感谢周志华先生写的《机器学习》这本书，还有网上的资源，使得我勉强能理解简单的人工神经网络的实现方法，觉得自己在这方面真是弱爆了。
先贴一下BP神经网络的实现代码：
package com.test.bpNeuronNetworkVer2;
/**
* Created by cjz on 2018/12/19.
*/
import java.io.File;
import java.io.FileNotFoundException;
import java.io.FileOutputStream;
import java.io.IOException;
import java.io.ObjectOutputStream;
import java.io.Serializable;
import java.util.Random;
public class BPDualLayer implements Serializable{
protected static final long serialVersionUID = 1;
public final double[] input; //输入层
public final double[] hiddenLayer1; //隐含层1
public final double[] hiddenLayer2; //隐含层2
public final double[] output; //输出层
public final double[] target; //预测的输出内容
/**隐含层的神经元的误差(每一个的)**/
public final double[] hidDeltaLayer1;
public final double[] hidDeltaLayer2;
/**输出层的神经元的误差(每一个的)**/
public final double[] optDeltaLayer1;
public final double[] optDeltaLayer2;
public final double eta; //学习率
/**动量参数**/
public final double momentum;
/**从输入层到隐含层1的矩阵权值**/
public final double[][] iptToHidLayer1Weights;
/**从到隐含层1到隐含层2的矩阵权值**/
public final double[][] HidLayer1ToHidLayer2Weights;
/**从隐含层2到输出层的矩阵权值**/
public final double[][] hidLayer2ToOptWeights;
/**更新之前的权值信息(输入层到隐含层1)**/
public final double[][] iptToHid1PrevUptWeights;
/**更新之前的权值信息(隐含层1到隐含层2)**/
public final double[][] hid1ToHid2PrevUptWeights;
/**更新之前的权值信息(隐含层2到输出层)**/
public final double[][] hid2ToOptPrevUptWeights;
public double optErrSum = 0d;
public double hidLayer1ErrSum = 0d;
public double hidLayer2ErrSum = 0d;
private final Random random; //主要是对权值采取的是随机产生的方法
//初始化
public BPDualLayer(int inputSize, int hiddenSize, int outputSize, double eta,
double momentum) {
input = new double[inputSize];
hiddenLayer1 = new double[hiddenSize];
hiddenLayer2 = new double[hiddenSize];
output = new double[outputSize];
target = new double[outputSize];
hidDeltaLayer1 = new double[hiddenSize];
hidDeltaLayer2 = new double[hiddenSize];
optDeltaLayer1 = new double[hiddenSize];
optDeltaLayer2 = new double[outputSize];
iptToHidLayer1Weights = new double[inputSize][hiddenSize];
HidLayer1ToHidLayer2Weights = new double[hiddenSize][hiddenSize];
hidLayer2ToOptWeights = new double[hiddenSize][outputSize];
random = new Random(100000); //使每次产生的随机数都是第一次的分配，这是有参数和没参数的区别
randomizeWeights(iptToHidLayer1Weights); //分配输入层到隐含层的神经元的权值
randomizeWeights(HidLayer1ToHidLayer2Weights); //分配输入层到隐含层的神经元的权值
randomizeWeights(hidLayer2ToOptWeights); //分配隐含层到输出层的神经元的权值
iptToHid1PrevUptWeights = new double[inputSize][hiddenSize]; //更新权值
hid1ToHid2PrevUptWeights = new double[hiddenSize][hiddenSize]; //更新权值
hid2ToOptPrevUptWeights = new double[hiddenSize][outputSize];
this.eta = eta; //学习率
this.momentum = momentum; //动态量
}
//初始化输入层，隐含层长度，和输出层长度
public BPDualLayer(int inputSize, int hiddenSize, int outputSize) {
this(inputSize, hiddenSize, outputSize, 0.25, 0.9);
}
//随机产生神经元之间的权值信息
private void randomizeWeights(double[][] matrix) {
for(int i = 0; i < matrix.length; i++){
for(int j = 0; j < matrix[i].length; j++){
