jens/Rbm
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

- use all-at-onec (v+c) method

Browse Source 
git-svn-id: http://moon:8086/svn/software/trunk/projects/Rbm@759 b431acfa-c32f-4a4a-93f1-934dc6c82436
This commit is contained in:
jens
2022-01-08 15:24:07 +00:00
parent 3272899888
commit 9fff2082d6
5 changed files with 48 additions and 127 deletions
Download Patch File Download Diff File Expand all files Collapse all files
source/Rbm.cpp
+19 -100
View File
@@ -18,12 +18,10 @@
Rbm::Rbm(size_t numVisible, size_t numHidden, size_t numContext)
: m_params()
, m_whv(numVisible, numHidden)
, m_whc(numContext, numHidden)
, m_whv(numVisible+numContext, numHidden)
, m_bhv(1, numHidden)
, m_bhc(1, numHidden)
, m_bv(1, numVisible)
, m_bc(1, numContext)
, m_bv(1, numVisible+numContext)
, m_ctx(1, numContext)
{
assert(numVisible > 0);
assert(numHidden > 0);
@@ -37,11 +35,9 @@ Rbm::Rbm(size_t numVisible, size_t numHidden, size_t numContext)
Rbm::Rbm(const Rbm& orig)
: m_params(orig.m_params)
, m_whv(orig.m_whv)
, m_whc(orig.m_whc)
, m_bhv(orig.m_bhv)
, m_bhc(orig.m_bhc)
, m_bv(orig.m_bv)
, m_bc(orig.m_bc)
, m_ctx(orig.m_ctx)
{
}
@@ -53,11 +49,9 @@ Rbm::~Rbm()
void Rbm::weightsInit(double stddev, double mu)
{
uniform(m_whv, stddev, mu);
uniform(m_whc, 0.01*stddev, mu);
uniform(m_bhv, stddev, mu);
uniform(m_bhc, 0.01*stddev, mu);
uniform(m_bv, stddev, mu);
uniform(m_bc, stddev, mu);
uniform(m_ctx, stddev, mu);
}
void Rbm::fromJson(Json::Value rbm)
@@ -74,9 +68,9 @@ Json::Value Rbm::toJson() const
return rbm;
}
void Rbm::gibbs_hv(arma::mat &hv_probs, arma::mat &v_probs)
void Rbm::gibbs(arma::mat &hv_probs, arma::mat &v_probs)
{
for (int gibbs=0; gibbs < m_params.numGibbs; gibbs++)
for (int i=0; i < m_params.numGibbs; i++)
{
// Create visible reconstruction (a fantasy...) given hid
if (m_params.gibbsDoSampleHidden)
@@ -100,33 +94,7 @@ void Rbm::gibbs_hv(arma::mat &hv_probs, arma::mat &v_probs)
}
}
void Rbm::gibbs_hc(arma::mat &hc_probs, arma::mat &c_probs)
{
for (int gibbs=0; gibbs < m_params.numGibbs; gibbs++)
{
// Create visible reconstruction (a fantasy...) given hid
if (m_params.gibbsDoSampleHidden)
{
c_probs = prob(h_to_c(sample(hc_probs)));
}
else
{
c_probs = prob(h_to_c(hc_probs));
}
// Create hidden representation given v
if (m_params.gibbsDoSampleVisible)
{
hc_probs = prob(c_to_h(sample(c_probs)));
}
else
{
hc_probs = prob(c_to_h(c_probs));
}
}
}
void Rbm::weightUpdate_hv(arma::mat const &v_states, arma::mat &dw, arma::mat &dbhv, arma::mat &dbv)
void Rbm::weightUpdate(arma::mat const &v_states, arma::mat &dw, arma::mat &dbhv, arma::mat &dbv)
{
arma::mat v_probs(v_states);
arma::mat h_states = v_to_h(v_states);
@@ -147,7 +115,7 @@ void Rbm::weightUpdate_hv(arma::mat const &v_states, arma::mat &dw, arma::mat &d
dbv = sum(v_states, 0);
dbhv = sum(h_states, 0);
gibbs_hv(h_probs, v_probs);
gibbs(h_probs, v_probs);
// Update weights (negative phase)
dw -= v_probs.t() * h_probs;
@@ -155,35 +123,6 @@ void Rbm::weightUpdate_hv(arma::mat const &v_states, arma::mat &dw, arma::mat &d
dbhv -= sum(h_probs, 0);
}
void Rbm::weightUpdate_hc(arma::mat const &c_states, arma::mat &dw, arma::mat &dbhc, arma::mat &dbc)
{
arma::mat c_probs(c_states);
arma::mat h_states = c_to_h(c_states);
arma::mat h_probs = prob(h_states);
// Sample hidden
if (m_params.doRaoBlackwell)
{
h_states = h_probs;
}
else
{
h_states = sample(h_probs);
}
// Update weights (positive phase)
dw = c_states.t() * h_states;
dbc = sum(c_states, 0);
dbhc = sum(h_states, 0);
gibbs_hc(h_probs, c_probs);
// Update weights (negative phase)
dw -= c_probs.t() * h_probs;
dbc -= sum(c_probs, 0);
dbhc -= sum(h_probs, 0);
}
void Rbm::train(const arma::mat& batch, IListener* pListener)
{
Status status;
@@ -193,17 +132,11 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
int batchRowIndex = 0;
arma::mat grad_bias_v(arma::zeros(1, m_bv.n_cols));
arma::mat grad_bias_c(arma::zeros(1, m_bc.n_cols));
arma::mat grad_bias_hv(arma::zeros(1, m_bhv.n_cols));
arma::mat grad_bias_hc(arma::zeros(1, m_bhc.n_cols));
arma::mat grad_weight_hv(arma::zeros(m_whv.n_rows, m_whv.n_cols));
arma::mat grad_weight_hc(arma::zeros(m_whc.n_rows, m_whc.n_cols));
arma::mat momentum_whv = arma::zeros(m_whv.n_rows, m_whv.n_cols);
