465 lines
21 KiB
Java
465 lines
21 KiB
Java
package org.dromara.easyai.nerveCenter;
|
||
|
||
import org.dromara.easyai.conv.ConvCount;
|
||
import org.dromara.easyai.i.CustomEncoding;
|
||
import org.dromara.easyai.matrixTools.Matrix;
|
||
import org.dromara.easyai.i.ActiveFunction;
|
||
import org.dromara.easyai.nerveEntity.*;
|
||
|
||
import java.util.ArrayList;
|
||
import java.util.HashMap;
|
||
import java.util.List;
|
||
import java.util.Map;
|
||
|
||
/**
|
||
* 神经网络管理工具
|
||
* 创建神经网络
|
||
*
|
||
* @author lidapeng
|
||
* @date 11:05 上午 2019/12/21
|
||
*/
|
||
public class NerveManager extends ConvCount {
|
||
private final int hiddenNerveNub;//隐层神经元个数
|
||
private final int sensoryNerveNub;//输入神经元个数
|
||
private final int outNerveNub;//输出神经元个数
|
||
private final int hiddenDepth;//隐层深度
|
||
private final List<SensoryNerve> sensoryNerves = new ArrayList<>();//感知神经元
|
||
private SensoryNerve convInput;//卷积网络输入神经元
|
||
private final List<List<Nerve>> depthNerves = new ArrayList<>();//隐层神经元
|
||
private List<Nerve> convDepthNerves = new ArrayList<>();//卷积隐层神经元
|
||
private final List<Nerve> outNerves = new ArrayList<>();//输出神经元
|
||
private final List<Nerve> softMaxList = new ArrayList<>();//softMax层
|
||
private boolean initPower;
|
||
private float studyPoint = 0.001f;//学习率
|
||
private float convStudyPoint = 0.001f;//卷积学习率
|
||
private float oneConvRate = 0.001f;
|
||
private final ActiveFunction activeFunction;
|
||
private final int rzType;//正则化类型，默认不进行正则化
|
||
private final float lParam;//正则参数
|
||
private final int coreNumber;
|
||
private final float gaMa;//自适应学习率
|
||
private final float gMaxTh;//梯度裁剪阈值
|
||
private final boolean auto;
|
||
|
||
public SensoryNerve getConvInput() {
|
||
return convInput;
|
||
}
|
||
|
||
private Map<String, Float> conversion(Map<Integer, Float> map) {
|
||
Map<String, Float> cMap = new HashMap<>();
|
||
for (Map.Entry<Integer, Float> entry : map.entrySet()) {
|
||
cMap.put(String.valueOf(entry.getKey()), entry.getValue());
|
||
}
|
||
return cMap;
|
||
}
|
||
|
||
private Map<Integer, Float> unConversion(Map<String, Float> map) {
|
||
Map<Integer, Float> cMap = new HashMap<>();
|
||
for (Map.Entry<String, Float> entry : map.entrySet()) {
|
||
cMap.put(Integer.parseInt(entry.getKey()), entry.getValue());
|
||
}
|
||
return cMap;
|
||
}
|
||
|
||
private ModelParameter getDymModelParameter() throws Exception {//获取动态神经元参数
|
||
ModelParameter modelParameter = new ModelParameter();
|
||
List<ConvDymNerveStudy> convStudies = new ArrayList<>();
|
||
modelParameter.setDymNerveStudies(convStudies);
|
||
for (Nerve convDepthNerve : convDepthNerves) {
|
||
ConvParameter convParameter = convDepthNerve.getConvParameter();
|
||
List<Matrix> nerveMatrixList = convParameter.getNerveMatrixList();//权重矩阵
|
||
ConvDymNerveStudy convDymNerveStudy = new ConvDymNerveStudy();
|
||
List<List<Float>> oneConvList = convParameter.getOneConvPower();
|
||
List<DymNerveStudy> dymNerveStudies = new ArrayList<>();//一个卷积层的所有权重参数
