1 代码说明
1）SVM原理介绍及Python代码：
https://blog.csdn.net/weixin_44575152/article/details/100184850
2）本代码使用需要配合SimpleTool.java与Classify.java，地址如下：
https://github.com/InkiInki/Java/tree/master/Code
3）代码中所使用的数据为内置测试数据，SimpleTool.java类中包含数据读取函数，但是返回值为String[][]
4）代码的使用方法如图1-1(目前仅适用于训练集与测试集均为同一数据集的情况，其他情况需做部分修改)
5）必须需要指定的参数如下：
- tempData：待分类数据集
- tempLabel：待分类数据集标签
- tempC：Upper bound
- tempKernel：RBF核大小
- tempToler：容差大小
- tempMaxIter ：最大迭代次数
- tempKerlerType：核函数类型，目前仅支持RBF与LINEAR，且指定为KERNEL_LIN时，无需指定tempKernel大小
- tempTrainingDataIndex：训练集索引
- tempTestingDataIndex：测试集索引
6）目前仅能用于二分类，且数据集标签为{-1,1}
7）本次实验为Leave one out实验，其他实验需另行设计
Data from SimpleTool.testData1:
The simple binary data set(num-instances: 100; label: 1 or -1).
The index-21 is error: (the predicted:0.8792991168603441, the real: -1.0)
The index-30 is error: (the predicted:0.11244419302719777, the real: -1.0)
The index-54 is error: (the predicted:-0.8202579798872351, the real: 1.0)
The index-59 is error: (the predicted:0.3123402309947668, the real: -1.0)
The index-76 is error: (the predicted:0.10301136514058307, the real: -1.0)
The index-87 is error: (the predicted:0.4138048760572841, the real: -1.0)
The error rate is: 0.06
2 代码
package classify;
import java.util.Arrays;
import util.SimpleTool;
/*
* @(#)SVM.java
* The class of SVM.
* Author: Yu-Xuan Zhang
* Email: inki.yinji@qq.com
* Created: October 5, 2019.
* Last modified: December 18, 2019.
*/
public class SVM extends Classify {
/**
* The lin kernel function.
*/
public static final int KERNEL_LIN = 1;
/**
* The rbf kernel function.
*/
public static final int KERNEL_RBF = 0;
/**
* The number of trainingInstances.
*/
public static int numTrainingData;
/**
* The number of trainingInstances.
*/
public int numTestData;
/**
* The index of training data.
*/
public static int[] trainingDataIndex;
/**
* The index of test data.
*/
public int[] testDataIndex;
/**
* The upper border.
*/
private double C;
/**
* The fault tolerance rate.
*/
private double toler;
/**
* The type of kernel function.
*/
private int kernelType;
/**
* The size of kernel.
*/
private double kernel;
/**
* The array of K.
*/
private double[][] K;
/**
* The alpha.
*/
private double[] alphas;
/**
* The eCache.
*/
private double[][] eCache;
/**
* The b.
*/
private double b;
/**
* The max-number of iteration.
*/
private int maxIter;
/**
*************
* The constructor of first SVM.
*
* @param paraData The data.
* @param paralabel The label of data.
* @param paraC The upper bound.
* @param paraToler The tolerance.
* @param paraKernelType The type of kernel function.
* @param paraKernel The size of kernel.
*************
*/
public SVM(double[][] paraData, double[] paralabel, double paraC, double paraToler, int paraMaxIter,
int paraKernelType, double paraKernel) {
super(paraData, paralabel);
C = paraC;
toler = paraToler;
maxIter = paraMaxIter;
kernelType = paraKernelType;
kernel = paraKernel;
SimpleTool.tracingStarting("Hello SVM(double[][], double[], double, double, int, int, double).");
}// The first constructor
/**
*************
* The constructor of second SVM.
*
* @param paraData The data.
* @param paralabel The label of data.
* @param paraC The upper bound.
* @param paraToler The tolerance.
* @param paraKernelType The type of kernel function.
