BP神经网络的Java实现（转）

http://fantasticinblur.iteye.com/blog/1465497

课程作业要求实现一个BPNN。这次尝试使用Java实现了一个。现共享之。版权属于大家。关于BPNN的原理，就不赘述了。

下面是BPNN的实现代码。类名为BP。

package ml;

import java.util.Random;

/**
 * BPNN.
 *
 * @author RenaQiu
 *
 */
public class BP {
    /**
     * input vector.
     */
    private final double[] input;
    /**
     * hidden layer.
     */
    private final double[] hidden;
    /**
     * output layer.
     */
    private final double[] output;
    /**
     * target.
     */
    private final double[] target;

    /**
     * delta vector of the hidden layer .
     */
    private final double[] hidDelta;
    /**
     * output layer of the output layer.
     */
    private final double[] optDelta;

    /**
     * learning rate.
     */
    private final double eta;
    /**
     * momentum.
     */
    private final double momentum;

    /**
     * weight matrix from input layer to hidden layer.
     */
    private final double[][] iptHidWeights;
    /**
     * weight matrix from hidden layer to output layer.
     */
    private final double[][] hidOptWeights;

    /**
     * previous weight update.
     */
    private final double[][] iptHidPrevUptWeights;
    /**
     * previous weight update.
     */
    private final double[][] hidOptPrevUptWeights;

    public double optErrSum = 0d;

    public double hidErrSum = 0d;

    private final Random random;

    /**
     * Constructor.
     * <p>
     * <strong>Note:</strong> The capacity of each layer will be the parameter
     * plus 1. The additional unit is used for smoothness.
     * </p>
     *
     * @param inputSize
     * @param hiddenSize
     * @param outputSize
     * @param eta
     * @param momentum
     * @param epoch
     */
    public BP(int inputSize, int hiddenSize, int outputSize, double eta,
            double momentum) {

        input = new double[inputSize + 1];
        hidden = new double[hiddenSize + 1];
        output = new double[outputSize + 1];
        target = new double[outputSize + 1];

        hidDelta = new double[hiddenSize + 1];
        optDelta = new double[outputSize + 1];

        iptHidWeights = new double[inputSize + 1][hiddenSize + 1];
        hidOptWeights = new double[hiddenSize + 1][outputSize + 1];

        random = new Random(19881211);
        randomizeWeights(iptHidWeights);
        randomizeWeights(hidOptWeights);

        iptHidPrevUptWeights = new double[inputSize + 1][hiddenSize + 1];
        hidOptPrevUptWeights = new double[hiddenSize + 1][outputSize + 1];

        this.eta = eta;
        this.momentum = momentum;
    }

    private void randomizeWeights(double[][] matrix) {
        for (int i = 0, len = matrix.length; i != len; i++)
            for (int j = 0, len2 = matrix[i].length; j != len2; j++) {
                double real = random.nextDouble();
                matrix[i][j] = random.nextDouble() > 0.5 ? real : -real;
            }
    }

    /**
     * Constructor with default eta = 0.25 and momentum = 0.3.
     *
     * @param inputSize
     * @param hiddenSize
     * @param outputSize
     * @param epoch
     */
    public BP(int inputSize, int hiddenSize, int outputSize) {
        this(inputSize, hiddenSize, outputSize, 0.25, 0.9);
    }

    /**
     * Entry method. The train data should be a one-dim vector.
     *
     * @param trainData
     * @param target
     */
    public void train(double[] trainData, double[] target) {
        loadInput(trainData);
        loadTarget(target);
        forward();
        calculateDelta();
        adjustWeight();
    }

    /**
     * Test the BPNN.
     *
     * @param inData
     * @return
     */
    public double[] test(double[] inData) {
        if (inData.length != input.length - 1) {
            throw new IllegalArgumentException("Size Do Not Match.");
        }
        System.arraycopy(inData, 0, input, 1, inData.length);
        forward();
        return getNetworkOutput();
    }

    /**
     * Return the output layer.
     *
     * @return
     */
    private double[] getNetworkOutput() {
        int len = output.length;
        double[] temp = new double[len - 1];
        for (int i = 1; i != len; i++)
            temp[i - 1] = output[i];
        return temp;
    }

    /**
     * Load the target data.
     *
     * @param arg
     */
    private void loadTarget(double[] arg) {
        if (arg.length != target.length - 1) {
            throw new IllegalArgumentException("Size Do Not Match.");
        }
        System.arraycopy(arg, 0, target, 1, arg.length);
    }

    /**
     * Load the training data.
     *
     * @param inData
     */
    private void loadInput(double[] inData) {
        if (inData.length != input.length - 1) {
            throw new IllegalArgumentException("Size Do Not Match.");
        }
        System.arraycopy(inData, 0, input, 1, inData.length);
    }

    /**
     * Forward.
     *
     * @param layer0
     * @param layer1
     * @param weight
     */
    private void forward(double[] layer0, double[] layer1, double[][] weight) {
        // threshold unit.
        layer0[0] = 1.0;
        for (int j = 1, len = layer1.length; j != len; ++j) {
            double sum = 0;
            for (int i = 0, len2 = layer0.length; i != len2; ++i)
                sum += weight[i][j] * layer0[i];
            layer1[j] = sigmoid(sum);
        }
    }