|
||
convDymNerveStudy.setOneConvPower(oneConvList);
|
||
convDymNerveStudy.setDymNerveStudyList(dymNerveStudies);
|
||
for (Matrix nerveMatrix : nerveMatrixList) {
|
||
DymNerveStudy deepNerveStudy = new DymNerveStudy();//动态神经元隐层
|
||
List<Float> list = deepNerveStudy.getList();
|
||
insertWList(nerveMatrix, list);
|
||
dymNerveStudies.add(deepNerveStudy);
|
||
}
|
||
convStudies.add(convDymNerveStudy);
|
||
}
|
||
getStaticModelParameter(modelParameter);
|
||
return modelParameter;
|
||
}
|
||
|
||
private void insertWList(Matrix matrix, List<Float> list) throws Exception {//
|
||
for (int i = 0; i < matrix.getX(); i++) {
|
||
for (int j = 0; j < matrix.getY(); j++) {
|
||
list.add(matrix.getNumber(i, j));
|
||
}
|
||
}
|
||
}
|
||
|
||
public ModelParameter getConvModel() throws Exception {
|
||
return getDymModelParameter();
|
||
}
|
||
|
||
public ModelParameter getDnnModel() throws Exception {
|
||
ModelParameter modelParameter = new ModelParameter();
|
||
getStaticModelParameter(modelParameter);
|
||
return modelParameter;
|
||
}
|
||
|
||
private void getStaticModelParameter(ModelParameter modelParameter) {//获取当前模型参数
|
||
List<List<NerveStudy>> studyDepthNerves = new ArrayList<>();//隐层神经元模型
|
||
List<NerveStudy> outStudyNerves = new ArrayList<>();//输出神经元
|
||
//隐层神经元
|
||
for (List<Nerve> depthNerve : depthNerves) {
|
||
//创建一层深度的隐层神经元模型
|
||
List<NerveStudy> deepNerve = new ArrayList<>();
|
||
for (Nerve nerve : depthNerve) {
|
||
//遍历某一层深度的所有隐层神经元
|
||
NerveStudy nerveStudy = new NerveStudy();
|
||
nerveStudy.setThreshold(nerve.getThreshold());
|
||
nerveStudy.setDendrites(conversion(nerve.getDendrites()));
|
||
deepNerve.add(nerveStudy);
|
||
}
|
||
studyDepthNerves.add(deepNerve);
|
||
}
|
||
for (Nerve nerve : outNerves) {
|
||
NerveStudy nerveStudy = new NerveStudy();
|
||
nerveStudy.setThreshold(nerve.getThreshold());
|
||
nerveStudy.setDendrites(conversion(nerve.getDendrites()));
|
||
outStudyNerves.add(nerveStudy);
|
||
}
|
||
modelParameter.setDepthNerves(studyDepthNerves);
|
||
modelParameter.setOutNerves(outStudyNerves);
|
||
}
|
||
|
||
public void insertConvModel(ModelParameter modelParameter) throws Exception {
|
||
insertConvolutionModelParameter(modelParameter);//动态神经元注入
|
||
}
|
||
|
||
public void insertDnnModel(ModelParameter modelParameter) {
|
||
insertBpModelParameter(modelParameter);//全连接层注入参数
|
||
}
|
||
|
||
//注入卷积层模型参数
|
||
private void insertConvolutionModelParameter(ModelParameter modelParameter) throws Exception {
|
||
List<ConvDymNerveStudy> allDymNerveStudyList = modelParameter.getDymNerveStudies();
|
||
for (int t = 0; t < allDymNerveStudyList.size(); t++) {
|
||
ConvParameter convParameter = convDepthNerves.get(t).getConvParameter();
|
||
List<Matrix> nerveMatrixList = convParameter.getNerveMatrixList();
|
||
ConvDymNerveStudy convDymNerveStudy = allDymNerveStudyList.get(t);
|
||
List<List<Float>> oneConvPower = convDymNerveStudy.getOneConvPower();