*************
*/
public SVM(double[][] paraData, double[] paralabel, double paraC, double paraToler, int paraMaxIter,
int paraKernelType) {
super(paraData, paralabel);
C = paraC;
toler = paraToler;
kernelType = paraKernelType;
SimpleTool.tracingStarting("Hello SVM(double[][], double[], double, double, int, int).");
}// The second constructor
/**
*************
* Initialize
*************
*/
private void initialize() {
b = 0;
K = new double[numM][];
// for (int i = 0; i < K.length; i++) {
// Arrays.fill(K[i], 0);
// }//Of for i
double[][] tempK = new double[numM][];
switch (kernelType) {
case KERNEL_LIN :
for (int i = 0; i < numM; i++) {
tempK[i] = SimpleTool.kernelLinear(data, data[i]);
} // Of for i
break;
case KERNEL_RBF:
for (int i = 0; i < numM; i++) {
tempK[i] = SimpleTool.kernelRBF(data, data[i], kernel);
} // Of for i
break;
default:
System.out.println("Error--> That Kernel is not recognized!");
break;
}
K = SimpleTool.arrayTranspose(tempK);
alphas = new double[numM];
eCache = new double[2][numM];
}// initialize
/**
*************
* Set index of training data.
*
* @param paraTrainingDataIndex The index of training data set.
*************
*/
public void setTrainingData(int[] paraTrainingDataIndex, int[] paraTestDataIndex) {
trainingDataIndex = paraTrainingDataIndex;
numTrainingData = trainingDataIndex.length;
testDataIndex = paraTestDataIndex;
numTestData = testDataIndex.length;
initialize();
}// Of setTrainingData
/**
************
* Select j not equal i.
*
* @param paraI The truth index i
* @return The j not equal i.
************
*/
private static int selcetJRandom(int paraI) {
int resultJ = paraI;
while (resultJ == paraI) {
resultJ = SimpleTool.random.nextInt(numTrainingData);
resultJ = trainingDataIndex[resultJ];
} // Of while
return resultJ;
}// Of selectJRandom
/**
************
* Numerical adjustment.
*
* @param paraAlphaJ The alpha of index j.
* @param paraHighest The upper bound.
* @param paraLowest The lower bound.
* @return The j not equal i.
************
*/
private double clipAlpha(double paraAlphaJ, double paraHighest, double paraLowest) {
if (paraAlphaJ > paraHighest) {
paraAlphaJ = paraHighest;
} // Of if
if (paraLowest > paraAlphaJ) {
paraAlphaJ = paraLowest;
} // Of if
return paraAlphaJ;
}// Of clipAlpha
/**
************
* Compute Error of index i.
*
* @param paraI The index of i.
* @return The Error of index i.
************
*/
private double computeErrorI(int paraI) {
double[] tempI = SimpleTool.arrayColumnValue(K, paraI);
double tempErrorI = SimpleTool.arrayMultiplyAndAdd(SimpleTool.arrayMultiply(alphas, label), tempI) + b;
return tempErrorI - label[paraI];
}// Of computeE
/**
************
* Select second j.
*
* @param paraI The index of i.
* @param errorI The error of i.
* @return The Error.
************
*/
public double[] selectJ(int paraI, double paraErrorI) {
int tempMaxI = -1;
double tempErrorJ = 0;
double tempMaxDiffError = 0;
eCache[0][paraI] = 1;
eCache[1][paraI] = paraErrorI;
int[] tempValidECacheList = SimpleTool.computeNonZeroIndex(eCache[0]);
int tempValidECacheListLen = tempValidECacheList.length;
// The error of i.
double tempErrorI = 0;
// The diffience of two error.
double tempDiffErrorI = 0;
double[] resultMaxErrorEAndErrorj = new double[2];
if (tempValidECacheListLen > 1) {
for (int i = 0; i < tempValidECacheListLen; i++) {
// Skip equal to paraI.
if (tempValidECacheList[i] == paraI) {
continue;
} // Of if
tempErrorI = computeErrorI(tempValidECacheList[i]);
tempDiffErrorI = Math.abs(paraErrorI - tempErrorI);
if (tempDiffErrorI >= tempMaxDiffError) {
tempMaxI = tempValidECacheList[i];
tempMaxDiffError = tempDiffErrorI;
tempErrorJ = tempErrorI;
} // Of if
} // Of for i
} else {
tempMaxI = selcetJRandom(paraI);
tempErrorJ = computeErrorI(tempMaxI);
} // Of if
resultMaxErrorEAndErrorj[0] = tempMaxI;
resultMaxErrorEAndErrorj[1] = tempErrorJ;
return resultMaxErrorEAndErrorj;
}// Of selectJ
/**
************
* Update the error.
*
* @param paraK The index of i.
