diff --git a/function.py b/function.py
index daa5782..c041cc6 100644
--- a/function.py
+++ b/function.py
@@ -28,7 +28,7 @@ def sigmoid(z):
     return 1.0 / (1 + np.exp(-z))
 
 # 这里写自己的代码实现
-def train_logreg(X, y, learning_rate=0.01, iterations=1000):
+def train_logreg(X, y, learning_rate=0.01, iterations=1000, batch_size=64):
     """
     对率回归训练函数。
 
@@ -42,15 +42,15 @@ def train_logreg(X, y, learning_rate=0.01, iterations=1000):
     """
     m, n = X.shape
     weights = np.zeros(n)
-    # 用于Adagrad的累积平方梯度
-    r = np.zeros(n)
-    epsilon = 1e-8
     for i in range(iterations):
-        z = np.dot(X, weights)
-        h = sigmoid(z)
-        gradient = np.dot(X.T, (h - y)) / m
-        r += gradient ** 2
-        weights -= learning_rate * gradient / (np.sqrt(r) + epsilon)
+        for j in range(0, m, batch_size):
+            end = min(j + batch_size, m)
+            X_batch = X[j:end]
+            y_batch = y[j:end]
+            z = np.dot(X_batch, weights)
+            h = sigmoid(z)
+            gradient = np.dot(X_batch.T, (h - y_batch)) / len(X_batch)
+            weights -= learning_rate * gradient
     return weights
 
 def predict_logreg(inX, weights):