden()));
|
|
|
|
m_progress = 0;
|
|
while (trainingSizeRemain)
|
|
{
|
|
cout << "trainingSizeRemain: " << trainingSizeRemain << endl;
|
|
size_t toSlice = std::min(m_params.m_miniBatchSize, trainingSizeRemain);
|
|
MatrixXd batch = m_batch.block(batchRowIndex, 0, toSlice, m_w.getNumVisible());
|
|
trainingSizeRemain -= toSlice;
|
|
batchRowIndex += toSlice;
|
|
size_t batchSize = batch.rows();
|
|
double mu_w = m_params.m_muWeights/std::min(m_params.m_miniBatchSize, trainingSize);
|
|
double mu_biasV = m_params.m_muWeights/std::min(m_params.m_miniBatchSize, trainingSize);
|
|
double mu_biasH = m_params.m_muWeights/std::min(m_params.m_miniBatchSize, trainingSize);
|
|
|
|
MatrixXd batch_sampled(batchSize, m_w.getNumVisible());
|
|
MatrixXd v_sampled(batchSize, m_w.getNumVisible());
|
|
MatrixXd vis(batchSize, m_w.getNumVisible());
|
|
MatrixXd hid(batchSize, m_w.getNumHidden());
|
|
|
|
if (m_params.m_doNormalizeData)
|
|
{
|
|
RowVectorXd mean = calcMean(batch);
|
|
for (i=0; i < batchSize; i++)
|
|
{
|
|
RowVectorXd x = batch.row(i);
|
|
batch.row(i) = normalizeData(x, mean, m_variableSigma);
|
|
}
|
|
}
|
|
|
|
for (epoch=0; epoch < numEpochs; epoch++)
|
|
{
|
|
onProgressChanged();
|
|
|
|
if (m_params.m_doSampleBatch)
|
|
{
|
|
// When the hidden units are being driven by data, always use stochastic binary states
|
|
sample(batch_sampled, batch);
|
|
|
|
// Create hidden layer base on sampled training data
|
|
hid = batch_sampled * m_w.weights();
|
|
hid += m_w.hiddenBias().replicate(batchSize, 1);
|
|
probsLogistic(hid);
|
|
}
|
|
else
|
|
{
|
|
// Create hidden layer base on training data
|
|
hid = batch * m_w.weights();
|
|
hid += m_w.hiddenBias().replicate(batchSize, 1);
|
|
probsLogistic(hid);
|
|
}
|
|
// Sample hidden
|
|
if (!m_params.m_doRaoBlackwell)
|
|
{
|
|
sample(hid);
|
|
}
|
|
|
|
// Update weights (positive phase)
|
|
dBiasV_curr = batch.colwise().sum();
|
|
dBiasH_curr = hid.colwise().sum();
|
|
dW_curr = batch.transpose() * hid;
|
|
|
|
for (gibbs=0; gibbs < m_params.m_numGibbs; gibbs++)
|
|
{
|
|
if (m_params.m_useHiddenGaussian)
|
|
{
|
|
sampleGaussian(hid, m_params.m_constantSigma);
|
|
}
|
|
else
|
|
{
|
|
sample(hid);
|
|
}
|
|
if (m_params.m_useVisibleGaussian)
|
|
{
|
|
// Create visible reconstruction (a fantasy...) given hid
|
|
vis = hid * m_w.weights().transpose();
|
|
vis += m_w.visibleBias().replicate(batchSize, 1);
|
|
sampleGaussian(v_sampled, vis, m_variableSigma.replicate(batchSize, 1));
|
|
hid = v_sampled * m_w.weights();
|
|
hid += m_w.hiddenBias().replicate(batchSize, 1);
|
|
probsLogistic(hid);
|
|
|
|
}
|
|
else
|
|
{
|
|
// Create visible reconstruction (a fantasy...) given hid
|
|
vis = hid * m_w.weights().transpose();
|
|
vis += m_w.visibleBias().replicate(batchSize, 1);
|
|
probsLogistic(vis, m_variableSigma.replicate(batchSize, 1));
|
|
if (m_params.m_doSampleVisible)
|
|
{
|
|
sample(v_sampled, vis);
|
|
