/// <summary>
/// Creates Outstar architecture with specified number of neurons in
/// output layer
/// </summary>
/// <param name="outputNeuronsCount">
/// number of neurons in output layer </param>
private void createNetwork(int outputNeuronsCount)
{
// set network type
this.NetworkType = NeuralNetworkType.OUTSTAR;
// init neuron settings for this type of network
NeuronProperties neuronProperties = new NeuronProperties();
neuronProperties.setProperty("transferFunction", TransferFunctionType.Step.ToString());
// create input layer
Layer inputLayer = LayerFactory.createLayer(1, neuronProperties);
this.addLayer(inputLayer);
// createLayer output layer
neuronProperties.setProperty("transferFunction", TransferFunctionType.Ramp.ToString());
Layer outputLayer = LayerFactory.createLayer(outputNeuronsCount, neuronProperties);
this.addLayer(outputLayer);
// create full conectivity between input and output layer
ConnectionFactory.fullConnect(inputLayer, outputLayer);
// set input and output cells for this network
NeuralNetworkFactory.DefaultIO = this;
// set outstar learning rule for this network
this.LearningRule = new OutstarLearning();
}
static PoolingLayer()
{
DEFAULT_NEURON_PROP.setProperty("useBias", true);
DEFAULT_NEURON_PROP.setProperty("transferFunction", (util.TransferFunctionType.Tanh.ToString()));
// DEFAULT_NEURON_PROP.setProperty("transferFunction", Ramp.class);
DEFAULT_NEURON_PROP.setProperty("inputFunction", typeof(Max));
}
/// <summary>
/// Creates an instance of Unsuervised Hebian net with specified number
/// of neurons in input layer and output layer, and transfer function
/// </summary>
/// <param name="inputNeuronsNum">
/// number of neurons in input layer </param>
/// <param name="outputNeuronsNum">
/// number of neurons in output layer </param>
/// <param name="transferFunctionType">
/// transfer function type </param>
private void createNetwork(int inputNeuronsNum, int outputNeuronsNum, TransferFunctionType transferFunctionType)
{
// init neuron properties
NeuronProperties neuronProperties = new NeuronProperties();
// neuronProperties.setProperty("bias", new Double(-Math
// .abs(Math.random() - 0.5))); // Hebbian network cann not work
// without bias
neuronProperties.setProperty("transferFunction", transferFunctionType.ToString());
neuronProperties.setProperty("transferFunction.slope", 1);
// set network type code
this.NetworkType = NeuralNetworkType.UNSUPERVISED_HEBBIAN_NET;
// createLayer input layer
Layer inputLayer = LayerFactory.createLayer(inputNeuronsNum, neuronProperties);
this.addLayer(inputLayer);
// createLayer output layer
Layer outputLayer = LayerFactory.createLayer(outputNeuronsNum, neuronProperties);
this.addLayer(outputLayer);
// createLayer full conectivity between input and output layer
ConnectionFactory.fullConnect(inputLayer, outputLayer);
// set input and output cells for this network
NeuralNetworkFactory.DefaultIO = this;
// set appropriate learning rule for this network
this.LearningRule = new UnsupervisedHebbianLearning();
//this.setLearningRule(new OjaLearning(this));
}
/// <summary>
/// Creates MaxNet network architecture
/// </summary>
/// <param name="neuronNum">
/// neuron number in network </param>
/// <param name="neuronProperties">
/// neuron properties </param>
private void createNetwork(int neuronsCount)
{
// set network type
this.NetworkType = NeuralNetworkType.MAXNET;
// createLayer input layer in layer
Layer inputLayer = LayerFactory.createLayer(neuronsCount, new NeuronProperties());
this.addLayer(inputLayer);
// createLayer properties for neurons in output layer
NeuronProperties neuronProperties = new NeuronProperties();
neuronProperties.setProperty("neuronType", typeof(CompetitiveNeuron));
neuronProperties.setProperty("transferFunction", TransferFunctionType.Ramp.ToString());
// createLayer full connectivity in competitive layer