************
*/
private void updateError(int paraI) {
eCache[0][paraI] = 1;
eCache[1][paraI] = computeErrorI(paraI);
}// Of updateError
/**
************
* SMO's inner loop.
*
* @param paraI The index of i.
* @return Does alpha change?
************
*/
private int innerL(int paraI) {
double tempErrorI = computeErrorI(paraI);
int tempJ = 0;
double tempErrorJ = 0;
double[] tempJAndErrorJ = new double[2];
double tempLowest = 0;
double tempHighest = 0;
double tempEta = 0;
double tempAlphaI = 0;
double tempAlphaJ = 0;
double tempB1 = 0;
double tempB2 = 0;
if (((label[paraI] * tempErrorI < -toler) && (alphas[paraI] < C))
|| ((label[paraI] * tempErrorI > toler) && (alphas[paraI] > 0))) {
tempJAndErrorJ = selectJ(paraI, tempErrorI);
tempJ = (int) tempJAndErrorJ[0];
tempErrorJ = tempJAndErrorJ[1];
tempAlphaI = alphas[paraI];
tempAlphaJ = alphas[tempJ];
if (label[paraI] != label[tempJ]) {
tempLowest = Math.max(0, alphas[tempJ] - alphas[paraI]);
tempHighest = Math.min(C, C + alphas[tempJ] - alphas[paraI]);
} else {
tempLowest = Math.max(0, alphas[tempJ] + alphas[paraI] - C);
tempHighest = Math.min(C, alphas[tempJ] + alphas[paraI]);
} // Of if
if (tempLowest == tempHighest) {
SimpleTool.tracingProgress("L==H");
return 0;
} // Of if
tempEta = 2 * K[paraI][tempJ] - K[paraI][paraI] - K[tempJ][tempJ];
if (tempEta >= 0) {
SimpleTool.tracingProgress("eta >= 0");
return 0;
} // Of if
alphas[tempJ] -= label[tempJ] * (tempErrorI - tempErrorJ) / tempEta;
alphas[tempJ] = clipAlpha(alphas[tempJ], tempHighest, tempLowest);
updateError(tempJ);
if (Math.abs(alphas[tempJ] - tempAlphaJ) < 1e-5) {
SimpleTool.tracingProgress("J not moving enough");
return 0;
} // Of if
alphas[paraI] += label[tempJ] * label[paraI] * (tempAlphaJ - alphas[tempJ]);
updateError(paraI);
tempB1 = b - tempErrorI - label[paraI] * (alphas[paraI] - tempAlphaI) * K[paraI][paraI]
- label[tempJ] * (alphas[tempJ] - tempAlphaJ) * K[paraI][tempJ];
tempB2 = b - tempErrorJ - label[paraI] * (alphas[paraI] - tempAlphaI) * K[paraI][paraI]
- label[tempJ] * (alphas[tempJ] - tempAlphaJ) * K[tempJ][tempJ];
if (0 < alphas[paraI] && C > alphas[paraI]) {
b = tempB1;
} else if (1e-5 < alphas[tempJ] && C > alphas[tempJ]) {
b = tempB2;
} else {
b = (tempB1 + tempB2) / 2.0;
} // Of if
return 1;
} else {
return 0;
} // Of if
}
/**
************
* The complete SMO algorithm outer loop.
*
* @return The b and alphas.
************
*/
public double[][] SMOP() {
int tempIter = 0;
// Whether to traverse the entire data set?
boolean tempEntireSet = true;
int tempAlphaPairsChanged = 0;
while ((tempIter < maxIter) && ((tempAlphaPairsChanged > 0) || tempEntireSet)) {
tempAlphaPairsChanged = 0;
if (tempEntireSet) {
for (int i = 0; i < numTrainingData; i++) {
tempAlphaPairsChanged += innerL(trainingDataIndex[i]);
SimpleTool.tracingProgress(
"Full set, iter: " + tempIter + ", i: " + i + "pairs changed " + tempAlphaPairsChanged);
} // Of for i
} else {
int[] tempNonBound = SimpleTool.computeFixedIntervalIndex(alphas, 0, C);
int tempNonBoundLen = tempNonBound.length;
for (int i = 0; i < tempNonBoundLen; i++) {
tempAlphaPairsChanged += innerL(tempNonBound[i]);
SimpleTool.tracingProgress(
"Non-bound, iter: " + tempIter + "i: " + i + ", pairs changed " + tempAlphaPairsChanged);
} // Of for i
} // Of if
tempIter++;
if (tempEntireSet) {
tempEntireSet = false;
} else if (tempAlphaPairsChanged == 0) {
tempEntireSet = true;
} // Of of
SimpleTool.tracingProgress("Iteration number: " + tempIter);
} // Of while
double[][] returnBAndAlphas = new double[2][];
returnBAndAlphas[0] = new double[1];
returnBAndAlphas[0][0] = b;
returnBAndAlphas[1] = new double[numM];
for (int i = 0; i < numM; i++) {
returnBAndAlphas[1][i] = alphas[i];
} // Of for i
SimpleTool.tracingProgress("The b is: " + b);
SimpleTool.tracingProgress("The alphas is: " + Arrays.toString(alphas));
return returnBAndAlphas;
}// Of SMOP
/**
************
* Compute the Weights.
