This commit is contained in:
2025-09-04 14:08:44 +08:00
parent 548aee4e1b
commit d7711aa5dc

View File

@@ -0,0 +1,458 @@
package org.dromara.easyai.nerveEntity;
import org.dromara.easyai.conv.ConvCount;
import org.dromara.easyai.conv.ConvResult;
import org.dromara.easyai.conv.DymStudy;
import org.dromara.easyai.entity.ThreeChannelMatrix;
import org.dromara.easyai.i.CustomEncoding;
import org.dromara.easyai.matrixTools.Matrix;
import org.dromara.easyai.matrixTools.MatrixList;
import org.dromara.easyai.matrixTools.MatrixOperation;
import org.dromara.easyai.config.RZ;
import org.dromara.easyai.i.ActiveFunction;
import org.dromara.easyai.i.OutBack;
import org.dromara.easyai.rnnJumpNerveCenter.CustomManager;
import java.util.*;
/**
* @author lidapeng
* 神经元,所有类别神经元都要继承的类,具有公用属性
* @date 9:36 上午 2019/12/21
*/
public abstract class Nerve extends ConvCount {
private final List<Nerve> son = new ArrayList<>();//轴突下一层的连接神经元
private final List<Nerve> father = new ArrayList<>();//树突上一层的连接神经元
private Nerve sonOnly;
private Nerve fatherOnly;
protected final float gaMa;//自适应学习率衰减系数
protected Map<Integer, Float> dendrites = new HashMap<>();//上一层权重(需要取出)
private final Map<Integer, Float> dymStudyRate = new HashMap<>();//动态学习率
protected Map<Integer, Float> wg = new HashMap<>();//上一层权重与梯度的积
private final int id;//同级神经元编号,注意在同层编号中ID应有唯一性
protected int upNub;//上一层神经元数量
protected int downNub;//下一层神经元的数量
protected Map<Long, List<Float>> features = new HashMap<>();//上一层神经元输入的数值
protected float threshold;//此神经元的阈值需要取出
protected String name;//该神经元所属类型
protected float outNub;//输出数值ps:只有训练模式的时候才可保存输出过的数值)
protected float E;//模板期望值
protected float gradient;//当前梯度
protected float studyPoint;
protected float sigmaW;//对上一层权重与上一层梯度的积进行求和
protected List<Matrix> sigmaMatrix;//对上层回传误差进行求和
private int backNub = 0;//当前节点被反向传播的次数
protected 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 kernLen;//核长
protected final int depth;//所处深度
protected final int matrixX;//卷积输出行数列数
protected final int matrixY;//卷积输出矩阵列数
private final MatrixOperation matrixOperation;
protected final int channelNo;//通道数
private final ConvParameter convParameter = new ConvParameter();//内存中卷积层模型及临时数据
protected final float oneConvRate;
private final boolean norm;//是否进行1v1卷积升降维
private final CustomEncoding customEncoding;//自定义编码模块
private final float gMaxTh;
private final DymStudy dymStudy;
private final boolean auto;
private final float GRate;//梯度衰减
public Map<Integer, Float> getDendrites() {
return dendrites;
}
public ConvParameter getConvParameter() {
return convParameter;
}
public void setDendrites(Map<Integer, Float> dendrites) {
this.dendrites = dendrites;
}
public float getThreshold() {
return threshold;
}
public void setThreshold(float threshold) {
this.threshold = threshold;
}
protected Nerve(int id, int upNub, String name, int downNub,
float studyPoint, boolean init, ActiveFunction activeFunction
, boolean isDynamic, int rzType, float lParam, int kernLen, int depth
, int matrixX, int matrixY, int coreNumber, int channelNo, float onConvRate, boolean norm, CustomEncoding customEncoding
, float gaMa, float gMaxTh, boolean auto, float GRate) throws Exception {//该神经元在同层神经元中的编号
if (auto) {
if (gaMa <= 0 || gaMa >= 1) {
throw new IllegalArgumentException("gaMa 取值范围是(0,1),当前值:" + gaMa);
}
if (gMaxTh <= 0) {
throw new IllegalArgumentException("gMaxTh 必须比0大,当前值:" + gMaxTh);
}
}
matrixOperation = new MatrixOperation(coreNumber);
dymStudy = new DymStudy(gaMa, gMaxTh, auto);