// Create hidden representation given sampled v
|
|
hid = v_sampled * m_w.weights();
|
|
hid += m_w.hiddenBias().replicate(batchSize, 1);
|
|
probsLogistic(hid);
|
|
}
|
|
else
|
|
{
|
|
// Create hidden representation given v
|
|
hid = vis * m_w.weights();
|
|
hid += m_w.hiddenBias().replicate(batchSize, 1);
|
|
probsLogistic(hid);
|
|
}
|
|
}
|
|
}
|
|
|
|
// Update weights (negative phase)
|
|
dBiasV_curr -= vis.colwise().sum();
|
|
dBiasH_curr -= hid.colwise().sum();
|
|
dW_curr -= vis.transpose() * hid;
|
|
|
|
m_w.visibleBias() += mu_biasV*(m_params.m_momentum*dBiasV + (1-m_params.m_momentum)*dBiasV_curr);
|
|
dBiasV = dBiasV_curr;
|
|
|
|
if (m_params.m_doSparse)
|
|
{
|
|
MatrixXd h1 = hid-MatrixXd::Ones(hid.rows(), hid.cols())*m_params.m_sparsity;
|
|
RowVectorXd hm = h1.colwise().mean();
|
|
m_w.hiddenBias() -= m_params.m_muSparsity * hm;
|
|
}
|
|
else
|
|
{
|
|
m_w.hiddenBias() += mu_biasH*(m_params.m_momentum*dBiasH + (1-m_params.m_momentum)*dBiasH_curr);
|
|
}
|
|
dBiasH = dBiasH_curr;
|
|
|
|
m_w.weights() += mu_w*((m_params.m_momentum*dW + (1-m_params.m_momentum)*dW_curr) - m_params.m_weightDecay*m_w.weights());
|
|
dW = dW_curr;
|
|
|
|
if (m_params.m_sigmaDecay > 0)
|
|
{
|
|
if (m_variableSigma[0] > sigmaMin)
|
|
{
|
|
m_variableSigma.array() *= (1-m_params.m_sigmaDecay);
|
|
}
|
|
}
|
|
|
|
m_progress += dProgress;
|
|
|
|
} // Number of epochs
|
|
|
|
MatrixXd diffErr = batch - vis;
|
|
diffErr.array() *= diffErr.array();
|
|
double err = diffErr.colwise().sum().sum();
|
|
cout << "error (per mini batch) = " << err << endl;
|
|
} // number of mini batches
|
|
|
|
updateHiddenBatch();
|
|
|
|
MatrixXd vis = m_h * m_w.weights().transpose();
|
|
vis += m_w.visibleBias().replicate(m_batch.rows(), 1);
|
|
probsLogistic(vis, m_variableSigma.replicate(m_batch.rows(), 1));
|
|
MatrixXd diffErr = m_batch - vis;
|
|
diffErr.array() *= diffErr.array();
|
|
double err = diffErr.colwise().sum().sum();
|
|
cout << "error (total) = " << err << endl;
|
|
|
|
onProgressChanged();
|
|
|
|
}
|
|
|
|
double Rbm::getProgress() const
|
|
{
|
|
return m_progress;
|
|
}
|
|
|
|
double Rbm::getEnergy(const VectorXd& visible, const VectorXd& hidden)
|
|
{
|
|
double energy;
|
|
double sigma = m_variableSigma.array().mean();
|
|
|
|
energy = m_w.visibleBias() * visible;
|
|
energy += m_w.hiddenBias() * hidden;
|
|
energy += visible.transpose() * m_w.weights() * hidden;
|
|
|
|
return -energy/(sigma*sigma);
|
|
}
|
|
|
|
void Rbm::toHidden(RowVectorXd &h, RowVectorXd const &v)
|
|
{
|
|
h = v * m_w.weights();
|
|
h += m_w.hiddenBias();
|
|
probsLogistic(h);
|
|
}
|
|
|
|
void Rbm::toVisible(RowVectorXd &v, RowVectorXd const &h)
|
|
{
|
|
v = h * m_w.weights().transpose();
|
|
v += m_w.visibleBias();
|
|
|
|
if (m_params.m_useVisibleGaussian)
|
|
{
|
|
// probsGaussian(v, m_variableSigma);
|
|
}
|
|
else
|
|
{
|
|
probsLogistic(v, m_variableSigma);
|
|
}
|
|
}
|
|
|
|
void Rbm::setConstantSigma(double value)
