Add File

src/main/java/org/dromara/easyai/matrixTools/MatrixNorm.java

@@ -0,0 +1,108 @@
+package org.dromara.easyai.matrixTools;
+
+import org.dromara.easyai.conv.DymStudy;
+import org.dromara.easyai.resnet.entity.NormModel;
+
+import java.util.Random;
+
+/**
+ * @author lidapeng
+ * @time 2025/4/11 15:27
+ * @des 矩阵批量归一化
+ */
+public class MatrixNorm {
+    private Matrix bTa;//偏移值
+    private Matrix power;//膨胀系数矩阵
+    private final Matrix bTaDymStudyRate;//偏移系数动态学习率
+    private final Matrix powerDymStudyRate;//膨胀系数动态学习率
+    private final DymStudy dymStudy;
+    private final float studyRate;//全局学习率
+    private final MatrixOperation matrixOperation = new MatrixOperation();
+    private Matrix norm;
+
+    public NormModel getModel() {
+        NormModel normModel = new NormModel();
+        normModel.setBtaParameter(bTa.getMatrixModel());
+        normModel.setPowerParameter(power.getMatrixModel());
+        return normModel;
+    }
+
+    public void insertModel(NormModel normModel) {
+        bTa.insertMatrixModel(normModel.getBtaParameter());
+        power.insertMatrixModel(normModel.getPowerParameter());
+    }
+
+    public MatrixNorm(int size, float studyRate, float gaMa, float gMaxTh, boolean auTo) throws Exception {
+        dymStudy = new DymStudy(gaMa, gMaxTh, auTo);
+        bTa = new Matrix(size, size);
+        power = new Matrix(size, size);
+        bTaDymStudyRate = new Matrix(size, size);
+        powerDymStudyRate = new Matrix(size, size);
+        this.studyRate = studyRate;
+        initPower(power);
+        initPower(bTa);
+    }
+
+    private void initPower(Matrix matrix) throws Exception {
+        int size = matrix.getX();
+        Random random = new Random();
+        for (int i = 0; i < size; i++) {
+            for (int j = 0; j < size; j++) {
+                matrix.setNub(i, j, random.nextFloat() / size);
+            }
+        }
+    }
+
+    private Matrix back(Matrix errorMatrix, Matrix myData) throws Exception {
+        Matrix subPower = matrixOperation.matrixMulPd(errorMatrix, myData, power, false);
+        Matrix sub = matrixOperation.matrixMulPd(errorMatrix, myData, power, true);
+        int x = sub.getX();
+        int y = sub.getY();
+        Matrix errorPower = dymStudy.getErrorMatrixByStudy(studyRate, powerDymStudyRate, subPower);
+        power = matrixOperation.add(errorPower, power);
+        float n = (float) Math.sqrt(x * y);
+        float nt = -n / (n - 1);
+        Matrix subMatrix = new Matrix(x, y);
+        for (int i = 0; i < x; i++) {
+            for (int j = 0; j < y; j++) {
+                float subValue = sub.getNumber(i, j);
+                float value = subValue * n + subMatrix.getNumber(i, j);
+                subMatrix.setNub(i, j, value);
+                for (int k = 0; k < x; k++) {
+                    for (int l = 0; l < y; l++) {
+                        if (k != i || l != j) {
+                            float otherValue = subValue * nt + subMatrix.getNumber(k, l);
+                            subMatrix.setNub(k, l, otherValue);
+                        }
+                    }
+                }
+            }
+        }
+        return subMatrix;
+    }
+
+    public Matrix backError(Matrix errorMatrix) throws Exception {
+        Matrix error = dymStudy.getErrorMatrixByStudy(studyRate, bTaDymStudyRate, errorMatrix);
+        bTa = matrixOperation.add(error, bTa);//更新bTa
+        return back(errorMatrix, norm);
+    }
+
+    public Matrix norm(Matrix matrix) throws Exception {
+        int x = matrix.getX();
+        int y = matrix.getY();
+        if (x != y) {
+            throw new Exception("必须是方阵才能进行全矩阵的归一化");
+        }
+        Matrix result = new Matrix(x, y);
+        float avg = matrix.getAVG();//平均值
+        float sd = matrixOperation.getSdByMatrix(matrix, avg, 0.0000001f);//标准差
+        for (int i = 0; i < x; i++) {
+            for (int j = 0; j < y; j++) {
+                float value = (matrix.getNumber(i, j) - avg) / sd;
+                result.setNub(i, j, value);
+            }
+        }
+        norm = result;
+        return matrixOperation.add(matrixOperation.mulMatrix(result, power), bTa);
+    }
+}