this.GRate = GRate;
this.gaMa = gaMa;
this.auto = auto;
this.gMaxTh = gMaxTh;
this.matrixX = matrixX;
this.customEncoding = customEncoding;
this.norm = norm;
this.matrixY = matrixY;
this.channelNo = channelNo;
this.id = id;
this.depth = depth;
this.upNub = upNub;
this.name = name;
this.downNub = downNub;
this.studyPoint = studyPoint;
this.activeFunction = activeFunction;
this.rzType = rzType;
this.lParam = lParam;
this.kernLen = kernLen;
this.oneConvRate = onConvRate;
initPower(init, isDynamic);//生成随机权重
}
protected void setStudyPoint(float studyPoint) {
this.studyPoint = studyPoint;
}
public void sendMessage(long eventId, float parameter, boolean isStudy, Map<Integer, Float> E
, OutBack outBack) throws Exception {
if (!son.isEmpty()) {
for (Nerve nerve : son) {
nerve.input(eventId, parameter, isStudy, E, outBack);
}
} else {
throw new Exception("this layer is lastIndex");
}
}
protected List<Matrix> conv(List<Matrix> matrix) throws Exception {//一次正向卷积,下取样
return downConvAndPooling(matrix, convParameter, channelNo, activeFunction, kernLen, true, -1);
}
protected void demRedByMatrixList(long eventId, List<Matrix> matrixList, boolean study,
Map<Integer, Float> E, OutBack outBack, boolean needMatrix) throws Exception {
if (study) {//训练临时保存
convParameter.setFeatureMatrixList(matrixList);
}
List<Matrix> feature;
if (norm) {
feature = manyOneConv(matrixList, convParameter.getOneConvPower());//降维后的特征矩阵
} else {
if (matrixList.size() != 3) {
throw new Exception("不进行维度调节输入的特征矩阵通道数必须为3");
}
feature = matrixList;
}
List<Matrix> convMatrix = conv(feature);
sendMatrix(eventId, convMatrix, study, E, outBack, needMatrix);
}
public void sendMatrixList(long eventId, List<Float> parameter, boolean isStudy,
Map<Integer, Float> E, OutBack outBack) throws Exception {
if (!son.isEmpty()) {
for (Nerve nerve : son) {
nerve.inputMatrixFeature(eventId, parameter, isStudy, E, outBack);
}
} else {
throw new Exception("this layer is lastIndex");
}
}
public void sendMatrix(long eventId, List<Matrix> parameter, boolean isStudy,
Map<Integer, Float> E, OutBack outBack, boolean needMatrix) throws Exception {
if (sonOnly != null) {
sonOnly.inputMatrix(eventId, parameter, isStudy, E, outBack, needMatrix);
} else {
throw new Exception("this layer is lastIndex");
}
}
public void sendThreeChannelMatrix(long eventId, ThreeChannelMatrix parameter, boolean isStudy,
Map<Integer, Float> E, OutBack outBack, boolean needMatrix) throws Exception {
if (sonOnly != null) {
sonOnly.inputThreeChannelMatrix(eventId, parameter, isStudy, E, outBack, needMatrix);
} else {
throw new Exception("this layer is lastIndex");
}
}
public void sendListMatrix(long eventId, List<Matrix> parameter, boolean isStudy,
Map<Integer, Float> E, OutBack outBack, boolean needMatrix) throws Exception {
if (sonOnly != null) {
sonOnly.demRedByMatrixList(eventId, parameter, isStudy, E, outBack, needMatrix);
} else {
throw new Exception("this layer is lastIndex");
}
}
private void backSendMessage(long eventId) throws Exception {//反向传播
if (!father.isEmpty()) {
for (int i = 0; i < father.size(); i++) {
father.get(i).backGetMessage(wg.get(i + 1), eventId);
}
} else if (fatherOnly != null && depth == 1) {//反矩阵误差
List<Matrix> errorMatrixList = new ArrayList<>();
int size = this.matrixX * this.matrixY;
int featureSize = wg.size() / size;
for (int i = 0; i < featureSize; ++i) {
List<Float> list = new ArrayList<>();
int startIndex = size * i;
int endIndex = startIndex + size;