|
|
{
|
|
m_params.m_constantSigma = value;
|
|
m_variableSigma.fill(m_params.m_constantSigma);
|
|
onParamsChanged();
|
|
}
|
|
|
|
RowVectorXd& Rbm::getVariableSigma()
|
|
{
|
|
return m_variableSigma;
|
|
}
|
|
|
|
void Rbm::setSigmaDecay(double value)
|
|
{
|
|
m_params.m_sigmaDecay = value;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setWeightDecay(double value)
|
|
{
|
|
m_params.m_weightDecay = value;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setLambda(double value)
|
|
{
|
|
m_params.m_lambda = value;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setSparsity(double value)
|
|
{
|
|
m_params.m_sparsity = value;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setUseVisibleGaussian(bool flag)
|
|
{
|
|
m_params.m_useVisibleGaussian = flag;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setUseHiddenGaussian(bool flag)
|
|
{
|
|
m_params.m_useHiddenGaussian = flag;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setDoRaoBlackwell(bool flag)
|
|
{
|
|
m_params.m_doRaoBlackwell = flag;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setDoSampleVisible(bool flag)
|
|
{
|
|
m_params.m_doSampleVisible = flag;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setDoSampleBatch(bool flag)
|
|
{
|
|
m_params.m_doSampleBatch = flag;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setDoSparse(bool flag)
|
|
{
|
|
m_params.m_doSparse = flag;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setNormalizeData(bool flag)
|
|
{
|
|
m_params.m_doNormalizeData = flag;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setDoLearnVariance(bool flag)
|
|
{
|
|
m_params.m_doLearnVariance = flag;
|
|
if (m_params.m_doLearnVariance && m_batch.rows())
|
|
{
|
|
m_variableSigma = calcSigma(m_batch);
|
|
}
|
|
else
|
|
{
|
|
m_variableSigma.fill(m_params.m_constantSigma);
|
|
}
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setNumGibbs(size_t value)
|
|
{
|
|
m_params.m_numGibbs = value;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setMiniBatchSize(size_t size)
|
|
{
|
|
m_params.m_miniBatchSize = size;
|
|
}
|
|
|
|
void Rbm::setMuWeights(double value)
|
|
{
|
|
m_params.m_muWeights = value;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setMuSparsity(double value)
|
|
{
|
|
m_params.m_muSparsity = value;
|
|
onParamsChanged();
|
|
}
|
|
|
|
void Rbm::setMomentum(double value)
|
|
{
|
|
m_params.m_momentum = value;
|
|
onParamsChanged();
|
|
}
|
|
|
|
MatrixXd const& Rbm::getHiddenBatch()
|
|
{
|
|
return m_h;
|
|
}
|
|
|
|
MatrixXd const& Rbm::getBatch()
|
|
{
|
|
return m_batch;
|
|
}
|
|
|
|
void Rbm::updateHiddenBatch()
|
|
{
|
|
if (m_batch.rows() == 0)
|
|
{
|
|
return;
|
|
}
|
|
m_h.resize(m_batch.rows(), m_w.getNumHidden());
|
|
m_h = m_batch * m_w.weights();
|
|
m_h += m_w.hiddenBias().replicate(m_batch.rows(), 1);
|
|
probsLogistic(m_h);
|
|
}
|
|
|
|
Rbm::Params const& Rbm::params()
|
|
{
|
|
return m_params;
|
|
}
|
|
|