arma::mat momentum_whc = arma::zeros(m_whc.n_rows, m_whc.n_cols);
arma::mat momentum_bias_v(arma::zeros(1, m_bv.n_cols));
arma::mat momentum_bias_c(arma::zeros(1, m_bc.n_cols));
arma::mat momentum_bias_hv(arma::zeros(1, m_bhv.n_cols));
arma::mat momentum_bias_hc(arma::zeros(1, m_bhc.n_cols));
arma::mat penalty_weights = arma::zeros(m_whv.n_rows, m_whv.n_cols);
int trainingSizeRemain = batch.n_rows;
@@ -212,7 +145,7 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
while (trainingSizeRemain && !shouldAbort)
{
int miniBatchSizeActual = std::min(m_params.miniBatchSize, trainingSizeRemain);
arma::mat miniBatch = batch.rows(batchRowIndex, batchRowIndex+miniBatchSizeActual-1);
arma::mat miniBatch_v = batch.rows(batchRowIndex, batchRowIndex+miniBatchSizeActual-1);
trainingSizeRemain -= miniBatchSizeActual;
batchRowIndex += miniBatchSizeActual;
int scaler = std::min(m_params.miniBatchSize, (int)batch.n_rows);
@@ -224,8 +157,9 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
arma::mat v_probs(miniBatchSizeActual, m_bv.n_cols);
arma::mat hid_states(miniBatchSizeActual, m_bhv.n_cols);
#endif
arma::mat c_states(arma::zeros(miniBatchSizeActual, m_ctx.n_cols));
arma::mat miniBatch(arma::join_rows(miniBatch_v, c_states));
arma::mat v_states(miniBatch);
arma::mat c_states(arma::zeros(miniBatchSizeActual, m_bc.n_cols));
// Create hidden layer base on training data
if (m_params.doSampleBatch)
@@ -253,7 +187,7 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
grad_bias_v = sum(v_states, 0);
grad_bias_hv = sum(hid_states, 0);
for (int gibbs_hv=0; gibbs_hv < m_params.numGibbs; gibbs_hv++)
for (int i=0; i < m_params.numGibbs; i++)
{
// Create visible reconstruction (a fantasy...) given hid
if (m_params.gibbsDoSampleHidden)
@@ -283,14 +217,13 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
grad_bias_v -= sum(v_probs, 0);
grad_bias_hv -= sum(h_probs, 0);
#else
if (m_bc.n_cols == 0)
if (m_ctx.n_cols == 0)
{
weightUpdate_hv(v_states, grad_weight_hv, grad_bias_hv, grad_bias_v);
weightUpdate(v_states, grad_weight_hv, grad_bias_hv, grad_bias_v);
}
else
{
weightUpdate_hv(v_states, grad_weight_hv, grad_bias_hv, grad_bias_v);
weightUpdate_hc(c_states, grad_weight_hc, grad_bias_hc, grad_bias_c);
weightUpdate(v_states, grad_weight_hv, grad_bias_hv, grad_bias_v);
}
#endif
@@ -300,17 +233,11 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
status.L2 = accu(m_whv % m_whv);
momentum_bias_v = m_params.momentum*momentum_bias_v + grad_bias_v;
momentum_bias_hv = m_params.momentum*momentum_bias_hv + grad_bias_hv;
momentum_bias_c = m_params.momentum*momentum_bias_c + grad_bias_c;
momentum_bias_hc = m_params.momentum*momentum_bias_hc + grad_bias_hc;
momentum_whv = m_params.momentum*momentum_whv + grad_weight_hv - status.L2*penalty_weights;
momentum_whc = m_params.momentum*momentum_whc + grad_weight_hc;
m_bv += learning_rate*momentum_bias_v;
m_bc += learning_rate*momentum_bias_c;
m_bhv += learning_rate*momentum_bias_hv;
m_bhc += learning_rate*momentum_bias_hc;
m_whv += learning_rate*momentum_whv;
m_whc += learning_rate*momentum_whc;
progress += dProgress*miniBatchSizeActual;
status.progress = (int)(progress + 0.5);
@@ -337,6 +264,7 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
} // number of mini batches
#if FIXED_BATCH
arma::mat diffErr = batch - prob(h_to_v(prob(v_to_h(batch))));
arma::mat diffErr_squared = diffErr % diffErr;
status.err_total = accu(diffErr_squared)/diffErr_squared.n_elem;
@@ -345,6 +273,7 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
{
pListener->onProgress(this, status);
}
#endif
}
arma::mat Rbm::prob(const arma::mat &src)
@@ -377,16 +306,6 @@ arma::mat Rbm::h_to_v(const arma::mat &hidden) const
return hidden * m_whv.t() + arma::repmat(m_bv, hidden.n_rows, 1);
}
arma::mat Rbm::c_to_h(const arma::mat &context) const
{
return context * m_whc + arma::repmat(m_bhc, context.n_rows, 1);
}
arma::mat Rbm::h_to_c(const arma::mat &hidden) const
{
return hidden * m_whc.t() + arma::repmat(m_bc, hidden.n_rows, 1);
}
arma::mat Rbm::normalize(const arma::mat& src)
{
double mean = arma::accu(src)/src.n_elem;
@@ -414,7 +333,7 @@ void Rbm::uniform(arma::mat& srcDst, double stdDev, double mu)
#endif
}
const arma::mat& Rbm::w() const
const arma::mat& Rbm::whv() const
{
return m_whv;
}