double real = random.nextDouble(); //随机分配着产生0-1之间的值
matrix[i][j] = random.nextDouble() > 0.5 ? real : -real;
}
}
}
public void train(double[] trainData, double[] target) { //训练数据
System.arraycopy(trainData, 0, this.input, 0, trainData.length); //加载输入的数据
System.arraycopy(target, 0, this.target, 0, target.length); //方法和之前的输入数据一样，都是调用复制数据的方法
forward(); //向前计算神经元权值(先算输入到隐含层的，然后再算隐含到输出层的权值)
calculateDelta(); //计算误差逆传播值
adjustWeight(); //调整更新神经元的权值
}
//测试自己弄的BP神经网络训练的效果咋样
public double[] test(double[] inData) {
System.arraycopy(inData, 0, input, 0, inData.length);
forward();
return getNetworkOutput();
}
//返回最后的输出层的结果
private double[] getNetworkOutput() {
return output;
}
//向前计算各个神经元的权值 (输入层到隐含层的)(参数一：输入层的数据,二：隐含层的内容，三：输入到隐含的神经元的权值)
//向前计算各个神经元的权值 (隐含层到输出层)(参数一：隐含层的数据,二：输出层的内容，三：隐含层到输出层的神经元的权值)
private void forward(double[] currentLayer, double[] nextLayer, double[][] weight) {
for(int j = 0; j < nextLayer.length; j++){
double sum = 0f;
for(int i = 0; i < currentLayer.length; i++){
sum += currentLayer[i] * weight[i][j]; //第i结点乘以第i结点的各个权值
}
nextLayer[j] = sigmoid(sum); //调用神经元的激活函数来得到结果(结果肯定是在0-1之间的)，并输出到下一层
}
}
//向前计算(先算输入到隐含层的，然后再算隐含到输出层的权值)
private void forward() {
forward(input, hiddenLayer1, iptToHidLayer1Weights); //计算输入层到隐含层的权值
forward(hiddenLayer1, hiddenLayer2, HidLayer1ToHidLayer2Weights); //计算输入层到隐含层的权值
forward(hiddenLayer2, output, hidLayer2ToOptWeights); //计算隐含层到输出层的权值
}
//计算输出层的误差
private void outputErr() {
double errSum = 0; //误传播值
for (int idx = 0; idx < optDeltaLayer2.length; idx++) {
double o = output[idx];
optDeltaLayer2[idx] = o * (1d - o) * (target[idx] - o); //书上p104的公式
errSum += Math.abs(optDeltaLayer2[idx]);
}
optErrSum = errSum;
}
//计算隐含层的误差
private void hiddenLayer2Err() {
double errSum = 0; //保存误差
for (int j = 0; j < hidDeltaLayer2.length; j++) {
double o = hiddenLayer2[j]; //神经元权值
double sum = 0;
for (int k = 0; k < optDeltaLayer2.length; k++){ //由输出层来反向计算
sum += hidLayer2ToOptWeights[j][k] * optDeltaLayer2[k];
}
optDeltaLayer1[j] = o * (1d - o) * sum; //书上p104的公式
hidDeltaLayer2[j] = o * (1d - o) * sum; //书上的P104的(5.15)公式
errSum += Math.abs(hidDeltaLayer2[j]);
}
hidLayer2ErrSum = errSum;
}
//计算隐含层的误差
private void hiddenLayer1Err() {
double errSum = 0; //保存误差
for (int j = 0; j < hidDeltaLayer1.length; j++) {
double o = hiddenLayer1[j]; //神经元权值
double sum = 0;
for (int k = 0; k < optDeltaLayer1.length; k++){ //由输出层来反向计算
sum += HidLayer1ToHidLayer2Weights[j][k] * optDeltaLayer1[k];
}
hidDeltaLayer1[j] = o * (1d - o) * sum; //书上的P104的(5.15)公式
errSum += Math.abs(hidDeltaLayer1[j]);
}
hidLayer1ErrSum = errSum;
}
//计算每一层的误差(因为在BP中，要达到使误差最小)(就是逆传播算法，书上有P101)
private void calculateDelta() {
outputErr(); //计算输出层的误差(因为要反过来算，所以先算输出层的)
hiddenLayer2Err(); //计算隐含层的误差
hiddenLayer1Err(); //计算隐含层的误差
}
//更新每层中的神经元的权值信息(这也就是不断的训练过程,)
private void adjustWeight(double[] delta, double[] layer,
double[][] weight, double[][] prevWeight) {
for (int i = 0; i < delta.length; i++) {
