Examples of org.encog.engine.network.activation.ActivationFunction

Package org.encog.engine.network.activation

Examples of org.encog.engine.network.activation.ActivationFunction

org.encog.engine.network.activation.ActivationFunction
This interface allows various activation functions to be used with the neural network. Activation functions are applied to the output from each layer of a neural network. Activation functions scale the output into the desired range. Methods are provided both to process the activation function, as well as the derivative of the function. Some training algorithms, particularly back propagation, require that it be possible to take the derivative of the activation function. Not all activation functions support derivatives. If you implement an activation function that is not derivable then an exception should be thrown inside of the derivativeFunction method implementation. Non-derivable activation functions are perfectly valid, they simply cannot be used with every training algorithm.


    Class<?> t = transferFunctions.iterator().next();


    double slope = 1;


    ActivationFunction activation = null;
    
    if (inputLayer)
      activation = new ActivationLinear();
    else if (t == Linear.class) {
      slope = ((Linear) transfer).getSlope();

View Full Code Here

  public final File SERIAL_FILENAME = TEMP_DIR.createFile("encogtest.ser");
    
  private NEATNetwork create()
  {
    List<NEATNeuron> neurons = new ArrayList<NEATNeuron>();
    ActivationFunction afSigmoid = new ActivationSigmoid();
    ActivationFunction afStep = new ActivationStep();
    
    // create the neurons
    NEATNeuron input1 = new NEATNeuron(
        NEATNeuronType.Input, 
        1,

View Full Code Here


        flat.setActivationFunctions(new ActivationFunction[flat
            .getLayerCounts().length]);


        for (final String line : section.getLines()) {
          ActivationFunction af = null;
          final List<String> cols = EncogFileSection
              .splitColumns(line);
          final String name = "org.encog.engine.network.activation."
              + cols.get(0);
          try {
            final Class<?> clazz = Class.forName(name);
            af = (ActivationFunction) clazz.newInstance();
          } catch (final ClassNotFoundException e) {
            throw new PersistError(e);
          } catch (final InstantiationException e) {
            throw new PersistError(e);
          } catch (final IllegalAccessException e) {
            throw new PersistError(e);
          }


          for (int i = 0; i < af.getParamNames().length; i++) {
            af.setParam(i,
                CSVFormat.EG_FORMAT.parse(cols.get(i + 1)));
          }


          flat.getActivationFunctions()[index++] = af;
        }

View Full Code Here

   *            True if this is a tanh activation, false for sigmoid.
   */
  public FlatNetwork(final int input, final int hidden1, final int hidden2,
      final int output, final boolean tanh) {


    final ActivationFunction linearAct = new ActivationLinear();
    FlatLayer[] layers;
    final ActivationFunction act = tanh ? new ActivationTANH()
        : new ActivationSigmoid();


    if ((hidden1 == 0) && (hidden2 == 0)) {
      layers = new FlatLayer[2];
      layers[0] = new FlatLayer(linearAct, input,

View Full Code Here


        flat.setActivationFunctions(new ActivationFunction[flat
            .getLayerCounts().length]);


        for (final String line : section.getLines()) {
          ActivationFunction af = null;
          final List<String> cols = EncogFileSection
              .splitColumns(line);
          final String name = "org.encog.engine.network.activation."
              + cols.get(0);
          try {
            final Class<?> clazz = Class.forName(name);
            af = (ActivationFunction) clazz.newInstance();
          } catch (final ClassNotFoundException e) {
            throw new PersistError(e);
          } catch (final InstantiationException e) {
            throw new PersistError(e);
          } catch (final IllegalAccessException e) {
            throw new PersistError(e);
          }


          for (int i = 0; i < af.getParamNames().length; i++) {
            af.setParam(i,
                CSVFormat.EG_FORMAT.parse(cols.get(i + 1)));
          }


          flat.getActivationFunctions()[index++] = af;
        }

View Full Code Here

    } else {
      name = fn.toLowerCase();
      params = new double[0];
    }


    ActivationFunction af = allocateAF(name);
    
    if( af==null ) {
      return null;
    }


    if (af.getParamNames().length != params.length) {
      throw new EncogError(name + " expected "
          + af.getParamNames().length + ", but " + params.length
          + " were provided.");
    }


