[RBM]

- moved sigma and mean from WEIGHTS to RBM
- use Gibbs slider also for reconstruction draw


git-svn-id: http://moon:8086/svn/software/trunk/projects/RBM@287 b431acfa-c32f-4a4a-93f1-934dc6c82436

Source/Rbm.hpp

@@ -41,9 +41,9 @@ public:
 	: m_w(weights)
 	, m_v(weights.getNumVisible())
 	, m_h(weights.getNumHidden())
+	, m_sigmas(weights.getNumVisible())
 	, m_pListener(pListener)
 	, m_progress(0)
-	, m_sigma(1.0)
 	, m_sigmaDecay(1.0)
 	, m_weightDecay(0.0)
 	, m_lambda(1.0)
@@ -222,7 +222,7 @@ public:
 		return t1;
 	}
 
-	void train(const LayerArray &vt, uint32_t numEpochs, uint32_t batchSize, double sigmaMin = 0.05)
+	void train(const LayerArray &vt, uint32_t numEpochs, double sigmaMin = 0.05)
 	{
 		uint32_t t, i;
 		uint32_t epoch;
@@ -230,7 +230,7 @@ public:
 		double dProgress = 1.0/numEpochs;
 		double kTrain = 1.0/vt.getSize();
 
-		batchSize = vt.getSize();
+		size_t batchSize = vt.getSize();
 
 		MatrixXd v(batchSize, m_w.getNumVisible());
 		MatrixXd h(batchSize, m_w.getNumHidden());
@@ -252,19 +252,18 @@ public:
 
 		batch = vt.data();
 
-		m_w.mean() = calcMean(batch);
-
 		if (m_doLearnVariance)
 		{
-			m_w.sigma() = calcSigma(batch);
+			m_sigmas = calcSigma(batch);
 		}
 
 		if (m_doNormalizeData)
 		{
+			RowVectorXd mean = calcMean(batch);
 			for (i=0; i < batchSize; i++)
 			{
 				RowVectorXd x = batch.row(i);
-				batch.row(i) = normalizeData(x, m_w.mean(), m_w.sigma());
+				batch.row(i) = normalizeData(x, mean, m_sigmas);
 			}
 		}
 
@@ -310,12 +309,12 @@ public:
 				{
 					if (!m_useProbsForHiddenReconstruction)
 					{
-						sampleGaussian(v, m_w.sigma().replicate(batchSize, 1));
+						sampleGaussian(v, m_sigmas.replicate(batchSize, 1));
 					}
 				}
 				else
 				{
-					probsLogistic(v, m_w.sigma().replicate(batchSize, 1));
+					probsLogistic(v, m_sigmas.replicate(batchSize, 1));
 					if (!m_useProbsForHiddenReconstruction)
 					{
 						sample(v);
@@ -367,9 +366,9 @@ public:
 			}
 			m_w.hiddenBias() += deltaBiasH;
 
-			if (m_w.sigma()[0] > sigmaMin)
+			if (m_sigmas[0] > sigmaMin)
 			{
-				m_w.sigma().array() *= m_sigmaDecay;
+				m_sigmas.array() *= m_sigmaDecay;
 			}
 
 			m_progress += dProgress;
@@ -395,12 +394,13 @@ public:
 	double getEnergy(const VectorXd& visible, const VectorXd& hidden)
 	{
 		double energy;
+		double sigma = m_sigmas.array().mean();
 
 		energy = m_w.visibleBias() * visible;
 		energy += m_w.hiddenBias() * hidden;
 		energy += visible.transpose() * m_w.weights() * hidden;
-
-		return -energy/(m_sigma*m_sigma);
+		
+		return -energy/(sigma*sigma);
 	}
 
 	RowVectorXd const & toHidden(const RowVectorXd& v)
@@ -419,18 +419,23 @@ public:
 
 		if (m_useVisibleGaussian)
 		{
-//			probsGaussian(v, m_w.sigma());
+//			probsGaussian(v, m_sigmas);
 		}
 		else
 		{
-			probsLogistic(m_v, m_w.sigma());
+			probsLogistic(m_v, m_sigmas);
 		}
 		return m_v;
 	}
 
    void setSigma(double value)
    {
-		m_sigma = value;
+		m_sigmas.fill(value);
+   }
+
+   RowVectorXd& getSigma()
+   {
+		return m_sigmas;
    }
 
    void setSigmaDecay(double value)
@@ -481,10 +486,6 @@ public:
    void setDoLearnVariance(bool flag)
    {
 	   m_doLearnVariance = flag;
-	   if (!flag)
-	   {
-			m_w.sigma().fill(m_sigma);
-	   }
    }
 
    void setNumGibbs(uint32_t value)
@@ -517,10 +518,11 @@ private:
 	Weights &m_w;
 	RowVectorXd m_h;
 	RowVectorXd m_v;
+	RowVectorXd m_sigmas;
+
 	RbmListener *m_pListener;
 	noise_gen_t m_noise;
 	double m_progress;
-	double m_sigma;
 	double m_sigmaDecay;
 	double m_weightDecay;
 	double m_lambda;