17-4 MLP 应用实战

 1从零实现的神经网络（第17-3课）
 2        ↓
 3MLPClassifier / MLPRegressor API
 4        ↓
 5核心参数详解
 6        ↓
 7分类案例：手写数字识别
 8        ↓
 9回归案例：房价预测
10        ↓
11模型调参与诊断

1from sklearn.neural_network import MLPClassifier, MLPRegressor
2from sklearn.datasets import load_digits, fetch_california_housing
3from sklearn.model_selection import train_test_split
4from sklearn.preprocessing import StandardScaler
5from sklearn.metrics import classification_report, mean_squared_error

1model = NeuralNetwork(n_input=10, n_hidden=32, n_output=1, learning_rate=0.1)
2model.fit(X_train, y_train, epochs=1000)

1model = MLPClassifier(hidden_layer_sizes=(32,), activation='relu',
2                      solver='sgd', learning_rate_init=0.1,
3                      max_iter=1000, random_state=42)
4model.fit(X_train.T, y_train.ravel())  # 注意：sklearn 要求 X: (n_samples, n_features)

1from sklearn.datasets import load_digits
2import matplotlib.pyplot as plt
3
4digits = load_digits()
5X, y = digits.data, digits.target
6
7print(f"数据形状: {X.shape}")  # (1797, 64)
8print(f"标签形状: {y.shape}")   # (1797,)
9print(f"类别数: {len(np.unique(y))}")  # 10

 1from sklearn.model_selection import train_test_split
 2from sklearn.preprocessing import StandardScaler
 3
 4# 划分训练集和测试集
 5X_train, X_test, y_train, y_test = train_test_split(
 6    X, y, test_size=0.2, random_state=42
 7)
 8
 9# 标准化（MLP 对特征尺度敏感）
10scaler = StandardScaler()
11X_train_scaled = scaler.fit_transform(X_train)
12X_test_scaled = scaler.transform(X_test)
13
14print(f"训练集: {X_train_scaled.shape}, 测试集: {X_test_scaled.shape}")

 1from sklearn.neural_network import MLPClassifier
 2from sklearn.metrics import accuracy_score, classification_report, confusion_matrix
 3import seaborn as sns
 4
 5# 实例化模型
 6mlp = MLPClassifier(
 7    hidden_layer_sizes=(64, 32),  # 两个隐藏层：64 → 32
 8    activation='relu',
 9    solver='adam',
10    alpha=0.0001,
11    learning_rate_init=0.001,
12    max_iter=500,
13    early_stopping=True,
14    validation_fraction=0.1,
15    random_state=42,
16    verbose=True
17)
18
19# 训练
20mlp.fit(X_train_scaled, y_train)
21
22# 预测与评估
23y_pred = mlp.predict(X_test_scaled)
24print(f"测试集准确率: {accuracy_score(y_test, y_pred):.4f}")
25print("\n分类报告:")
26print(classification_report(y_test, y_pred))

 1# 提取训练过程中的损失值
 2loss_curve = mlp.loss_curve_
 3
 4plt.figure(figsize=(8, 5))
 5plt.plot(loss_curve)
 6plt.xlabel('迭代次数')
 7plt.ylabel('损失值')
 8plt.title('MLPClassifier 训练损失曲线')
 9plt.grid(True, alpha=0.3)
10plt.show()

1import seaborn as sns
2
3cm = confusion_matrix(y_test, y_pred)
4plt.figure(figsize=(8, 6))
5sns.heatmap(cm, annot=True, fmt='d', cmap='Blues')
6plt.xlabel('预测标签')
7plt.ylabel('真实标签')
8plt.title('混淆矩阵')
9plt.show()

 1# 找出预测错误的样本
 2errors = np.where(y_pred != y_test)[0]
 3print(f"错误样本数: {len(errors)}")
 4
 5# 展示前 5 个错误样本
 6fig, axes = plt.subplots(1, 5, figsize=(10, 3))
 7for i, idx in enumerate(errors[:5]):
 8    axes[i].imshow(X_test[idx].reshape(8, 8), cmap='gray')
 9    axes[i].set_title(f"真实: {y_test[idx]}\n预测: {y_pred[idx]}")
10    axes[i].axis('off')
11plt.tight_layout()
12plt.show()

1from sklearn.datasets import fetch_california_housing
2
3housing = fetch_california_housing()
4X, y = housing.data, housing.target
5
6print(f"数据形状: {X.shape}")  # (20640, 8)
7print(f"特征名称: {housing.feature_names}")
8print(f"目标值范围: [{y.min():.2f}, {y.max():.2f}]")

1X_train, X_test, y_train, y_test = train_test_split(
2    X, y, test_size=0.2, random_state=42
3)
4
5# 标准化
6scaler = StandardScaler()
7X_train_scaled = scaler.fit_transform(X_train)
8X_test_scaled = scaler.transform(X_test)

 1from sklearn.neural_network import MLPRegressor
 2from sklearn.metrics import mean_squared_error, r2_score
 3
 4mlp_reg = MLPRegressor(
 5    hidden_layer_sizes=(64, 32),
 6    activation='relu',
 7    solver='adam',
 8    alpha=0.001,
 9    learning_rate_init=0.001,
10    max_iter=500,
11    early_stopping=True,
12    validation_fraction=0.1,
13    random_state=42,
14    verbose=True
15)
16
17mlp_reg.fit(X_train_scaled, y_train)
18
19# 预测与评估
20y_pred = mlp_reg.predict(X_test_scaled)
21mse = mean_squared_error(y_test, y_pred)
22r2 = r2_score(y_test, y_pred)
23
24print(f"测试集 MSE: {mse:.4f}")
25print(f"测试集 R²: {r2:.4f}")

 1plt.figure(figsize=(8, 5))
 2plt.scatter(y_test, y_pred, alpha=0.3)
 3plt.plot([y_test.min(), y_test.max()], [y_test.min(), y_test.max()],
 4         'r--', linewidth=2, label='理想预测')
 5plt.xlabel('真实值')
 6plt.ylabel('预测值')
 7plt.title('房价预测：真实值 vs 预测值')
 8plt.legend()
 9plt.grid(True, alpha=0.3)
10plt.show()