    for (int i = 0; i < af.getParamNames().length; i++) {
      af.setParam(i, params[i]);
    }


    return af;
  }

View Full Code Here

   */
  public FlatNetwork(final int input, final int hidden1, final int hidden2,
      final int output, final boolean tanh) {
    final double[] params = new double[1];
    FlatLayer[] layers;
    final ActivationFunction act = tanh ? new ActivationTANH()
        : new ActivationSigmoid();
    params[0] = 1; // slope


    if ((hidden1 == 0) && (hidden2 == 0)) {
      layers = new FlatLayer[2];

View Full Code Here

    final int toLayerIndex = this.layerIndex[currentLevel];
    final int fromLayerSize = this.layerCounts[currentLevel + 1];
    final int toLayerSize = this.layerFeedCounts[currentLevel];


    final int index = this.weightIndex[currentLevel];
    final ActivationFunction activation = this.network
        .getActivationFunctions()[currentLevel + 1];


    // handle weights
    int yi = fromLayerIndex;
    for (int y = 0; y < fromLayerSize; y++) {
      final double output = this.layerOutput[yi];
      double sum = 0;
      int xi = toLayerIndex;
      int wi = index + y;
      for (int x = 0; x < toLayerSize; x++) {
        this.gradients[wi] += output * this.layerDelta[xi];
        sum += this.weights[wi] * this.layerDelta[xi];
        wi += fromLayerSize;
        xi++;
      }


      this.layerDelta[yi] = sum
          * activation.derivativeFunction(this.layerOutput[yi]);
      yi++;
    }
  }

View Full Code Here

    final int toLayerIndex = this.layerIndex[currentLevel];
    final int fromLayerSize = this.layerCounts[currentLevel + 1];
    final int toLayerSize = this.layerFeedCounts[currentLevel];


    final int index = this.weightIndex[currentLevel];
    final ActivationFunction activation = this.network
        .getActivationFunctions()[currentLevel + 1];
    final double currentFlatSpot = this.flatSpot[currentLevel + 1];


    // handle weights
    int yi = fromLayerIndex;
    for (int y = 0; y < fromLayerSize; y++) {
      final double output = this.layerOutput[yi];
      double sum = 0;
      int xi = toLayerIndex;
      int wi = index + y;
      for (int x = 0; x < toLayerSize; x++) {
        this.gradients[wi] += output * this.layerDelta[xi];
        sum += this.weights[wi] * this.layerDelta[xi];
        wi += fromLayerSize;
        xi++;
      }


      this.layerDelta[yi] = sum
          * (activation.derivativeFunction(this.layerSums[yi],this.layerOutput[yi])+currentFlatSpot);
      yi++;
    }
  }

View Full Code Here

    // fix flat spot, if needed
    this.flatSpot = new double[this.network.getActivationFunctions().length];


    if (this.shouldFixFlatSpot) {
      for (int i = 0; i < this.network.getActivationFunctions().length; i++) {
        final ActivationFunction af = this.network
            .getActivationFunctions()[i];
        
        if( af instanceof ActivationSigmoid ) {
          this.flatSpot[i] = 0.1;
        } else {

View Full Code Here

0 1 2 3

TOP

Related Classes of org.encog.engine.network.activation.ActivationFunction

org.encog.app.generate.generators.AbstractTemplateGenerator

org.encog.app.generate.generators.js.GenerateEncogJavaScript

org.encog.engine.network.flat.FlatNetwork

org.encog.engine.network.train.gradient.GradientWorkerCPU

org.encog.engine.opencl.kernels.KernelNetworkCalc

org.encog.engine.opencl.kernels.KernelNetworkTrain

org.encog.mathutil.matrices.hessian.ChainRuleWorker

org.encog.mathutil.randomize.NguyenWidrowRandomizer

org.encog.ml.factory.method.FeedforwardFactory

org.encog.ml.factory.method.NEATFactory

All source code are property of their respective owners. Java is a trademark of Sun Microsystems, Inc and owned by ORACLE Inc. Contact coftware#gmail.com.