    /**
     * Forward.
     */
    private void forward() {
        forward(input, hidden, iptHidWeights);
        forward(hidden, output, hidOptWeights);
    }

    /**
     * Calculate output error.
     */
    private void outputErr() {
        double errSum = 0;
        for (int idx = 1, len = optDelta.length; idx != len; ++idx) {
            double o = output[idx];
            optDelta[idx] = o * (1d - o) * (target[idx] - o);
            errSum += Math.abs(optDelta[idx]);
        }
        optErrSum = errSum;
    }

    /**
     * Calculate hidden errors.
     */
    private void hiddenErr() {
        double errSum = 0;
        for (int j = 1, len = hidDelta.length; j != len; ++j) {
            double o = hidden[j];
            double sum = 0;
            for (int k = 1, len2 = optDelta.length; k != len2; ++k)
                sum += hidOptWeights[j][k] * optDelta[k];
            hidDelta[j] = o * (1d - o) * sum;
            errSum += Math.abs(hidDelta[j]);
        }
        hidErrSum = errSum;
    }

    /**
     * Calculate errors of all layers.
     */
    private void calculateDelta() {
        outputErr();
        hiddenErr();
    }

    /**
     * Adjust the weight matrix.
     *
     * @param delta
     * @param layer
     * @param weight
     * @param prevWeight
     */
    private void adjustWeight(double[] delta, double[] layer,
            double[][] weight, double[][] prevWeight) {

        layer[0] = 1;
        for (int i = 1, len = delta.length; i != len; ++i) {
            for (int j = 0, len2 = layer.length; j != len2; ++j) {
                double newVal = momentum * prevWeight[j][i] + eta * delta[i]
                        * layer[j];
                weight[j][i] += newVal;
                prevWeight[j][i] = newVal;
            }
        }
    }

    /**
     * Adjust all weight matrices.
     */
    private void adjustWeight() {
        adjustWeight(optDelta, hidden, hidOptWeights, hidOptPrevUptWeights);
        adjustWeight(hidDelta, input, iptHidWeights, iptHidPrevUptWeights);
    }

    /**
     * Sigmoid.
     *
     * @param val
     * @return
     */
    private double sigmoid(double val) {
        return 1d / (1d + Math.exp(-val));
    }
}

 为了验证正确性，我写了一个测试用例，目的是对于任意的整数（int型），BPNN在经过训练之后，能够准确地判断出它是奇数还是偶数，正数还是负数。首先对于训练的样本（是随机生成的数字），将它转化为一个32位的向量，向量的每个分量就是其二进制形式对应的位上的0或1。将目标输出视作一个4维的向量，[1,0,0,0]代表正奇数，[0,1,0,0]代表正偶数，[0,0,1,0]代表负奇数，[0,0,0,1]代表负偶数。

训练样本为1000个，学习200次。

package ml;

import java.io.IOException;
import java.util.ArrayList;
import java.util.List;
import java.util.Random;

public class Test {

    /**
     * @param args
     * @throws IOException
     */
    public static void main(String[] args) throws IOException {
        BP bp = new BP(32, 15, 4);

        Random random = new Random();
        List<Integer> list = new ArrayList<Integer>();
        for (int i = 0; i != 1000; i++) {
            int value = random.nextInt();
            list.add(value);
        }

        for (int i = 0; i != 200; i++) {
            for (int value : list) {
                double[] real = new double[4];
                if (value >= 0)
                    if ((value & 1) == 1)
                        real[0] = 1;
                    else
                        real[1] = 1;
                else if ((value & 1) == 1)
                    real[2] = 1;
                else
                    real[3] = 1;
                double[] binary = new double[32];
                int index = 31;
                do {
                    binary[index--] = (value & 1);
                    value >>>= 1;
                } while (value != 0);

                bp.train(binary, real);
            }
        }

        System.out.println("训练完毕，下面请输入一个任意数字，神经网络将自动判断它是正数还是复数，奇数还是偶数。");

        while (true) {
            byte[] input = new byte[10];
            System.in.read(input);
            Integer value = Integer.parseInt(new String(input).trim());
            int rawVal = value;
            double[] binary = new double[32];
            int index = 31;
            do {
                binary[index--] = (value & 1);
                value >>>= 1;
            } while (value != 0);

            double[] result = bp.test(binary);

            double max = -Integer.MIN_VALUE;
            int idx = -1;

            for (int i = 0; i != result.length; i++) {
                if (result[i] > max) {
                    max = result[i];
                    idx = i;
                }
            }

            switch (idx) {
            case 0:
                System.out.format("%d是一个正奇数
", rawVal);
                break;
            case 1:
                System.out.format("%d是一个正偶数
", rawVal);
                break;
            case 2:
                System.out.format("%d是一个负奇数
", rawVal);
                break;
            case 3:
                System.out.format("%d是一个负偶数
", rawVal);
                break;
            }
        }
    }

}

 运行结果截图如下：

 这个测试的例子非常简单。大家可以根据自己的需要去使用BP这个类。