更新 function.py

function.py

@@ -34,33 +34,41 @@ def load_data(filepath):
     with open(filepath, 'r') as file:
         for line in file:
             fields = line.strip().split()
-            if '?' in fields:  # 处理缺失值
-                fields = [0 if f == '?' else float(f) for f in fields] 
+            fields = [0 if f == '?' else float(f) for f in fields]  # 缺失值填充
             features = np.array(fields[:-1], dtype=float)
-            label = int(fields[-1])  # 确保标签未合并
+            label = int(fields[-1])
             data.append(features)
             labels.append(label)
-    return np.array(data), np.array(labels)
+    labels = np.array(labels)
 
-def train_svm(X, y, learning_rate=0.01, iterations=1000):
+    # 检查标签是否为三类
+    unique_labels = np.unique(labels)
+    if len(unique_labels) != 3:
+        raise ValueError("标签数量不为3，请确保标签未合并！")
+    
+    return np.array(data), labels
+
+def train_svm(X, y, learning_rate=0.01, iterations=1000, regularization_strength=0.01):
     n_samples, n_features = X.shape
     n_classes = len(np.unique(y))
     weights = np.zeros((n_classes, n_features))
     
+    # 将标签转化为多类别（one-vs-rest）形式
+    y_one_vs_rest = np.array([[1 if y[i] == c else -1 for c in range(n_classes)] for i in range(n_samples)])
+
     for _ in range(iterations):
-        for idx, x_i in enumerate(X):
-            for c in range(n_classes):
-                y_i = 1 if y[idx] == c else -1
-                if y_i * (np.dot(weights[c], x_i)) < 1:
-                    weights[c] += learning_rate * (y_i * x_i - (2 * 1/iterations) * weights[c])
-                else:
-                    weights[c] -= learning_rate * (2 * 1/iterations) * weights[c]
+        # 计算预测值
+        margins = y_one_vs_rest * (X @ weights.T)
+        mask = margins < 1
+        grad = -np.einsum('ij,ik->kj', y_one_vs_rest * mask, X) / n_samples
+        weights -= learning_rate * (grad + regularization_strength * weights)
     
     return weights
 
-def predict_svm(inX, weights):
-    scores = [np.dot(w, inX) for w in weights]
-    return np.argmax(scores)
+def predict_svm(X, weights):
+    scores = X @ weights.T
+    return np.argmax(scores, axis=1)
+