for (int j = 0; j < layer.length; j++) {
double newVal = momentum * prevWeight[j][i] + eta * delta[i] * layer[j]; //通过公式计算误差限=(动态量*之前的该神经元的阈值+学习率*误差*对应神经元的阈值)，来进行更新权值
weight[j][i] += newVal; //得到新的神经元之间的权值
prevWeight[j][i] = newVal; //保存这一次得到的权值，方便下一次进行更新
}
}
}
// //更新每层中的神经元的权值信息
// private void adjustWeight() {
// adjustWeight(optDeltaLayer2, hiddenLayer2, hidLayer2ToOptWeights, hid2ToOptPrevUptWeights);
// adjustWeight(optDeltaLayer1, hiddenLayer1, HidLayer1ToHidLayer2Weights, hid1ToHid2PrevUptWeights);
// adjustWeight(hidDeltaLayer1, input, iptToHidLayer1Weights, iptToHid1PrevUptWeights);
// }
//更新每层中的神经元的权值信息
private void adjustWeight() {
final Object a = new Object();
new Thread(new Runnable() {
@Override
public void run() {
synchronized (a) {
adjustWeight(optDeltaLayer2, hiddenLayer2, hidLayer2ToOptWeights, hid2ToOptPrevUptWeights);
}
}
}).start();
new Thread(new Runnable() {
@Override
public void run() {
synchronized (a) {
adjustWeight(optDeltaLayer1, hiddenLayer1, HidLayer1ToHidLayer2Weights, hid1ToHid2PrevUptWeights);
}
}
}).start();
new Thread(new Runnable() {
@Override
public void run() {
synchronized (a) {
adjustWeight(hidDeltaLayer1, input, iptToHidLayer1Weights, iptToHid1PrevUptWeights);
}
}
}).start();
try {
Thread.sleep(1);
synchronized (a) {
//System.out.println("finished");
}
} catch (InterruptedException e) {
e.printStackTrace();
}
}
//我这里用的是sigmoid激活函数，当然也可以用阶跃函数，看自己选择吧
private double sigmoid(double val) {
return 1d / (1d + Math.exp(-val)); //这函数书上有P98页
}
public void outputParameter() throws IOException{
File outputFile = new File("E:\\work\\ExampleProject\\人工神经网络\\手写图案的识别数据\\NeuronParameter.txt");
if(outputFile.exists()){
outputFile.delete();
}
outputFile.createNewFile();
FileOutputStream fileOutputStream = new FileOutputStream(outputFile);
writeToFileOs(fileOutputStream, "input:" + input.length + ";\n");
writeToFileOs(fileOutputStream, "hidden:" + hiddenLayer1.length + ";\n");
writeToFileOs(fileOutputStream, "output:" + output.length + ";\n");
writeToFileOs(fileOutputStream, "hidDelta:" + "\n");
for(int i = 0; i < hidDeltaLayer1.length; i ++){
writeToFileOs(fileOutputStream, String.format("%.12f", hidDeltaLayer1[i]) + (i == hidDeltaLayer1.length - 1 ? ";" : ","));
}
writeToFileOs(fileOutputStream, "\n");
writeToFileOs(fileOutputStream, "optDelta:" + "\n");
for(int i = 0; i < optDeltaLayer1.length; i ++){
writeToFileOs(fileOutputStream, String.format("%.12f", optDeltaLayer1[i]) + (i == optDeltaLayer1.length - 1 ? ";" : ","));
}
writeToFileOs(fileOutputStream, "\n");
writeToFileOs(fileOutputStream, "eta:" + eta + ";\n");
writeToFileOs(fileOutputStream, "momentum:" + momentum + ";\n");
writeToFileOs(fileOutputStream, "iptHidWeights:" + ";\n");
for(int i = 0; i < iptToHidLayer1Weights.length; i++){
for(int j = 0; j < iptToHidLayer1Weights[i].length; j++){
writeToFileOs(fileOutputStream, String.format("%.12f", iptToHidLayer1Weights[i][j]) + (j == iptToHidLayer1Weights[i].length - 1 ? ";" : ","));
