Add File

src/main/java/org/dromara/easyai/conv/DymStudy.java

@@ -0,0 +1,108 @@
+package org.dromara.easyai.conv;
+
+import org.dromara.easyai.matrixTools.Matrix;
+import org.dromara.easyai.matrixTools.MatrixOperation;
+import org.dromara.easyai.nerveEntity.ConvParameter;
+
+import java.util.List;
+import java.util.Map;
+
+/**
+ * @author lidapeng
+ * @time 2025/4/19 09:24
+ * @des 学习率处理
+ */
+public class DymStudy {
+    private final float gaMa;//梯度衰减率
+    private final float gMaxTh;//梯度最大阈值
+    private final boolean auto;//是否自动调整学习率
+    private final MatrixOperation matrixOperation = new MatrixOperation();
+
+    public DymStudy(float gaMa, float gMaxTh, boolean auto) {
+        this.gaMa = gaMa;
+        this.gMaxTh = gMaxTh;
+        this.auto = auto;
+        if (auto) {
+            if (gaMa <= 0 || gaMa >= 1) {
+                throw new IllegalArgumentException("gaMa 取值范围是(0,1)");
+            }
+            if (gMaxTh <= 0) {
+                throw new IllegalArgumentException("gMaxTh 必须比0大");
+            }
+        }
+    }
+
+    private float getErrorNotAuto(float studyRate, float error) {
+        return error * studyRate;
+    }
+
+    public float getNerveStudyError(Map<Integer, Float> dymStudyRate, int key, float g, float studyRate) {
+        if (!auto) {
+            return getErrorNotAuto(studyRate, g);
+        }
+        float gc = gCropping(g);
+        float s = dymStudyRate.get(key);
+        float sNext = gaMa * s + (1 - gaMa) * (float) Math.pow(gc, 2);
+        float myStudyRate = studyRate / (float) Math.sqrt(sNext + 0.00000001f);
+        dymStudyRate.put(key, sNext);
+        return myStudyRate * gc;
+    }
+
+    public float getOneValueError(float studyRate, float g, ConvParameter convParameter) {
+        if (!auto) {
+            return getErrorNotAuto(studyRate, g);
+        }
+        float gc = gCropping(g);
+        float s = convParameter.getStudyRateTh();
+        float sNext = gaMa * s + (1 - gaMa) * (float) Math.pow(gc, 2);
+        convParameter.setStudyRateTh(sNext);
+        float myStudyRate = studyRate / (float) Math.sqrt(sNext + 0.00000001f);
+        return myStudyRate * gc;
+    }
+
+    private float gCropping(float g) {//梯度裁剪
+        if (Math.abs(g) > gMaxTh) {
+            if (g < 0) {
+                return -1 * gMaxTh;
+            } else {
+                return 1 * gMaxTh;
+            }
+        }
+        return g;
+    }
+
+    public float getErrorValueByStudy(float studyRate, List<Float> sList, float g, int t) {
+        if (!auto) {
+            return getErrorNotAuto(studyRate, g);
+        }
+        float s = sList.get(t);
+        g = gCropping(g);
+        float sNext = gaMa * s + (1 - gaMa) * (float) Math.pow(g, 2);
+        float myStudyRate = studyRate / (float) Math.sqrt(sNext + 0.00000001f);
+        sList.set(t, sNext);
+        return g * myStudyRate;
+    }
+
+    public Matrix getErrorMatrixByStudy(float studyRate, Matrix sMatrix, Matrix gMatrix) throws Exception {//获取动态学习率
+        if (!auto) {
+            return matrixOperation.mathMulBySelf(gMatrix, studyRate);
+        }
+        int x = sMatrix.getX();
+        int y = sMatrix.getY();
+        Matrix errorMatrix = new Matrix(x, y);
+        if (x == gMatrix.getX() && y == gMatrix.getY()) {
+            for (int i = 0; i < x; i++) {
+                for (int j = 0; j < y; j++) {
+                    float s = sMatrix.getNumber(i, j);
+                    float g = gCropping(gMatrix.getNumber(i, j));
+                    float sNext = gaMa * s + (1 - gaMa) * (float) Math.pow(g, 2);
+                    sMatrix.setNub(i, j, sNext);
+                    float myStudyRate = studyRate / (float) Math.sqrt(sNext + 0.00000001f);
+                    errorMatrix.setNub(i, j, myStudyRate * g);
+                }
+            }
+        }
+        return errorMatrix;
+    }
+
+}