|
||
if (oneConvPower != null && !oneConvPower.isEmpty()) {
|
||
convParameter.setOneConvPower(oneConvPower);
|
||
}
|
||
List<DymNerveStudy> dymNerveStudyList = convDymNerveStudy.getDymNerveStudyList();
|
||
if (dymNerveStudyList.size() != nerveMatrixList.size()) {
|
||
throw new Exception("卷积层数量参数与模型不匹配");
|
||
}
|
||
for (int i = 0; i < dymNerveStudyList.size(); i++) {
|
||
List<Float> list = dymNerveStudyList.get(i).getList();
|
||
Matrix nerveMatrix = nerveMatrixList.get(i);
|
||
insertMatrix(nerveMatrix, list);
|
||
}
|
||
}
|
||
insertBpModelParameter(modelParameter);//全连接层注入参数
|
||
}
|
||
|
||
private void insertMatrix(Matrix matrix, List<Float> list) throws Exception {
|
||
for (int i = 0; i < list.size(); i++) {
|
||
matrix.setNub(i, 0, list.get(i));
|
||
}
|
||
}
|
||
|
||
//注入全连接模型参数
|
||
private void insertBpModelParameter(ModelParameter modelParameter) {
|
||
List<List<NerveStudy>> depthStudyNerves = modelParameter.getDepthNerves();//隐层神经元
|
||
List<NerveStudy> outStudyNerves = modelParameter.getOutNerves();//输出神经元
|
||
//隐层神经元参数注入
|
||
for (int i = 0; i < depthNerves.size(); i++) {
|
||
List<NerveStudy> depth = depthStudyNerves.get(i);//对应的学习结果
|
||
List<Nerve> depthNerve = depthNerves.get(i);//深度隐层神经元
|
||
for (int j = 0; j < depthNerve.size(); j++) {//遍历当前深度神经元
|
||
Nerve nerve = depthNerve.get(j);
|
||
NerveStudy nerveStudy = depth.get(j);
|
||
//学习结果
|
||
Map<Integer, Float> studyDendrites = unConversion(nerveStudy.getDendrites());
|
||
//神经元参数注入
|
||
Map<Integer, Float> dendrites = nerve.getDendrites();
|
||
nerve.setThreshold(nerveStudy.getThreshold());//注入隐层阈值
|
||
for (Map.Entry<Integer, Float> entry : dendrites.entrySet()) {
|
||
int key = entry.getKey();
|
||
dendrites.put(key, studyDendrites.get(key));//注入隐层权重
|
||
}
|
||
}
|
||
}
|
||
//输出神经元参数注入
|
||
for (int i = 0; i < outNerves.size(); i++) {
|
||
Nerve outNerve = outNerves.get(i);
|
||
NerveStudy nerveStudy = outStudyNerves.get(i);
|
||
outNerve.setThreshold(nerveStudy.getThreshold());
|
||
Map<Integer, Float> dendrites = outNerve.getDendrites();
|
||
Map<Integer, Float> studyDendrites = unConversion(nerveStudy.getDendrites());
|
||
for (Map.Entry<Integer, Float> outEntry : dendrites.entrySet()) {
|
||
int key = outEntry.getKey();
|
||
dendrites.put(key, studyDendrites.get(key));
|
||
}
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 初始化神经元参数
|
||
*
|
||
* @param sensoryNerveNub 输入神经元个数
|
||
* @param hiddenNerveNub 隐层神经元个数
|
||
* @param outNerveNub 输出神经元个数
|
||
* @param hiddenDepth 隐层深度
|
||
* @param activeFunction 激活函数
|
||
* @param studyPoint 线性分类器学习率
|
||
* @param rzType 正则函数
|
||
* @param lParam 正则系数
|
||
* @param coreNumber 并行计算核心数
|
||
* @param gaMa 自适应学习率衰减系数
|
||
* @param gMaxTh 梯度裁剪阈值
|
||
* @param auTo 是否使用自适应学习率
|
||
* @throws Exception 如果参数错误则抛异常
|
||
*/
|
||
public NerveManager(int sensoryNerveNub, int hiddenNerveNub, int outNerveNub
|
||
, int hiddenDepth, ActiveFunction activeFunction, float studyPoint, int rzType, float lParam