}
writeToFileOs(fileOutputStream, "\n");
}
writeToFileOs(fileOutputStream, "\n");
writeToFileOs(fileOutputStream, "hidOptWeights:" + "\n");
for(int i = 0; i < hidLayer2ToOptWeights.length; i++){
for(int j = 0; j < hidLayer2ToOptWeights[i].length; j++){
writeToFileOs(fileOutputStream, String.format("%.12f", hidLayer2ToOptWeights[i][j]) + (j == hidLayer2ToOptWeights[i].length - 1 ? ";" : ","));
}
writeToFileOs(fileOutputStream, "\n");
}
writeToFileOs(fileOutputStream, "\n");
writeToFileOs(fileOutputStream, "iptHidPrevUptWeights:" + "\n");
for(int i = 0; i < iptToHid1PrevUptWeights.length; i++){
for(int j = 0; j < iptToHid1PrevUptWeights[i].length; j++){
writeToFileOs(fileOutputStream, String.format("%.12f", iptToHid1PrevUptWeights[i][j]) + (j == iptToHid1PrevUptWeights[i].length - 1 ? ";" : ","));
}
writeToFileOs(fileOutputStream, "\n");
}
writeToFileOs(fileOutputStream, "\n");
writeToFileOs(fileOutputStream, "hidOptPrevUptWeights:" + "\n");
for(int i = 0; i < hid2ToOptPrevUptWeights.length; i++){
for(int j = 0; j < hid2ToOptPrevUptWeights[i].length; j++){
writeToFileOs(fileOutputStream, String.format("%.12f", hid2ToOptPrevUptWeights[i][j]) + (j == hid2ToOptPrevUptWeights[i].length - 1 ? ";" : ","));
}
writeToFileOs(fileOutputStream, "\n");
}
writeToFileOs(fileOutputStream, "\n");
writeToFileOs(fileOutputStream, "optErrSum:" + String.format("%.12f", optErrSum) + ";\n");
writeToFileOs(fileOutputStream, "hidErrSum:" + String.format("%.12f", hidLayer1ErrSum) + ";\n");
}
public void outputBpObject(File file){
try {
if(!file.exists()){
file.createNewFile();
}
ObjectOutputStream objectOutputStream = new ObjectOutputStream(new FileOutputStream(file));
objectOutputStream.writeObject(this);
objectOutputStream.flush();
objectOutputStream.close();
System.out.println("保存BP神经网络成功");
} catch (Exception e) {
System.out.println("保存BP神经网络失败");
// TODO Auto-generated catch block
e.printStackTrace();
}
}
private void writeToFileOs(FileOutputStream fileOutputStream, String content){
try {
fileOutputStream.write(content.getBytes());
} catch (IOException e) {
// TODO Auto-generated catch block
e.printStackTrace();
}
}
}先占个位置，以后再解释原理。为了偷懒，两个隐层的细胞数我设定成一样的，读者可以根据需求进行更改。
我测试时利用了很多预先画好的图案集合导出的csv，和一个交代每一行训练像素数据的标签的csv。读者也可以更改为自己需要的集合，例如数字集合等等进行训练：
测试代码：
我的测试代码使用了两个28细胞的隐含层，目标识别正确率为超过0.990，如果达到才跳出循环。另外，trainingresult代表期望的正确输出值，例如训练集某一项数据对应的Label为0，那么trainingresult[0]则要为1，其他下标均为0，从而标记出正确的输出应该是怎样的，训练过程就会使得神经网络逼近这个结果而运行。
最后，你可以在训练到目标正确率之后，调用bp神经网络对象的outputBpObject方法得到序列化后的对象文件，然后在安卓平台上建立一样的包路径，并放到assets路径中备用，就可以把电脑上训练好的神经网络放到安卓上使用了：
package com.test.bpNeuronNetworkVer2;
import java.io.BufferedReader;
import java.io.File;
import java.io.FileInputStream;
import java.io.FileNotFoundException;
import java.io.FileReader;
import java.io.IOException;
import java.util.Scanner;
public class TestRecoginzeShapeDualLayer {
/**
* @param args
* @throws Exception
*/
// private static int hidCount = 16;
