Add File

2025-09-04 14:09:33 +08:00
parent 76432a0fcd
commit bc48177e3a
1 changed files with 87 additions and 0 deletions
--- a/src/main/java/org/dromara/easyai/resnet/ResnetManager.java
+++ b/src/main/java/org/dromara/easyai/resnet/ResnetManager.java
@@ -0,0 +1,87 @@
+package org.dromara.easyai.resnet;
+
+import org.dromara.easyai.config.ResnetConfig;
+import org.dromara.easyai.conv.ResConvCount;
+import org.dromara.easyai.i.ActiveFunction;
+import org.dromara.easyai.nerveCenter.NerveManager;
+import org.dromara.easyai.nerveEntity.SensoryNerve;
+import org.dromara.easyai.resnet.entity.ResBlockModel;
+import org.dromara.easyai.resnet.entity.ResnetModel;
+
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * @author lidapeng
+ * @time 2025/4/11 10:51
+ * @des resnet管理器
+ */
+public class ResnetManager extends ResConvCount {
+    private final NerveManager nerveManager;
+    private final List<ResBlock> resBlockList = new ArrayList<>();//残差集合
+    private final ResnetInput restNetInput;
+
+    public ResnetInput getRestNetInput() {
+        return restNetInput;
+    }
+
+    public ResnetModel getModel() throws Exception {
+        ResnetModel resnetModel = new ResnetModel();
+        List<ResBlockModel> resBlockModelList = new ArrayList<>();
+        resnetModel.setResBlockModelList(resBlockModelList);
+        for (ResBlock resBlock : resBlockList) {
+            resBlockModelList.add(resBlock.getModel());
+        }
+        resnetModel.setParameter(nerveManager.getDnnModel());
+        return resnetModel;
+    }
+
+    public void insertModel(ResnetModel resnetModel) {
+        List<ResBlockModel> resBlockModelList = resnetModel.getResBlockModelList();
+        int size = resBlockList.size();
+        for (int i = 0; i < size; i++) {
+            resBlockList.get(i).insertModel(resBlockModelList.get(i));
+        }
+        nerveManager.insertDnnModel(resnetModel.getParameter());
+    }
+
+    public ResnetManager(ResnetConfig resNetConfig, ActiveFunction activeFunction) throws Exception {
+        int deep = getConvDeep(resNetConfig.getSize(), resNetConfig.getMinFeatureSize());//获取深度
+        int channelNo = resNetConfig.getChannelNo();//通道数
+        int lastSize = getFeatureSize(deep, resNetConfig.getSize(), true);//最后一层特征大小
+        //全局学习率
+        float studyRate = resNetConfig.getStudyRate();
+        if (deep < 2) {
+            throw new Exception("图像尺寸太小了，不能用resnet进行训练");
+        }
+        int featureLength = (int) (channelNo * Math.pow(2, deep - 1));//卷积层输出特征大小
+        nerveManager = new NerveManager(featureLength, resNetConfig.getHiddenNerveNumber(), resNetConfig.getTypeNumber(), resNetConfig.getHiddenDeep()
+                , activeFunction, studyRate, resNetConfig.getRegularModel(), resNetConfig.getRegular(), 0, resNetConfig.getGaMa()
+                , resNetConfig.getGMaxTh(), resNetConfig.isAuto());
+        ResNetConnectionLine resNetConnectionLine = new ResNetConnectionLine();
+        nerveManager.init(true, resNetConfig.isShowLog(), resNetConfig.isSoftMax(), resNetConnectionLine);
+        for (int i = 0; i < deep; i++) {
+            List<SensoryNerve> sensoryNerves = null;
+            if (i == deep - 1) {
+                sensoryNerves = nerveManager.getSensoryNerves();
+            }
+            ResBlock resBlock = new ResBlock(channelNo, i + 1, studyRate, resNetConfig.getSize(), sensoryNerves, resNetConfig.getGaMa()
+                    , resNetConfig.getGMaxTh(), resNetConfig.isAuto(), resNetConfig.getGRate());
+            resBlockList.add(resBlock);
+        }
+        restNetInput = new ResnetInput(resBlockList.get(0), resNetConfig.getSize());
+        connection();//残差块进行互相连接
+        resNetConnectionLine.setLastBlock(resBlockList.get(deep - 1), lastSize, resNetConfig.getHiddenNerveNumber(), featureLength);
+    }
+
+    private void connection() {//残差块相互连接
+        int size = resBlockList.size();
+        for (int i = 0; i < size - 1; i++) {
+            ResBlock resBlock = resBlockList.get(i);
+            ResBlock nextResBlock = resBlockList.get(i + 1);
+            resBlock.setSonResBlock(nextResBlock);
+            nextResBlock.setFatherResBlock(resBlock);
+        }
+    }
+
+}