************
*/
public double[] computeWeights() {
double[] returnW = new double[numN];
for (int i = 0; i < numN; i++) {
returnW = SimpleTool.arrayAdd(returnW, SimpleTool.arrayMultiply(alphas[i] * label[i], data[i]));
} // Of for i
return returnW;
}// Of computeWeights
/**
************
* Predictint label.
*
* @return The predicted label.
************
*/
public double[] predictedLabel() {
double[][] resultBAndAlphas = SMOP();
double b = resultBAndAlphas[0][0];
double[] alphas = resultBAndAlphas[1];
int[] nonZeroAlphas = SimpleTool.computeNonZeroIndex(alphas);
int tempNonZeroAlphasLength = nonZeroAlphas.length;
double[][] tempDataSupported = new double[tempNonZeroAlphasLength][numN];
double[] tempDataLabelSupported = new double[tempNonZeroAlphasLength];
double[] tempAlphaSupported = new double[tempNonZeroAlphasLength];
double tempPredicted;
for (int j = 0; j < tempNonZeroAlphasLength; j++) {
for (int j1 = 0; j1 < numN; j1++) {
tempDataSupported[j][j1] = data[nonZeroAlphas[j]][j1];
} // Of j2
tempDataLabelSupported[j] = label[nonZeroAlphas[j]];
tempAlphaSupported[j] = alphas[nonZeroAlphas[j]];
} // Of for j
double[] returnPredictedLabel = new double[numTestData];
for (int i = 0; i < numTestData; i++) {
double[] kernelEval = SimpleTool.kernelRBF(tempDataSupported, data[(int) testDataIndex[i]], kernel);
tempPredicted = SimpleTool.arrayMultiplyAndAdd(kernelEval,
SimpleTool.arrayMultiply(tempDataLabelSupported, tempAlphaSupported)) + b;
returnPredictedLabel[i] = tempPredicted;
} // Of for i
return returnPredictedLabel;
}// Of predictedLabel
/**
************
* The test of SVM.
************
*/
public static void testSvm() {
double[][] tempData = SimpleTool.testData1();
double[] tempLabel = SimpleTool.testLabel1();
double tempC = 1.5;
double tempKernel = .9;
double tempToler = 1e-4;
int tempMaxIter = 100;
int tempKernelType = KERNEL_RBF;
SVM svm = new SVM(tempData, tempLabel, tempC, tempToler, tempMaxIter, tempKernelType, tempKernel);
int[] tempIndex = SimpleTool.arrayIndexAuto(numM);
int[] tempTrainingDataIndex = new int[numM - 1];
int[] tempTestingDataIndex = new int[1];
double tempPredicted;
double error = 0;
for (int i = 0; i < tempLabel.length; i++) {
tempTestingDataIndex[0] = i;
tempTrainingDataIndex = SimpleTool.deleteElement(tempIndex, i);
svm.setTrainingData(tempTrainingDataIndex, tempTestingDataIndex);
tempPredicted = svm.predictedLabel()[0];
if (SimpleTool.sign(tempPredicted) != tempLabel[i]) {
SimpleTool.tracingPredictedError("The index-" + i + " is error: (the predicted:" + tempPredicted
+ ", the real: " + tempLabel[i] + ")");
error++;
} // Of if
} // Of for i
SimpleTool.tracingFinalResult("The error rate is: " + error / numM);
}// Of testSvm
/**
************
* The main
************
*/
public static void main(String[] args) {
testSvm();
} // Of main
}// Of class SVM
扫一扫
906
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