// private static int hidCount = 18;
// private static int hidCount = 30;
// private static int hidCount = 50;
private static int hidCount = 28;
// private static int hidCount = 27;
// private static int hidCount = 120;
private static int trainCount = 0;
private static double targetRightRate = 0.99f;
private static int outputCount = 4;
public static void main(String[] args) throws Exception {
System.out.println("双层神经网络,默认");
// TODO Auto-generated method stub
System.out.println("隐层神经元数目" + hidCount);
BPDualLayer bp = new BPDualLayer(50*50, hidCount, outputCount, 0.3, 0.9); // 分别代表784个输入层，一层隐含层其中是784个神经元，outputCount个输出结果
float rightRate;
new Thread(new Runnable() {
@Override
public void run() {
Scanner scanner = new Scanner(System.in);
while(true){
targetRightRate = Float.valueOf(scanner.nextLine());
System.out.println("目标正确率更改为:" + targetRightRate);
}
}
}).start();
do {
System.out.println("开始训练" + "第" + (trainCount++) + "次");
//训练数据的标签
CSVFileUtil util = new CSVFileUtil("E:\\work\\ExampleProject\\人工神经网络\\手写图案的识别数据\\trainLabels.csv");
int resultRow = util.getRowNum(); // 得到训练标签行数
// int resultRow = 20; // 得到训练标签行数
int resultColumn = util.getColNum(); // 得到训练标签列数
//训练数据,50*50=784像素
CSVFileUtil util2 = new CSVFileUtil("E:\\work\\ExampleProject\\人工神经网络\\手写图案的识别数据\\trainData.csv");
int inputRow = util2.getRowNum(); // 得到训练数据行数
int inputColumn = util2.getColNum(); // 得到训练数据列数
int prevPercent = 0;
for (int index = 0; index < resultRow; index++) {
int nowPercent = (int)(Float.valueOf(index) /Float.valueOf(resultRow) * 100);
if(nowPercent != prevPercent) {
System.out.print("+");
}
prevPercent = (int)(Float.valueOf(index) /Float.valueOf(resultRow) * 100);
double[] trainingresult = new double[outputCount]; // 因为结果有outputCount种情况
String getresult = util.getString(index, 0); // 得到训练集的第i的标签
trainingresult[Integer.parseInt(getresult)] = 1; // 表示当前的这个训练集的结果是对应的下标的值,1代表可能性为百分之几，满为1
// 将训练数据的每一个转换成只有0和1的形式
double[] binary = new double[50 * 50]; // 转成二进制进行处理(2500)
String currentTrainData = util2.getRow(index);
String currentTrainDataNum[] = currentTrainData.split("\\,");
for(int i = 0; i < currentTrainDataNum.length; i++){
// float val = Float.valueOf(currentTrainDataNum[i]) / 255f;
// binary[i] = val;
float val = Float.valueOf(currentTrainDataNum[i]);
if(val >= 255f / 2){
binary[i] = 1;
} else {
binary[i] = 0;
}
//System.out.println("::" + binary[i]);
}
bp.train(binary, trainingresult); // 训练数据，各个标签的可能性
}
System.out.println("神经网络拓扑结构训练好了，可以进行测试");
rightRate = recognize(bp);
System.out.println(String.format("正确率:%.20f", rightRate));
if(rightRate < targetRightRate){
System.out.println("正确率过低,重新训练");
if(trainCount > 500){
break;
}
} else {
System.out.println("训练完成");
}
} while(rightRate < targetRightRate);
//bp.outputParameter();
bp.outputBpObject(new File("E:\\work\\ExampleProject\\人工神经网络\\手写图案的识别数据\\neuroNetworkObject_" + System.currentTimeMillis() + "_neuronCount_" + hidCount + ".bp"));