|
||
, int coreNumber, float gaMa, float gMaxTh, boolean auTo) throws Exception {
|
||
if (sensoryNerveNub > 0 && hiddenNerveNub > 0 && outNerveNub > 0 && hiddenDepth > 0 && activeFunction != null) {
|
||
this.coreNumber = coreNumber;
|
||
this.gaMa = gaMa;
|
||
this.auto = auTo;
|
||
this.gMaxTh = gMaxTh;
|
||
this.hiddenNerveNub = hiddenNerveNub;
|
||
this.sensoryNerveNub = sensoryNerveNub;
|
||
this.outNerveNub = outNerveNub;
|
||
this.hiddenDepth = hiddenDepth;
|
||
this.activeFunction = activeFunction;
|
||
this.rzType = rzType;
|
||
this.lParam = lParam;
|
||
if (studyPoint > 0 && studyPoint < 1) {
|
||
this.studyPoint = studyPoint;
|
||
}
|
||
} else {
|
||
throw new Exception("param is null");
|
||
}
|
||
}
|
||
|
||
public List<SensoryNerve> getSensoryNerves() {//获取感知神经元集合
|
||
return sensoryNerves;
|
||
}
|
||
|
||
private List<Nerve> initConDepthNerve(int kernLen, int conHiddenDepth, ActiveFunction convFunction, int channelNo, boolean norm, float GRate) throws Exception {//初始化隐层神经元1
|
||
List<Nerve> depthNerves = new ArrayList<>();
|
||
for (int i = 0; i < conHiddenDepth; i++) {//遍历深度
|
||
float studyPoint = this.convStudyPoint;
|
||
if (studyPoint <= 0 || studyPoint > 1) {
|
||
throw new Exception("studyPoint Values range from 0 to 1");
|
||
}
|
||
int downNub = 1;
|
||
boolean isConvFinish = false;
|
||
if (i == conHiddenDepth - 1) {//卷积层最后一层
|
||
downNub = hiddenNerveNub;
|
||
isConvFinish = true;
|
||
}
|
||
HiddenNerve hiddenNerve = new HiddenNerve(1, i + 1, 1, downNub, studyPoint, initPower, convFunction, true
|
||
, rzType, lParam, kernLen, 0, 0, isConvFinish, coreNumber, channelNo, oneConvRate, norm,
|
||
null, gaMa, gMaxTh, auto, GRate);
|
||
depthNerves.add(hiddenNerve);
|
||
}
|
||
for (int i = 0; i < conHiddenDepth - 1; i++) {//遍历深度
|
||
Nerve hiddenNerve = depthNerves.get(i);//当前遍历隐层神经元
|
||
Nerve nextHiddenNerve = depthNerves.get(i + 1);//当前遍历的下一层神经元
|
||
hiddenNerve.connectSonOnly(nextHiddenNerve);
|
||
nextHiddenNerve.connectFatherOnly(hiddenNerve);
|
||
}
|
||
return depthNerves;
|
||
}
|
||
|
||
|
||
private int getNerveNub(int deep, int size, int kernLen) {
|
||
int x = size;
|
||
int step = 1;
|
||
for (int i = 0; i < deep; i++) {
|
||
x = (x - (kernLen - step)) / step;
|
||
x = x / 2 + x % 2;
|
||
}
|
||
return x;
|
||
}
|
||
|
||
/**
|
||
* 初始化卷积层神经网络
|
||
*
|
||
* @param channelNo 通道数，当该数值为1 则采用多通道降维模式 推荐值1
|
||
* @param kernLen 卷积核大小 建议为3
|
||
* @param xSize 检测窗口行高
|
||
* @param ySize 检测窗口行宽
|
||
* @param convStudyPoint 卷积层学习率
|
||
* @param convFunction 卷积层激活函数
|
||
* @param isShowLog 是否打印学习参数
|
||
* @param isSoftMax 最后一层是否用softMax激活
|
||
* @param minFeatureValue 卷积层最小特征数量的开方 取值范围 [1,50]
|
||
* @param norm 是否进行维度调节，true 进行调节， false不进行维度调节
|
||
* @param oneConvRate 降维层学习率
|
||
* @param GRate 每层的梯度衰减阈值
|
||
*/
|
||
public void initImageNet(int channelNo, int kernLen, int xSize, int ySize, boolean isSoftMax, boolean isShowLog,
|
||
