diff --git a/src/main/java/org/dromara/easyai/rnnJumpNerveEntity/Nerve.java b/src/main/java/org/dromara/easyai/rnnJumpNerveEntity/Nerve.java new file mode 100644 index 0000000..9830bff --- /dev/null +++ b/src/main/java/org/dromara/easyai/rnnJumpNerveEntity/Nerve.java @@ -0,0 +1,385 @@ +package org.dromara.easyai.rnnJumpNerveEntity; + +import org.dromara.easyai.matrixTools.Matrix; +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 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 { + private final Map> son = new HashMap<>();//轴突下一层的连接神经元 + private final Map> father = new HashMap<>();//树突上一层的连接神经元 + private final List rnnOut = new ArrayList<>();//rnn隐层输出神经元集合 + protected Map dendrites = new HashMap<>();//上一层权重(需要取出) + protected Map 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 + boolean fromOutNerve = false;//是否是输出神经元 + 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 + protected Map> features = new HashMap<>();//上一层神经元输入的数值 + protected Matrix nerveMatrix;//权重矩阵可获取及注入 + 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 + 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 int myUpNumber;//统计参数数量 + protected int depth;//所处深度 + protected int allDepth;//总深度 + protected boolean creator;//是否为创建网络 + protected int startDepth;//开始深度 + + public int getDepth() { + return depth; + } + + public Map getDendrites() { + return dendrites; + } + + public Matrix getNerveMatrix() { + return nerveMatrix; + } + + public void setNerveMatrix(Matrix nerveMatrix) { + this.nerveMatrix = nerveMatrix; + } + + public void setDendrites(Map dendrites) { + this.dendrites = dendrites; + } + + public float getThreshold() { + return threshold; + } + + public void setThreshold(float threshold) { + this.threshold = threshold; + } + + protected Nerve(int id, String name, + float studyPoint, boolean init, ActiveFunction activeFunction, int rzType, float lParam, int sensoryNerveNub + , int hiddenNerveNub, int outNerveNub, int allDepth, boolean creator, int startDepth) throws Exception {//该神经元在同层神经元中的编号 + this.id = id; + this.creator = creator; + this.startDepth = startDepth; + this.allDepth = allDepth; + this.hiddenNerveNub = hiddenNerveNub;//隐层神经元个数 + this.sensoryNerveNub = sensoryNerveNub;//输入神经元个数 + this.outNerveNub = outNerveNub;//输出神经元个数 + this.name = name; + this.studyPoint = studyPoint; + this.activeFunction = activeFunction; + this.rzType = rzType; + this.lParam = lParam; + if (name.equals("OutNerve")) { + fromOutNerve = true; + } + initPower(init);//生成随机权重 + } + + protected void setStudyPoint(float studyPoint) { + this.studyPoint = studyPoint; + } + + private int getNextStorey(int[] storeys, int index) { + int nextStorey = -1; + int nextIndex = index + 1; + if (storeys.length > nextIndex) {//可以继续前进 + nextStorey = storeys[nextIndex]; + } + return nextStorey; + } + + protected void sendSoftMaxBack(long eventId, float parameter, Matrix rnnMatrix, OutBack outBack, String myWord) throws Exception { + if (!son.isEmpty()) { + List nerverList = son.get(0); + for (Nerve nerve : nerverList) { + nerve.sendAppointSoftMax(eventId, parameter, rnnMatrix, outBack, myWord); + } + } else { + throw new Exception("this storey is lastIndex"); + } + } + + protected void sendSoftMax(long eventId, float parameter, boolean isStudy, Map E + , OutBack outBack, Matrix rnnMatrix, int[] storeys, int index) throws Exception { + if (!son.isEmpty()) { + List nerverList = son.get(0); + for (Nerve nerve : nerverList) { + nerve.input(eventId, parameter, isStudy, E, outBack, rnnMatrix, storeys, index, 0); + } + } else { + throw new Exception("this storey is lastIndex"); + } + } + + protected void clearData(long eventId) { + } + + protected void sendMyTestMessage(long eventId, Matrix featureMatrix, OutBack outBack, String word) throws Exception { + + } + + protected void sendAppointSoftMax(long eventId, float parameter, Matrix rnnMatrix, OutBack outBack, String myWord) throws Exception { + } + + protected void sendAppointTestMessage(long eventId, float parameter, Matrix featureMatrix, OutBack outBack, String myWord) throws Exception { + } + + protected void sendTestMessage(long eventId, float parameter, Matrix featureMatrix, OutBack outBack, String myWord) throws Exception { + if (!son.isEmpty()) { + List nerveList = son.get(depth + 1); + if (nerveList != null) { + for (Nerve nerve : nerveList) { + nerve.sendAppointTestMessage(eventId, parameter, featureMatrix, outBack, myWord); + } + } else { + throw new Exception("Insufficient layer:" + depth + 1); + } + } else { + throw new Exception("this layer is lastIndex"); + } + } + + protected void sendRnnTestMessage(long eventId, float parameter, Matrix featureMatrix, OutBack outBack, String myWord) throws Exception { + if (!rnnOut.isEmpty()) { + for (Nerve nerve : rnnOut) { + nerve.sendAppointTestMessage(eventId, parameter, featureMatrix, outBack, myWord); + } + } else { + throw new Exception("this layer is lastIndex"); + } + } + + protected void sendMessage(long eventId, float parameter, boolean isStudy, Map E + , OutBack outBack, Matrix rnnMatrix, int[] storeys, int index, int questionLength) throws Exception { + List nerveList = null; + int nextStorey = 0; + if (storeys == null) { + nerveList = son.get(0); + } else { + nextStorey = getNextStorey(storeys, index); + if (nextStorey > -1) {//可以继续向前 + nerveList = son.get(nextStorey); + index++; + if (nerveList == null) { + throw new Exception("向前->要查找的层数不存在链接，序列：" + index + "层数:" + nextStorey + + ",当前所在层数:" + depth + ",我的身份:" + name); + } + } + } + if (nerveList != null) { + if (creator && !isStudy && nextStorey == startDepth) { + for (Nerve nerve : nerveList) { + nerve.sendAppointTestMessage(eventId, parameter, rnnMatrix, outBack, null); + } + } else { + for (Nerve nerve : nerveList) { + nerve.input(eventId, parameter, isStudy, E, outBack, rnnMatrix, storeys, index, questionLength); + } + } + } else {//发送到输出神经元 + sendRnnMessage(eventId, parameter, isStudy, E, outBack, rnnMatrix, storeys, index); + } + } + + private void sendRnnMessage(long eventId, float parameter, boolean isStudy, Map E + , OutBack outBack, Matrix rnnMatrix, int[] storeys, int index) throws Exception { + if (!rnnOut.isEmpty()) { + for (Nerve nerve : rnnOut) { + nerve.input(eventId, parameter, isStudy, E, outBack, rnnMatrix, storeys, index, 0); + } + } else { + throw new Exception("this layer is lastIndex"); + } + } + + private void backSendMessage(long eventId, boolean fromOutNerve, int[] storeys, int index) throws Exception {//反向传播 + if (!father.isEmpty()) {//要先判定是不是可以继续往后传 + List nerveList = null; + if (storeys == null) {//常规依次回退 + nerveList = father.get(0); + } else if (index > 0) {//可以继续向后传递 + nerveList = father.get(storeys[index]); + if (nerveList == null) { + throw new Exception("向后->要查找的层数不存在链接，序列：" + index + "目标层数:" + storeys[index] + + ",当前所在层数:" + depth + ",我的身份:" + name); + } + index--; + } + if (nerveList != null) { + for (int i = 0; i < nerveList.size(); i++) { + nerveList.get(i).backGetMessage(wg.get(i + 1), eventId, fromOutNerve, storeys, index); + } + } + } + } + + protected void input(long eventId, float parameter, boolean isStudy + , Map E, OutBack imageBack, Matrix rnnMatrix + , int[] storeys, int index, int questionLength) throws Exception {//输入参数 + + } + + private void backGetMessage(float parameter, long eventId, boolean fromOutNerve, int[] storeys, int index) throws Exception {//反向传播 + backNub++; + sigmaW = sigmaW + parameter; + int number; + if (fromOutNerve) { + number = outNerveNub; + } else { + number = hiddenNerveNub; + } + if (backNub == number) {//进行新的梯度计算 + backNub = 0; + gradient = activeFunction.functionG(outNub) * sigmaW; + updatePower(eventId, storeys, index);//修改阈值 + } + } + + protected void updatePower(long eventId, int[] storeys, int index) throws Exception {//修改阈值 + float h = gradient * studyPoint; + threshold = threshold - h; + updateW(h, eventId); + sigmaW = 0;//求和结果归零 + backSendMessage(eventId, fromOutNerve, storeys, index); + } + + 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 list = features.get(eventId); + float param = 0; + if (rzType != RZ.NOT_RZ) { + float sigma = 0; + for (Map.Entry entry : dendrites.entrySet()) { + if (rzType == RZ.L2) { + sigma = sigma + (float) Math.pow(entry.getValue(), 2); + } else { + sigma = sigma + (float) Math.abs(entry.getValue()); + } + } + param = sigma * lParam * studyPoint; + } + for (Map.Entry 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 wp = bn * h; + float dm = w * gradient; + 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 boolean insertParameter(long eventId, float parameter) throws Exception {//添加参数 + boolean allReady = false; + List featuresList; + if (features.containsKey(eventId)) { + featuresList = features.get(eventId); + } else { + featuresList = new ArrayList<>(); + features.put(eventId, featuresList); + } + featuresList.add(parameter); + if (featuresList.size() == myUpNumber) { + allReady = true; + } else if (featuresList.size() > myUpNumber) { + throw new Exception("接收参数数量异常"); + } + return allReady; + } + + protected void destroyParameter(long eventId) {//销毁参数 + features.remove(eventId); + } + + protected float calculation(long eventId) {//计算当前输出结果 + float sigma = 0; + List featuresList = features.get(eventId); + for (int i = 0; i < featuresList.size(); i++) { + float value = featuresList.get(i); + float w = dendrites.get(i + 1);//当value不为0的时候把w取出来 + sigma = w * value + sigma; + } + return sigma - threshold; + } + + private void initPower(boolean init) throws Exception {//初始化权重及阈值 + Random random = new Random(); + //设置初始化权重范围收缩系数 + if (name.equals("HiddenNerve")) {//隐层神经元 + myUpNumber = sensoryNerveNub; + } else if (name.equals("OutNerve")) {//输出神经元 + myUpNumber = hiddenNerveNub; + } else {//softmax + myUpNumber = outNerveNub; + } + if (myUpNumber > 0) {//输入个数 + float sh = (float) Math.sqrt(myUpNumber); + for (int i = 1; i < myUpNumber + 1; i++) { + float nub = 0; + if (init) { + nub = random.nextFloat() / sh; + } + dendrites.put(i, nub);//random.nextFloat() + } + //生成随机阈值 + float nub = 0; + if (init) { + nub = random.nextFloat() / sh; + } + threshold = nub; + } + } + + public int getId() { + return id; + } + + + public void connect(int depth, List nerveList) { + son.put(depth, nerveList);//连接下一层 + } + + public void connectOut(List nerveList) { + rnnOut.addAll(nerveList); + } + + public void connectFather(int depth, List nerveList) { + father.put(depth, nerveList); + } +}