CompetitiveLayer competitiveLayer = new CompetitiveLayer(neuronsCount, neuronProperties);
// add competitive layer to network
this.addLayer(competitiveLayer);
double competitiveWeight = -(1 / (double)neuronsCount);
// createLayer full connectivity within competitive layer
ConnectionFactory.fullConnect(competitiveLayer, competitiveWeight, 1);
// createLayer forward connectivity from input to competitive layer
ConnectionFactory.forwardConnect(inputLayer, competitiveLayer, 1);
// set input and output cells for this network
NeuralNetworkFactory.DefaultIO = this;
}
/// <summary>
/// Creates new Hopfield network with specified neuron number
/// </summary>
/// <param name="neuronsCount">
/// neurons number in Hopfied network </param>
public Hopfield(int neuronsCount)
{
// init neuron settings for hopfield network
NeuronProperties neuronProperties = new NeuronProperties();
neuronProperties.setProperty("neuronType", typeof(InputOutputNeuron));
neuronProperties.setProperty("bias", 0);
neuronProperties.setProperty("transferFunction", TransferFunctionType.Step.ToString());
neuronProperties.setProperty("transferFunction.yHigh", 1);
neuronProperties.setProperty("transferFunction.yLow", 0);
this.createNetwork(neuronsCount, neuronProperties);
}
/// <summary>
/// Creates an instance of Supervised Hebbian Network with specified number
/// of neurons in input layer, output layer and transfer function
/// </summary>
/// <param name="inputNeuronsNum">
/// number of neurons in input layer </param>
/// <param name="outputNeuronsNum">
/// number of neurons in output layer </param>
/// <param name="transferFunctionType">
/// transfer function type </param>
private void createNetwork(int inputNeuronsNum, int outputNeuronsNum, TransferFunctionType transferFunctionType)
{
// init neuron properties
NeuronProperties neuronProperties = new NeuronProperties();
neuronProperties.setProperty("transferFunction", transferFunctionType.ToString());
neuronProperties.setProperty("transferFunction.slope", 1);
neuronProperties.setProperty("transferFunction.yHigh", 1);
neuronProperties.setProperty("transferFunction.xHigh", 1);
neuronProperties.setProperty("transferFunction.yLow", -1);
neuronProperties.setProperty("transferFunction.xLow", -1);
// set network type code
this.NetworkType = NeuralNetworkType.SUPERVISED_HEBBIAN_NET;
// createLayer input layer
Layer inputLayer = LayerFactory.createLayer(inputNeuronsNum, neuronProperties);
this.addLayer(inputLayer);
// createLayer output layer
Layer outputLayer = LayerFactory.createLayer(outputNeuronsNum, neuronProperties);
this.addLayer(outputLayer);
// createLayer full conectivity between input and output layer
ConnectionFactory.fullConnect(inputLayer, outputLayer);
// set input and output cells for this network
NeuralNetworkFactory.DefaultIO = this;
// set appropriate learning rule for this network
this.LearningRule = new SupervisedHebbianLearning();
}
// three layers: input, hidden, output
// as mlp add context layer
// jordan connect output of output layer to input of context layer
// output of context to input of hidden layer
private void createNetwork(int inputNeuronsCount, int hiddenNeuronsCount, int contextNeuronsCount, int outputNeuronsCount)
{
// create input layer
InputLayer inputLayer = new InputLayer(inputNeuronsCount);
inputLayer.addNeuron(new BiasNeuron());
addLayer(inputLayer);
NeuronProperties neuronProperties = new NeuronProperties();
// neuronProperties.setProperty("useBias", true);
neuronProperties.setProperty("transferFunction", TransferFunctionType.Sigmoid.ToString()); // use linear or logitic function! (TR-8604.pdf)
Layer hiddenLayer = new Layer(hiddenNeuronsCount, neuronProperties);
hiddenLayer.addNeuron(new BiasNeuron());
addLayer(hiddenLayer);
ConnectionFactory.fullConnect(inputLayer, hiddenLayer);
Layer contextLayer = new Layer(contextNeuronsCount, neuronProperties);
addLayer(contextLayer); // we might also need bias for context neurons?