float convStudyPoint, ActiveFunction convFunction, int minFeatureValue, float oneConvRate
|
||
, boolean norm, float GRate) throws Exception {
|
||
this.initPower = true;
|
||
this.oneConvRate = oneConvRate;
|
||
if (minFeatureValue < 1 || minFeatureValue > 50) {
|
||
throw new Exception("minFeatureValue 取值范围是[1,50]");
|
||
}
|
||
if (channelNo < 1) {
|
||
throw new Exception("通道数不能小于1");
|
||
}
|
||
if (!norm) {//如果不进行维度调节，通道数必须为3
|
||
channelNo = 3;
|
||
}
|
||
this.convStudyPoint = convStudyPoint;
|
||
int deep = getConvMyDep(xSize, ySize, kernLen, minFeatureValue);//卷积层深度
|
||
if (deep < 2) {
|
||
throw new Exception("minFeatureValue 设置过大");
|
||
}
|
||
List<Nerve> myDepthNerves = initConDepthNerve(kernLen, deep, convFunction, channelNo, norm, GRate);//初始化卷积层隐层
|
||
Nerve convFirstNerve = myDepthNerves.get(0);//卷积第一层隐层神经元
|
||
Nerve convLastNerve = myDepthNerves.get(myDepthNerves.size() - 1);//卷积最后一层隐层神经元
|
||
convDepthNerves = myDepthNerves;
|
||
convInput = new SensoryNerve(1, 0, channelNo);//输入神经元
|
||
//感知神经元与卷积第一层隐层神经元进行连接
|
||
convInput.connectSonOnly(convFirstNerve);
|
||
initDepthNerve(kernLen, getNerveNub(deep, xSize, kernLen), getNerveNub(deep, ySize, kernLen), channelNo, null);//初始化深度隐层神经元 depthNerves
|
||
List<Nerve> firstNerves = depthNerves.get(0);//线性层第一层隐层神经元
|
||
List<Nerve> lastNerveList = depthNerves.get(depthNerves.size() - 1);//线性层最后一层隐层神经元
|
||
convLastNerve.connect(firstNerves);//卷积最后一层链接线性层第一层
|
||
for (Nerve nerve : firstNerves) {//线性层第一层链接卷积层最后一层
|
||
nerve.connectFatherOnly(convLastNerve);
|
||
}
|
||
List<OutNerve> myOutNerveList = new ArrayList<>();
|
||
//初始化输出神经元
|
||
for (int i = 1; i < outNerveNub + 1; i++) {
|
||
OutNerve outNerve = new OutNerve(i, hiddenNerveNub, 0, studyPoint, initPower,
|
||
activeFunction, false, isShowLog, rzType, lParam, isSoftMax, 0
|
||
, coreNumber, gaMa, gMaxTh, auto, 1);
|
||
//输出层神经元连接最后一层隐层神经元
|
||
outNerve.connectFather(lastNerveList);
|
||
outNerves.add(outNerve);
|
||
myOutNerveList.add(outNerve);
|
||
}
|
||
//生成softMax层
|
||
if (isSoftMax) {//增加softMax层
|
||
SoftMax softMax = new SoftMax(outNerveNub, false, myOutNerveList, isShowLog, coreNumber);
|
||
softMaxList.add(softMax);
|
||
for (Nerve nerve : outNerves) {
|
||
nerve.connect(softMaxList);
|
||
}
|
||
}
|
||
//最后一层隐层神经元 与输出神经元进行连接
|
||
for (Nerve nerve : lastNerveList) {
|
||
nerve.connect(outNerves);
|
||
}
|
||
|
||
}
|
||
|
||
/**
|
||
* 初始化
|
||
*
|
||
* @param initPower 是否是第一次注入
|
||
* @param isShowLog 是否打印学习参数
|
||
* @param isSoftMax 最后一层是否用softMax激活
|
||
*/
|
||
public void init(boolean initPower, boolean isShowLog, boolean isSoftMax, CustomEncoding customEncoding) throws Exception {//进行神经网络的初始化构建
|
||
this.initPower = initPower;
|
||
initDepthNerve(0, 0, 0, 0, customEncoding);//初始化深度隐层神经元
|
||
List<Nerve> nerveList = depthNerves.get(0);//第一层隐层神经元
|
||
//最后一层隐层神经元啊
|
||
List<Nerve> lastNerveList = depthNerves.get(depthNerves.size() - 1);
|
||
List<OutNerve> myOutNerveList = new ArrayList<>();
