从keras.wrappers.sci-learn迁移到Scikeras的技术求助

关键迁移步骤 #

• 安装Scikeras：pip install scikeras
• 替换导入：使用scikeras.wrappers.KerasRegressor替代旧版Keras wrapper
• 参数命名规则：GridSearchCV的参数网格需给模型参数加上model__前缀
• 模型定义优化：避免硬编码训练集形状，让模型更通用

修改后的完整代码 #

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense
from tensorflow.keras.optimizers import SGD
from tensorflow.keras.regularizers import L2
from sklearn.model_selection import TimeSeriesSplit, GridSearchCV
from sklearn.metrics import r2_score
import numpy as np
import matplotlib.pyplot as plt

# 模型构建函数：参数保持与原定义一致
def create_model(lstm_1, lstm_2, dense, input_shape):
    model = Sequential()
    model.add(LSTM(lstm_1, input_shape=input_shape, return_sequences=True))
    model.add(LSTM(lstm_2))
    model.add(Dense(dense, activation='linear'))
    model.add(Dense(1, activation='linear'))
    model.compile(optimizer='adam', loss='mean_squared_error')
    return model

# 定义参数网格：给每个模型参数加上model__前缀
param_grid = {
    'model__lstm_1': [32, 64, 128],
    'model__lstm_2': [32, 64, 128],
    'model__dense': [16, 32, 64]
}

# 时间序列交叉验证拆分
tscv = TimeSeriesSplit(n_splits=5)

# 用Scikeras的KerasRegressor包装模型
# 传入固定参数input_shape，可调参数通过GridSearchCV传递
estimator = tf.keras.wrappers.scikit_learn.KerasRegressor(
    build_fn=create_model,
    input_shape=(X_train.shape[1], X_train.shape[2]),
    verbose=0
)

# 网格搜索：使用包装后的estimator
grid = GridSearchCV(estimator=estimator, param_grid=param_grid, cv=tscv, n_jobs=-1)
grid_result = grid.fit(X_train, y_train)

print("最优参数: ", grid_result.best_params_)

# 获取最优模型（与原代码逻辑一致）
best_model = grid_result.best_estimator_.model

# 预测与评估
NN_predictions = best_model.predict(X_test)
NN_mse = np.mean((y_test - NN_predictions.flatten()) ** 2)
NN_rmse = np.sqrt(NN_mse)
print("神经网络模型RMSE: ", NN_rmse)

NN_r_squared = r2_score(y_test, NN_predictions)
n = len(y_test)
p = 8
NN_adjusted_r_squared = 1 - (1 - NN_r_squared) * (n - 1) / (n - p - 1)
print('OOS R²: ', NN_r_squared)
print('OOS 调整后R²: ', NN_adjusted_r_squared)

# 可视化
plt.figure(figsize=(12, 6))
plt.plot(y_test, label='隐含波动率')
plt.plot(NN_predictions, label='神经网络预测')
plt.xlabel('日期索引')
plt.ylabel('隐含波动率')
plt.title('实际值 vs 预测值')
plt.legend()
plt.show()

# 带正则化的模型训练
LSTM1 = grid_result.best_params_['model__lstm_1']
LSTM2 = grid_result.best_params_['model__lstm_2']
DENSE = grid_result.best_params_['model__dense']
K = 1  # 原代码中K未定义，这里补充为1（与输出层一致）
N_epochs = 50  # 原代码中未定义，需根据实际情况设置
N_batch_size = 32  # 原代码中未定义，需根据实际情况设置

model_REG = Sequential([
    LSTM(LSTM1, input_shape=(X_train.shape[1], X_train.shape[2]), return_sequences=True, kernel_regularizer=L2(0.001)),
    LSTM(LSTM2),
    Dense(DENSE, activation='linear'),
    Dense(K, activation='linear')
])

model_REG.compile(optimizer='adam', loss='mean_squared_error')
history_REG = model_REG.fit(X_train, y_train, epochs=N_epochs, batch_size=N_batch_size, validation_data=(X_test, y_test))

# 正则化模型评估
NN_REG_predictions = model_REG.predict(X_test)
NN_REG_mse = np.mean((y_test - NN_REG_predictions.flatten()) ** 2)
NN_REG_rmse = np.sqrt(NN_REG_mse)
print("带正则化的神经网络模型RMSE: ", NN_REG_rmse)

NN_r_squared = r2_score(y_test, NN_REG_predictions)
NN_adjusted_r_squared = 1 - (1 - NN_r_squared) * (n - 1) / (n - p - 1)
print('OOS R²: ', NN_r_squared)
print('OOS 调整后R²: ', NN_adjusted_r_squared)

# 可视化
plt.figure(figsize=(12, 6))
plt.plot(y_test, label='隐含波动率')
plt.plot(NN_REG_predictions, label='正则化神经网络预测')
plt.xlabel('日期索引')
plt.ylabel('隐含波动率')
plt.title('实际值 vs 预测值（正则化模型）')
plt.legend()
plt.show()

关键说明 #

1. 参数前缀：Scikeras通过model__前缀区分模型参数与wrapper自身参数（如epochs、batch_size），因此参数网格必须添加该前缀
2. 模型包装：使用KerasRegressor（回归任务）或KerasClassifier（分类任务）包装模型构建函数，固定参数（如input_shape）直接传入wrapper，可调参数通过GridSearchCV传递
3. 未定义变量处理：原代码中K、N_epochs、N_batch_size未定义，已补充默认值，需根据你的数据集调整
4. 冗余代码清理：原代码中TwoLayerFeedForward类与当前任务无关，已移除

内容的提问来源于stack exchange，提问作者Guesse