- crate and train Stack

git-svn-id: http://moon:8086/svn/software/trunk/projects/Rbm@586 b431acfa-c32f-4a4a-93f1-934dc6c82436

Makefile

@@ -12,8 +12,9 @@ CXXFLAGS += -std=c++11
 CXXFLAGS_debug := ${CXXFLAGS} -O0 -g
 CXXFLAGS_release := ${CXXFLAGS} -O2
 	
+DEFINES := -DARMA_OPENMP_THREADS=1
 ${BUILD_DIR}/main.elf: ${BUILD_DIR} ${SRCS}
-	g++ ${CXXFLAGS_${CONFIG}} ${SRCS} -o $@ ${LIBS}
+	g++ ${CXXFLAGS_${CONFIG}} ${DEFINES} ${SRCS} -o $@ ${LIBS}
 
 ${BUILD_DIR}:
 	mkdir -p $@

source/Stack.cpp

@@ -50,12 +50,12 @@ void Stack::addLayer(RbmLayer *pOtherLayer)
 	}
 }
 
-RbmLayer* Stack::getLayer(size_t id) const
+RbmLayer* Stack::getLayer(size_t layerId) const
 {
 	RbmLayer *pLayer = m_pLayers;
 	while(pLayer)
 	{
-		if (pLayer->id() == id)
+		if (pLayer->id() == layerId)
 		{
 			return pLayer;
 		}
@@ -88,4 +88,43 @@ void Stack::save(size_t numTraining)
 	ofs << writer.write(project);
 }
 
+void Stack::saveWeights()
+{
+	RbmLayer *pLayer = m_pLayers;
+	while(pLayer)
+	{
+		pLayer->saveWeights();
+	}
+}
+
+void Stack::train(const arma::mat& batch, size_t miniBatchSize, size_t numEpochs, Rbm::IListener* pListener)
+{
+	RbmLayer *pLayer = m_pLayers;
+	while(pLayer)
+	{
+		train(pLayer->id(), batch, miniBatchSize, numEpochs, pListener);
+		pLayer = pLayer->upper;
+	}
+}
+
+void Stack::train(size_t layerId, const arma::mat& batch, size_t miniBatchSize, size_t numEpochs, Rbm::IListener* pListener)
+{
+	arma::mat thisBatch = batch;
+	RbmLayer *pLayer = m_pLayers;
+	while(pLayer)
+	{
+		if (pLayer->id() == layerId)
+		{
+			break;
+		}
+		thisBatch = pLayer->toHiddenProbs(thisBatch);
+		pLayer = pLayer->upper;
+	}
+	
+	if (pLayer)
+	{
+		std::cout << m_prjname << ": " << " Training of layer " << std::to_string(layerId) << std::endl;
+		pLayer->train(thisBatch, miniBatchSize, numEpochs, pListener);
+	}
+}
 

source/Stack.hpp

@@ -28,8 +28,12 @@ public:
 	virtual ~Stack();
 	
 	void addLayer(RbmLayer *pLayer);
-	RbmLayer* getLayer(size_t id) const;
+	RbmLayer* getLayer(size_t layerId) const;
+
+	void train(size_t layerId, const arma::mat& batch, size_t miniBatchSize, size_t numEpochs, Rbm::IListener* pListener);
+	void train(const arma::mat& batch, size_t miniBatchSize, size_t numEpochs, Rbm::IListener* pListener);
 	void save(size_t numTraining);
+	void saveWeights();
 	
 private:
 	const std::string &m_prjname;
@@ -37,5 +41,6 @@ private:
 	
 };
 
+
 #endif /* STACK_HPP */
 

source/main.cpp

@@ -90,19 +90,28 @@ int main()
 	size_t numTraining = batch.n_rows;
 	size_t numVisibleX = 28;
 	size_t numVisibleY = 28;
-	size_t numHidden = 64;
+	size_t numHidden = 256;
 
 	printf("Loaded %d training samples\n", (int)numTraining);
 
-	for (int i=0; i < 8; i++)
+	int i=0;
+	RbmLayer *lowerLayer = new RbmLayer(project, i, numVisibleX, numVisibleY, numHidden, rbmParams);
+	stack.addLayer(lowerLayer);
+	numHidden >>= 1;
+	
+	i++;
+	for (i; i < 4; i++)
 	{
-		RbmLayer *layer = new RbmLayer(project, i, numVisibleX, numVisibleY, numHidden, rbmParams);
+		RbmLayer *layer = new RbmLayer(project, i, lowerLayer->bh().n_elem, 1, numHidden, rbmParams);
+		lowerLayer = layer;
 		stack.addLayer(layer);
+		numHidden >>= 1;
 	}
 	stack.save(numTraining);
+	stack.train(batch, 1000, 100, &statusDisplay);
+	stack.saveWeights();
+
 	RbmLayer *layer = stack.getLayer(0);
-	layer->train(batch, 1000, 100, &statusDisplay);
-	layer->saveWeights();
 	arma::mat v = arma::randu(numTraining, numVisibleX*numVisibleY);
 	arma::mat h = layer->toHiddenProbs(v);
 	arma::mat r = layer->toVisibleProbs(h);