|
||
//初始化输出神经元
|
||
for (int i = 1; i < outNerveNub + 1; i++) {
|
||
OutNerve outNerve = new OutNerve(i, hiddenNerveNub, 0, studyPoint, initPower,
|
||
activeFunction, false, isShowLog, rzType, lParam, isSoftMax, 0
|
||
, coreNumber, gaMa, gMaxTh, auto, 1);
|
||
//输出层神经元连接最后一层隐层神经元
|
||
outNerve.connectFather(lastNerveList);
|
||
outNerves.add(outNerve);
|
||
myOutNerveList.add(outNerve);
|
||
}
|
||
//生成softMax层
|
||
if (isSoftMax) {//增加softMax层
|
||
SoftMax softMax = new SoftMax(outNerveNub, false, myOutNerveList, isShowLog, coreNumber);
|
||
softMaxList.add(softMax);
|
||
for (Nerve nerve : outNerves) {
|
||
nerve.connect(softMaxList);
|
||
}
|
||
}
|
||
//最后一层隐层神经元 与输出神经元进行连接
|
||
for (Nerve nerve : lastNerveList) {
|
||
nerve.connect(outNerves);
|
||
}
|
||
|
||
//初始化感知神经元
|
||
for (int i = 1; i < sensoryNerveNub + 1; i++) {
|
||
SensoryNerve sensoryNerve = new SensoryNerve(i, 0, 0);
|
||
//感知神经元与第一层隐层神经元进行连接
|
||
sensoryNerve.connect(nerveList);
|
||
sensoryNerves.add(sensoryNerve);
|
||
}
|
||
|
||
}
|
||
|
||
private void initDepthNerve(int kernLen, int matrixX, int matrixY, int channelNo, CustomEncoding customEncoding) throws Exception {//初始化隐层神经元1
|
||
for (int i = 0; i < hiddenDepth; i++) {//遍历深度
|
||
List<Nerve> hiddenNerveList = new ArrayList<>();
|
||
float studyPoint = this.studyPoint;
|
||
if (studyPoint <= 0 || studyPoint > 1) {
|
||
throw new Exception("studyPoint Values range from 0 to 1");
|
||
}
|
||
CustomEncoding myCustomEncoding = null;
|
||
if (i == 0) {
|
||
myCustomEncoding = customEncoding;
|
||
}
|
||
for (int j = 1; j < hiddenNerveNub + 1; j++) {//遍历同级
|
||
int upNub;
|
||
int downNub;
|
||
int myMatrixX = 0;
|
||
int myMatrixY = 0;
|
||
if (i == 0) {
|
||
myMatrixX = matrixX;
|
||
myMatrixY = matrixY;
|
||
if (matrixX > 0 && matrixY > 0) {
|
||
upNub = matrixX * matrixY * channelNo;
|
||
} else {
|
||
upNub = sensoryNerveNub;
|
||
}
|
||
} else {
|
||
upNub = hiddenNerveNub;
|
||
}
|
||
if (i == hiddenDepth - 1) {//最后一层隐层神经元z
|
||
downNub = outNerveNub;
|
||
} else {
|
||
downNub = hiddenNerveNub;
|
||
}
|
||
HiddenNerve hiddenNerve = new HiddenNerve(j, i + 1, upNub, downNub, studyPoint, initPower, activeFunction, false
|
||
, rzType, lParam, kernLen, myMatrixX, myMatrixY, false, coreNumber, 0, oneConvRate, false
|
||
, myCustomEncoding, gaMa, gMaxTh, auto, 1);
|
||
hiddenNerveList.add(hiddenNerve);
|
||
}
|
||
depthNerves.add(hiddenNerveList);
|
||
}
|
||
initHiddenNerve();
|
||
}
|
||
|
||
private void initHiddenNerve() {//初始化隐层神经元2
|
||
for (int i = 0; i < hiddenDepth - 1; i++) {//遍历深度
|
||
List<Nerve> hiddenNerveList = depthNerves.get(i);//当前遍历隐层神经元
|
||
List<Nerve> nextHiddenNerveList = depthNerves.get(i + 1);//当前遍历的下一层神经元
|
||
for (Nerve hiddenNerve : hiddenNerveList) {
|
||
hiddenNerve.connect(nextHiddenNerveList);
|
||
}
|
||
for (Nerve nextHiddenNerve : nextHiddenNerveList) {
|
||
nextHiddenNerve.connectFather(hiddenNerveList);
|
||
}
|
||
}
|
||
}
|
||
}
|