使用Scikeras KerasRegressor遇'model'参数无效错误，求排查

问题场景 #

使用scikeras.wrappers.KerasRegressor结合GridSearchCV进行超参数调优时，触发如下错误：

ValueError: Invalid parameter 'model' for estimator KerasRegressor(
build_fn=<function auto_CreateNeural at 0x71a81c62f600>

相同逻辑在TensorFlow原生的KerasRegressor包装器中可正常运行，相关代码如下：

from scikeras.wrappers import KerasRegressor
import numpy as np
from sklearn import datasets
from sklearn.model_selection import train_test_split, GridSearchCV, KFold
from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import Adam


diabetes = datasets.load_diabetes()
X = diabetes.data
y = diabetes.target

# Split data into training, validation, and test sets
X_train_val, X_test, y_train_val, y_test = train_test_split(X, y, test_size=0.2, random_state=1)
X_train, X_val, y_train, y_val = train_test_split(X_train_val, y_train_val, test_size=0.125, random_state=1)  # 0.125 x 0.8 = 0.1

# Define a model builder function
def auto_CreateNeural(optimizer='adam', activation='relu'):
    regressor = Sequential()
    regressor.add(Dense(10, input_dim=X.shape[1], activation=activation))
    regressor.add(Dense(1))
    regressor.compile(loss='mean_squared_error', optimizer=optimizer)

    return regressor

# Wrap the model with KerasRegressor
regressor = KerasRegressor(build_fn=auto_CreateNeural, verbose=1)

# Define parameters for GridSearchCV
param_grid = {
    'model__optimizer': ['adam', 'sgd'],
    'model__activation': ['relu', 'tanh'],
    'model__batch_size': [4, 8],
    'model__epochs': [10, 20]
}

# Setup cross-validation
kf = KFold(n_splits=3, shuffle=True, random_state=1)
grid = GridSearchCV(estimator=regressor, param_grid=param_grid, cv=kf, scoring='neg_mean_squared_error', return_train_score=True)

# Perform Grid Search
grid_result = grid.fit(X_train_val, y_train_val)

# Evaluate the best model on the test set
best_model = grid.best_estimator_
test_loss = best_model.score(X_test, y_test)

# Output results
print("Best GridSearchCV score: {:.2f}".format(grid_result.best_score_))
print("Best parameters: {}".format(grid_result.best_params_))
print("Test set loss: {:.2f}".format(test_loss))

# Optionally, check how it performs on the validation set if needed
validation_loss = best_model.score(X_val, y_val)
print("Validation set loss: {:.2f}".format(validation_loss))

错误原因 #

Scikeras与TensorFlow原生KerasRegressor的参数传递规则存在差异：

• 原生包装器需要用model__前缀传递模型构建函数的参数，但Scikeras不需要该前缀，直接使用参数名即可
• batch_size和epochs是KerasRegressor自身的拟合参数，不属于模型构建函数的输入，不需要添加任何前缀

修正方案 #

1. 移除参数网格中所有model__前缀，直接使用参数名
2. 确保模型构建函数只接收自身需要的参数（optimizer、activation），batch_size和epochs作为KerasRegressor的参数单独在网格中定义

修正后的完整代码 #

from scikeras.wrappers import KerasRegressor
import numpy as np
from sklearn import datasets
from sklearn.model_selection import train_test_split, GridSearchCV, KFold
from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import Adam


diabetes = datasets.load_diabetes()
X = diabetes.data
y = diabetes.target

# Split data into training, validation, and test sets
X_train_val, X_test, y_train_val, y_test = train_test_split(X, y, test_size=0.2, random_state=1)
X_train, X_val, y_train, y_val = train_test_split(X_train_val, y_train_val, test_size=0.125, random_state=1)  # 0.125 x 0.8 = 0.1

# Define a model builder function
def auto_CreateNeural(optimizer='adam', activation='relu'):
    regressor = Sequential()
    regressor.add(Dense(10, input_dim=X.shape[1], activation=activation))
    regressor.add(Dense(1))
    regressor.compile(loss='mean_squared_error', optimizer=optimizer)

    return regressor

# Wrap the model with KerasRegressor
regressor = KerasRegressor(build_fn=auto_CreateNeural, verbose=1)

# Define parameters for GridSearchCV - 移除model__前缀
param_grid = {
    'optimizer': ['adam', 'sgd'],
    'activation': ['relu', 'tanh'],
    'batch_size': [4, 8],
    'epochs': [10, 20]
}

# Setup cross-validation
kf = KFold(n_splits=3, shuffle=True, random_state=1)
grid = GridSearchCV(estimator=regressor, param_grid=param_grid, cv=kf, scoring='neg_mean_squared_error', return_train_score=True)

# Perform Grid Search
grid_result = grid.fit(X_train_val, y_train_val)

# Evaluate the best model on the test set
best_model = grid.best_estimator_
test_loss = best_model.score(X_test, y_test)

# Output results
print("Best GridSearchCV score: {:.2f}".format(grid_result.best_score_))
print("Best parameters: {}".format(grid_result.best_params_))
print("Test set loss: {:.2f}".format(test_loss))

# Optionally, check how it performs on the validation set if needed
validation_loss = best_model.score(X_val, y_val)
print("Validation set loss: {:.2f}".format(validation_loss))

内容的提问来源于stack exchange，提问作者Deep