for (int j = startIndex; j < endIndex; j++) {
list.add(this.wg.get(j + 1));
}
Matrix errorMatrix = this.matrixOperation.ListToMatrix(list, this.matrixX, this.matrixY);
errorMatrixList.add(errorMatrix);
}
fatherOnly.backMatrix(errorMatrixList);
} else if (customEncoding != null && depth == 1) {//最后一层返回给第一层
customEncoding.backError(wg, id);
}
}
private void backMatrixMessage(List<Matrix> g) throws Exception {//反向传播矩阵
if (fatherOnly != null) {
fatherOnly.backMatrix(g);
}
}
protected void input(long eventId, float parameter, boolean isStudy
, Map<Integer, Float> E, OutBack imageBack) throws Exception {//输入参数
}
protected void inputMatrixFeature(long eventId, List<Float> parameters, boolean isStudy
, Map<Integer, Float> E, OutBack imageBack) throws Exception {//卷积层向网络发送参数
}
protected void inputMatrix(long eventId, List<Matrix> matrix, boolean isKernelStudy, Map<Integer, Float> E,
OutBack outBack, boolean needMatrix) throws Exception {//输入动态矩阵
}
protected void inputThreeChannelMatrix(long eventId, ThreeChannelMatrix picture, boolean isKernelStudy, Map<Integer, Float> E,
OutBack outBack, boolean needMatrix) throws Exception {//输入动态矩阵
}
private void backGetMessage(float parameter, long eventId) throws Exception {//反向传播
backNub++;
sigmaW = sigmaW + parameter;
if (backNub == downNub) {//进行新的梯度计算
backNub = 0;
gradient = activeFunction.functionG(outNub) * sigmaW;
updatePower(eventId);//修改阈值
}
}
protected void backMatrix(List<Matrix> t) throws Exception {//回传梯度
backNub++;
if (sigmaMatrix == null) {
sigmaMatrix = t;
} else {
sigmaMatrix = matrixOperation.addMatrixList(t, sigmaMatrix);
}
if (backNub == downNub) {
backNub = 0;
List<Matrix> errorMatrix = backDownPoolingByList(sigmaMatrix, convParameter.getOutX(), convParameter.getOutY());//池化误差返回
matrixOperation.mathMulByList(errorMatrix, GRate);
List<Matrix> myErrorMatrix = backAllDownConv(convParameter, errorMatrix, studyPoint, activeFunction, channelNo, kernLen,
gaMa, gMaxTh, auto);
sigmaMatrix = null;
if (depth == 1) {//1*1 卷积调整
if (norm) {
backOneConvByList(myErrorMatrix, convParameter.getFeatureMatrixList(), convParameter.getOneConvPower(), oneConvRate
, convParameter.getOneDymStudyRateList(), gaMa, gMaxTh, auto);
}
} else {
//将梯度继续回传
backMatrixMessage(myErrorMatrix);
}
}
}
protected void updatePower(long eventId) throws Exception {//修改阈值
float thError = dymStudy.getOneValueError(studyPoint, gradient, convParameter);
threshold = threshold - thError;
updateW(gradient, eventId);
sigmaW = 0;//求和结果归零
backSendMessage(eventId);
}
private float regularization(float w, float param) {//正则化类型
float re = 0.0f;
if (rzType != RZ.NOT_RZ) {
if (rzType == RZ.L2) {
re = param * -w;
} else if (rzType == RZ.L1) {
if (w > 0) {
re = -param;
} else if (w < 0) {
re = param;
}
}
}
return re;
}
private void updateW(float h, long eventId) {//h是学习率 * 当前g梯度
List<Float> list = features.get(eventId);
float param = 0;
if (rzType != RZ.NOT_RZ) {
float sigma = 0;
for (Map.Entry<Integer, Float> entry : dendrites.entrySet()) {
if (rzType == RZ.L2) {
sigma = sigma + (float) Math.pow(entry.getValue(), 2);
} else {
sigma = sigma + Math.abs(entry.getValue());
}
}
param = sigma * lParam * studyPoint;
}
for (Map.Entry<Integer, Float> entry : dendrites.entrySet()) {
int key = entry.getKey();//上层隐层神经元的编号
float w = entry.getValue();//接收到编号为KEY的上层隐层神经元的权重
float bn = list.get(key - 1);//接收到编号为KEY的上层隐层神经元的输入
float error = dymStudy.getNerveStudyError(dymStudyRate, key, h, studyPoint);
float wp = bn * error;
