keras.wrappers模块缺失，求GridSearchCV优化LSTM模型的替代方案

问题原因 #

keras.wrappers.scikit_learn模块在TensorFlow整合Keras后已被迁移，单独的Keras包也不再维护该模块。目前有两种靠谱的替代方案：

方案1：使用TensorFlow自带的Scikit-learn wrapper #

TensorFlow的Keras模块内置了兼容scikit-learn的wrapper，只需调整导入路径即可：

1. 修改所有Keras相关的导入语句，统一使用tensorflow.keras路径
2. 补全缺失的Sequential导入
3. 替换KerasRegressor的导入来源

修改后的完整代码：

from tensorflow.keras.layers import Dense, LSTM, Dropout
from tensorflow.keras import optimizers
from tensorflow.keras.models import Sequential  # 补全导入

from sklearn.model_selection import GridSearchCV
from tensorflow.keras.wrappers.scikit_learn import KerasRegressor

def create_model(unit, dropout_rate, lr ):
    model=Sequential()
    model.add(LSTM(unit,return_sequences=True, input_shape=(1,5)))
    model.add(Dropout(dropout_rate))
    model.add(LSTM(unit))
    model.add(Dropout(dropout_rate))
    model.add(Dense(1))
    adam= optimizers.Adam(learning_rate=lr)  # 新版Adam推荐用learning_rate参数，旧版lr也兼容
    model.compile(optimizer=adam, loss='mean_squared_error')

    return model

my_regressor = KerasRegressor(build_fn=create_model, verbose=2)

grid_param_LSTM = {
    'unit':   [50, 70, 120],
    'batch_size': [12, 24, 48],
    'epochs': [200],
    'lr': [0.001, 0.01, 0.1],
    'dropout_rate':[0.1, 0.2, 0.3]
}

grid_GBR = GridSearchCV(estimator=my_regressor, param_grid = grid_param_LSTM, scoring = 'neg_root_mean_squared_error',  cv = 2)
grid_GBR.fit(X_train, y_train)

print("Best: %f using %s" % (grid_GBR.best_score_, grid_GBR.best_params_))

方案2：使用Scikeras（推荐） #

Scikeras是专门为scikit-learn和Keras/TensorFlow打造的兼容库，维护更活跃，功能更完善，官方也更推荐使用它替代旧的Keras wrapper：

1. 先在Colab中安装Scikeras：

!pip install scikeras

2. 修改导入语句，调整KerasRegressor的参数（新版用model替代build_fn）

修改后的完整代码：

from tensorflow.keras.layers import Dense, LSTM, Dropout
from tensorflow.keras import optimizers
from tensorflow.keras.models import Sequential

from sklearn.model_selection import GridSearchCV
from scikeras.wrappers import KerasRegressor

def create_model(unit, dropout_rate, lr ):
    model=Sequential()
    model.add(LSTM(unit,return_sequences=True, input_shape=(1,5)))
    model.add(Dropout(dropout_rate))
    model.add(LSTM(unit))
    model.add(Dropout(dropout_rate))
    model.add(Dense(1))
    adam= optimizers.Adam(learning_rate=lr)
    model.compile(optimizer=adam, loss='mean_squared_error')

    return model

# 注意：Scikeras推荐用model参数替代build_fn
my_regressor = KerasRegressor(model=create_model, verbose=2)

grid_param_LSTM = {
    'unit':   [50, 70, 120],
    'batch_size': [12, 24, 48],
    'epochs': [200],
    'lr': [0.001, 0.01, 0.1],
    'dropout_rate':[0.1, 0.2, 0.3]
}

grid_GBR = GridSearchCV(estimator=my_regressor, param_grid = grid_param_LSTM, scoring = 'neg_root_mean_squared_error',  cv = 2)
grid_GBR.fit(X_train, y_train)

print("Best: %f using %s" % (grid_GBR.best_score_, grid_GBR.best_params_))

额外提示 #

• 新版TensorFlow的Adam优化器推荐使用learning_rate参数替代旧的lr，不过旧参数也能兼容
• 确保你的Colab环境中TensorFlow版本是2.x，目前Colab默认就是2.x版本

内容的提问来源于stack exchange，提问作者Bouram