// realRecognize(bp);
}
private static float recognize(BPDualLayer bp){
try {
// BufferedReader fileReaderTrainData = new BufferedReader(new FileReader(new File("E:\\work\\ExampleProject\\人工神经网络\\手写图案的识别数据\\trainData.csv")));
// BufferedReader fileReaderTrainDataLables = new BufferedReader(new FileReader(new File("E:\\work\\ExampleProject\\人工神经网络\\手写图案的识别数据\\trainLabels.csv")));
BufferedReader fileReaderTrainData = new BufferedReader(new FileReader(new File("E:\\work\\ExampleProject\\人工神经网络\\手写图案的识别数据\\testData.csv")));
BufferedReader fileReaderTrainDataLables = new BufferedReader(new FileReader(new File("E:\\work\\ExampleProject\\人工神经网络\\手写图案的识别数据\\testLabels.csv")));
int dataCount = 0;
int rightCount = 0;
String data = fileReaderTrainData.readLine();
String label = fileReaderTrainDataLables.readLine();
while (data != null) {
String[] dataSplited = data.split("\\,");
//System.out.println("数据长度:" + dataSplited.length);
double dataDouble[] = new double[dataSplited.length];
for (int i = 0; i < dataDouble.length; i++) {
dataDouble[i] = Double.parseDouble(dataSplited[i]);
}
double[] result = bp.test(dataDouble); // 测试的结果
int maxPossibleChoice = 0;
double maxPossible = 0;
for (int i = 0; i < result.length; i++) {
//System.out.println(i + ",识别可能性:" + String.format("%.6f", result[i]));
if(result[i] > maxPossible) {
maxPossible = result[i];
maxPossibleChoice = i;
}
}
if(Integer.parseInt(label) == maxPossibleChoice){
rightCount ++;
} else {
}
dataCount ++;
data = fileReaderTrainData.readLine();
label = fileReaderTrainDataLables.readLine();
}
//返回正确率
return Float.valueOf(rightCount) / Float.valueOf(dataCount);
} catch (FileNotFoundException e) {
// TODO Auto-generated catch block
e.printStackTrace();
} catch (IOException e) {
// TODO Auto-generated catch block
e.printStackTrace();
}
return 0f;
}
public enum Shapes{
TRANIGLE,
//LEFT_ISO_TRANIGLE,
//RIGHT_ISO_TRANIGLE,
//TOP_LEFT_ISO_TRANIGLE,
//TOP_RIGHT_ISO_TRANIGLE,
RECTANGLE,
CIRCLE,
DIAMON
//PENTAGON
}
private static void realRecognize(BPDualLayer bp){
Scanner scanner = new Scanner(System.in);
while(true){
String data = scanner.nextLine();
String[] dataSplited = data.split("\\,");
//System.out.println("数据长度:" + dataSplited.length);
double dataDouble[] = new double[dataSplited.length];
for (int i = 0; i < dataDouble.length; i++) {
dataDouble[i] = Double.parseDouble(dataSplited[i]);
}
double[] result = bp.test(dataDouble); // 测试的结果
int maxPossibleChoice = 0;
double maxPossible = 0;
for (int i = 0; i < result.length; i++) {
//System.out.println(i + ",识别可能性:" + String.format("%.6f", result[i]));
if(result[i] > maxPossible) {
maxPossible = result[i];
maxPossibleChoice = i;
}
}
System.out.println("这是一个：" + Shapes.values()[maxPossibleChoice]);
}
}
// private static double[] transPixelToBinary(){
//
// }
}
训练过程：
使用效果：
训练集以及全部代码的下载地址：
链接: https://pan.baidu.com/s/1up5_IAYCU06_oX2ePojZNA 提取码: xju8
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