float dm = w * h;
float regular = regularization(w, param);//正则化抑制权重s
w = w + regular;
w = w + wp;
wg.put(key, dm);//保存上一层权重与梯度的积
dendrites.put(key, w);//保存修正结果
}
features.remove(eventId); //清空当前上层输入参数参数
}
protected void insertParameters(long eventId, List<Float> parameters) {//添加参数
List<Float> featuresList;
if (features.containsKey(eventId)) {
featuresList = features.get(eventId);
} else {
featuresList = new ArrayList<>();
features.put(eventId, featuresList);
}
featuresList.addAll(parameters);
}
protected boolean insertParameter(long eventId, float parameter) {//添加参数
boolean allReady = false;
List<Float> featuresList;
if (features.containsKey(eventId)) {
featuresList = features.get(eventId);
} else {
featuresList = new ArrayList<>();
features.put(eventId, featuresList);
}
featuresList.add(parameter);
if (featuresList.size() >= upNub) {
allReady = true;
}
return allReady;
}
protected void destoryParameter(long eventId) {//销毁参数
features.remove(eventId);
}
protected float calculation(long eventId) throws Exception {//计算当前输出结果
float sigma = 0;
List<Float> featuresList = features.get(eventId);
if (dendrites.size() != featuresList.size()) {
throw new Exception("权重数量:" + dendrites.size() + ",特征数量:" + featuresList.size());
}
for (int i = 0; i < featuresList.size(); i++) {
float value = featuresList.get(i);
float w = dendrites.get(i + 1);
sigma = w * value + sigma;
}
return sigma - threshold;
}
private void initPower(boolean init, boolean isDynamic) throws Exception {//初始化权重及阈值
Random random = new Random();
if (!isDynamic) {//静态神经元
//设置初始化权重范围收缩系数
if (upNub > 0) {//输入个数
for (int i = 1; i < upNub + 1; i++) {
float nub = 0;
if (init) {
nub = random.nextFloat() / (float) Math.sqrt(upNub);
}
dendrites.put(i, nub);//random.nextFloat()
dymStudyRate.put(i, 0f);
}
//生成随机阈值
float nub = 0;
if (init) {
nub = random.nextFloat() / (float) Math.sqrt(upNub);
}
threshold = nub;
}
} else {//动态神经元
initMatrixPower(random);
}
}
private void initMatrixPower(Random random) throws Exception {
int nerveNub = kernLen * kernLen;
List<Matrix> nerveMatrixList = convParameter.getNerveMatrixList();//一层当中所有的深度卷积核
List<Matrix> dymStudyRateList = convParameter.getDymStudyRateList();
List<ConvSize> convSizeList = convParameter.getConvSizeList();
List<List<Float>> onePowers = new ArrayList<>();//1*1卷积核
List<List<Float>> oneDymStudyRateList = new ArrayList<>();
for (int k = 0; k < channelNo; k++) {//遍历通道
Matrix nerveMatrix = new Matrix(nerveNub, 1);//一组通道创建一组卷积核
convSizeList.add(new ConvSize());
for (int i = 0; i < nerveMatrix.getX(); i++) {//初始化深度卷积核权重
float nub = random.nextFloat() / kernLen;
nerveMatrix.setNub(i, 0, nub);
}
nerveMatrixList.add(nerveMatrix);
dymStudyRateList.add(new Matrix(nerveNub, 1));
if (depth == 1) {
List<Float> oneConvPowerList = new ArrayList<>();
List<Float> oneDymStudyRate = new ArrayList<>();
for (int i = 0; i < 3; i++) {
oneConvPowerList.add(random.nextFloat() / 3);
oneDymStudyRate.add(0f);
}
oneDymStudyRateList.add(oneDymStudyRate);
onePowers.add(oneConvPowerList);
}
}
if (depth == 1) {
convParameter.setOneDymStudyRateList(oneDymStudyRateList);
convParameter.setOneConvPower(onePowers);
}
}
public int getId() {
return id;
}
public void connect(List<Nerve> nerveList) {
son.addAll(nerveList);//连接下一层
}
public void connectSonOnly(Nerve nerve) {
sonOnly = nerve;
}
public void connectFatherOnly(Nerve nerve) {
fatherOnly = nerve;
}
public void connectFather(List<Nerve> nerveList) {
father.addAll(nerveList);//连接上一层
}
}