Examples with ActivationFunction org.encog.engine.network.activation.ActivationFunction used on opensource projects

Example 6 with ActivationFunction

use of org.encog.engine.network.activation.ActivationFunction in project shifu by ShifuML.

the class PersistBasicFloatNetwork method readNetwork.

public BasicFloatNetwork readNetwork(final DataInput in) throws IOException {
    final BasicFloatNetwork result = new BasicFloatNetwork();
    final FlatNetwork flat = new FlatNetwork();
    // read properties
    Map<String, String> properties = new HashMap<String, String>();
    int size = in.readInt();
    for (int i = 0; i < size; i++) {
        properties.put(ml.shifu.shifu.core.dtrain.StringUtils.readString(in), ml.shifu.shifu.core.dtrain.StringUtils.readString(in));
    }
    result.getProperties().putAll(properties);
    // read fields
    flat.setBeginTraining(in.readInt());
    flat.setConnectionLimit(in.readDouble());
    flat.setContextTargetOffset(readIntArray(in));
    flat.setContextTargetSize(readIntArray(in));
    flat.setEndTraining(in.readInt());
    flat.setHasContext(in.readBoolean());
    flat.setInputCount(in.readInt());
    flat.setLayerCounts(readIntArray(in));
    flat.setLayerFeedCounts(readIntArray(in));
    flat.setLayerContextCount(readIntArray(in));
    flat.setLayerIndex(readIntArray(in));
    flat.setLayerOutput(readDoubleArray(in));
    flat.setOutputCount(in.readInt());
    flat.setLayerSums(new double[flat.getLayerOutput().length]);
    flat.setWeightIndex(readIntArray(in));
    flat.setWeights(readDoubleArray(in));
    flat.setBiasActivation(readDoubleArray(in));
    // read activations
    flat.setActivationFunctions(new ActivationFunction[flat.getLayerCounts().length]);
    int acSize = in.readInt();
    for (int i = 0; i < acSize; i++) {
        String name = ml.shifu.shifu.core.dtrain.StringUtils.readString(in);
        if (name.equals("ActivationReLU")) {
            name = ActivationReLU.class.getName();
        } else if (name.equals("ActivationLeakyReLU")) {
            name = ActivationLeakyReLU.class.getName();
        } else if (name.equals("ActivationSwish")) {
            name = ActivationSwish.class.getName();
        } else if (name.equals("ActivationPTANH")) {
            name = ActivationPTANH.class.getName();
        } else {
            name = "org.encog.engine.network.activation." + name;
        }
        ActivationFunction af = null;
        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);
        }
        double[] params = readDoubleArray(in);
        for (int j = 0; j < params.length; j++) {
            af.setParam(j, params[j]);
        }
        flat.getActivationFunctions()[i] = af;
    }
    // read subset
    int subsetSize = in.readInt();
    Set<Integer> featureList = new HashSet<Integer>();
    for (int i = 0; i < subsetSize; i++) {
        featureList.add(in.readInt());
    }
    result.setFeatureSet(featureList);
    result.getStructure().setFlat(flat);
    return result;
}

Example 7 with ActivationFunction

use of org.encog.engine.network.activation.ActivationFunction in project shifu by ShifuML.

the class SubGradient method processLevel.

/**
 * Process one level.
 *
 * @param currentLevel
 *            The current level.
 */
private void processLevel(final int currentLevel) {
    final int fromLayerIndex = this.layerIndex[currentLevel + 1];
    final int toLayerIndex = this.layerIndex[currentLevel];
    final int fromLayerSize = this.layerCounts[currentLevel + 1];
    final int toLayerSize = this.layerFeedCounts[currentLevel];
    double nonDropoutRate = 1.0d;
    if (this.layerDropoutRates.length > currentLevel + 1 && this.layerDropoutRates[currentLevel + 1] != 0) {
        nonDropoutRate = 1.0d - this.layerDropoutRates[currentLevel + 1];
    }
    final int index = this.weightIndex[currentLevel];
    final ActivationFunction activation = this.getNetwork().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 wi = index + y;
        int xi = toLayerIndex;
        for (int x = 0; x < toLayerSize; x++) {
            this.gradients[wi] += output * this.getLayerDelta()[xi];
            sum += this.weights[wi] * this.getLayerDelta()[xi];
            wi += fromLayerSize;
            xi++;
        }
        if (Double.compare(nonDropoutRate, 1.0d) == 0 || this.dropoutNodes == null) {
            this.getLayerDelta()[yi] = sum * (activation.derivativeFunction(this.layerSums[yi], this.layerOutput[yi]) + currentFlatSpot);
        } else {
            if (this.dropoutNodes.contains(yi)) {
                this.getLayerDelta()[yi] = 0d;
            } else {
                this.getLayerDelta()[yi] = (sum / nonDropoutRate) * (activation.derivativeFunction(this.layerSums[yi], this.layerOutput[yi] * nonDropoutRate) + currentFlatSpot);
            }
        }
        yi++;
    }
}