Layer outputLayer = new Layer(outputNeuronsCount, neuronProperties);
addLayer(outputLayer);
ConnectionFactory.fullConnect(hiddenLayer, outputLayer);
ConnectionFactory.fullConnect(outputLayer, contextLayer);
ConnectionFactory.fullConnect(contextLayer, hiddenLayer);
// set input and output cells for network
NeuralNetworkFactory.DefaultIO = this;
// set learnng rule
this.LearningRule = new BackPropagation();
}
/// <summary>
/// Creates adaline network architecture with specified number of input neurons
/// </summary>
/// <param name="inputNeuronsCount">
/// number of neurons in input layer </param>
private void createNetwork(int inputNeuronsCount)
{
// set network type code
this.NetworkType = NeuralNetworkType.ADALINE;
// create input layer neuron settings for this network
NeuronProperties inNeuronProperties = new NeuronProperties();
inNeuronProperties.setProperty("transferFunction", TransferFunctionType.Linear.ToString());
// createLayer input layer with specified number of neurons
Layer inputLayer = LayerFactory.createLayer(inputNeuronsCount, inNeuronProperties);
inputLayer.addNeuron(new BiasNeuron()); // add bias neuron (always 1, and it will act as bias input for output neuron)
this.addLayer(inputLayer);
// create output layer neuron settings for this network
NeuronProperties outNeuronProperties = new NeuronProperties();
outNeuronProperties.setProperty("transferFunction", TransferFunctionType.Ramp.ToString());
outNeuronProperties.setProperty("transferFunction.slope", 1);
outNeuronProperties.setProperty("transferFunction.yHigh", 1);
outNeuronProperties.setProperty("transferFunction.xHigh", 1);
outNeuronProperties.setProperty("transferFunction.yLow", -1);
outNeuronProperties.setProperty("transferFunction.xLow", -1);
// createLayer output layer (only one neuron)
Layer outputLayer = LayerFactory.createLayer(1, outNeuronProperties);
this.addLayer(outputLayer);
// createLayer full conectivity between input and output layer
ConnectionFactory.fullConnect(inputLayer, outputLayer);
// set input and output cells for network
NeuralNetworkFactory.DefaultIO = this;
// set LMS learning rule for this network
this.LearningRule = new LMS();
}
public virtual void createDemoNetwork()
{
int productsCount = 20;
int typesCount = 3;
int brandsCount = 3;
int priceCount = 3;
int promoCount = 3;
this.NetworkType = NeuralNetworkType.RECOMMENDER;
//this.getLayers().clear();
// init neuron settings for this type of network
NeuronProperties neuronProperties = new NeuronProperties();
neuronProperties.setProperty("transferFunction", TransferFunctionType.Ramp.ToString());
// for sigmoid and tanh transfer functions
neuronProperties.setProperty("transferFunction.slope", 1);
// create input layer
Layer inputLayer = LayerFactory.createLayer(productsCount, neuronProperties);
this.addLayer(inputLayer);
createProductLabels(inputLayer);
// create product types layer
Layer typeLayer = LayerFactory.createLayer(typesCount, neuronProperties);
createTypeLabels(typeLayer);
this.addLayer(typeLayer);
// create brands layer
Layer brandLayer = LayerFactory.createLayer(brandsCount, neuronProperties);
createBrandLabels(brandLayer);
this.addLayer(brandLayer);
// create price layer
Layer priceLayer = LayerFactory.createLayer(priceCount, neuronProperties);
createPriceLabels(priceLayer);
this.addLayer(priceLayer);
// create price layer
Layer promoLayer = LayerFactory.createLayer(promoCount, neuronProperties);
createPromoLabels(promoLayer);
this.addLayer(promoLayer);
// create output layer
Layer outputLayer = LayerFactory.createLayer(productsCount, neuronProperties);
this.addLayer(outputLayer);
createProductLabels(outputLayer);
createTypeConnections();
createBrandConnections();
createPriceConnections();
createPromoConnections();
// create reccurent self connections in output layer
foreach (Neuron neuron in this.getLayerAt(outputLayerIdx).Neurons)
{
neuron.addInputConnection(neuron, 1);
}
// set input and output cells for this network
NeuralNetworkFactory.DefaultIO = this;
// dont learn the self connections
// moze cak i posle svakog prolaza da se primenjuje hebbianovo pravilo a ne samo nakon kupovine
// napravi vise varijanti
// ako kupuje onda moze da se primenjje winner takes all hebbian learning
this.LearningRule = new UnsupervisedHebbianLearning();
}
static InputMapsLayer()
{
DEFAULT_NEURON_PROP.setProperty("neuronType", typeof(InputNeuron));
DEFAULT_NEURON_PROP.setProperty("transferFunction", util.TransferFunctionType.Linear.ToString());
}