使用GridSearchCV调Keras模型时保存模型报错，需添加.keras扩展名

问题场景 #

使用GridSearchCV自动化搜索Keras模型超参数，执行grid_result.fit(X_train, y_train, verbose=0)时触发报错，提示需添加.keras扩展名，但找不到类内的文件路径设置项。

报错原因 #

TensorFlow/Keras新版本（2.10+）默认采用.keras作为模型保存格式，而Scikeras的KerasClassifier在交叉验证过程中会自动临时保存模型，默认配置未指定新格式，与新版本Keras的保存要求冲突。

解决步骤 #

1. 初始化KerasClassifier时指定保存格式：通过model__save_format='keras'参数，告知Scikeras使用.keras格式保存临时模型；
2. 修复模型创建函数的全局变量依赖：原create_model函数依赖全局X_train，改为动态传入输入维度参数，避免跨作用域问题；
3. 调整最佳模型的保存方式：改用Keras推荐的.keras格式保存最终模型；
4. 补充模型编译步骤：原代码遗漏模型编译，GridSearchCV无法正常训练模型。

修改后的完整代码 #

from sklearn.metrics import make_scorer
from sklearn.metrics import accuracy_score, precision_score, recall_score
from sklearn.model_selection import RandomizedSearchCV
import librosa
import pandas as pd
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from scikeras.wrappers import KerasClassifier 
from sklearn.model_selection import GridSearchCV
import os
import dill as pickle
import csv
import joblib
# Preprocessing
from keras.layers import BatchNormalization
from keras import backend as K
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder, StandardScaler
# Keras
from keras import models
from keras import layers
import keras

def generate_dataset():
    # 生成数据集
    header = 'filename chroma_stft rmse spectral_centroid spectral_bandwidth rolloff zero_crossing_rate'
    for i in range(1, 21):
        header += f' mfcc{i}'
    header += ' label'
    header = header.split()

    file = open('data.csv', 'w', newline='')
    with file:
        writer = csv.writer(file)
        writer.writerow(header)

    # 修改标签类别
    types = ['ukrainian', 'other']

    for t in types:
        for filename in os.listdir(f'./music/{t}/'):
            songname = f'./music/{t}/{filename}'
            y, sr = librosa.load(songname, mono=True)
            chroma_stft = librosa.feature.chroma_stft(y=y, sr=sr)
            rmse=librosa.feature.rms(y=y)[0]
            spec_cent = librosa.feature.spectral_centroid(y=y, sr=sr)
            spec_bw = librosa.feature.spectral_bandwidth(y=y, sr=sr)
            rolloff = librosa.feature.spectral_rolloff(y=y, sr=sr)
            zcr = librosa.feature.zero_crossing_rate(y)
            mfcc = librosa.feature.mfcc(y=y, sr=sr)
            to_append = f'{filename} {np.mean(chroma_stft)} {np.mean(rmse)} {np.mean(spec_cent)} {np.mean(spec_bw)} {np.mean(rolloff)} {np.mean(zcr)}'    
            for e in mfcc:
                to_append += f' {np.mean(e)}'
            # 为每行添加标签
            to_append += f' {t}'
            file = open('data.csv', 'a', newline='', encoding='UTF-8')
            with file:
                writer = csv.writer(file)
                writer.writerow(to_append.split())
    return pd.read_csv('data.csv', encoding='latin-1', on_bad_lines='skip')

def open_dataset():
     # 从CSV读取数据集
    data = pd.read_csv('data.csv', encoding='latin-1', on_bad_lines='skip')
    # 删除不必要的列
    data = data.drop(['filename'], axis=1)
    return data
 
def create_model(input_dim, optimizer='adam', activation='relu'):
    # 动态传入输入维度，避免依赖全局变量
    model = tf.keras.models.Sequential()
    model.add(layers.Dense(1024, activation, input_shape=(input_dim,)))
    model.add(BatchNormalization())
    model.add(layers.Dense(512, activation))
    model.add(BatchNormalization())
    model.add(layers.Dense(256, activation))
    model.add(BatchNormalization())
    model.add(layers.Dense(32, activation))
    model.add(BatchNormalization())
    # 根据标签数量调整输出单元
    model.add(layers.Dense(2, activation='sigmoid'))
    
    # 编译模型，GridSearchCV需要模型已编译
    model.compile(optimizer=optimizer, loss='sparse_categorical_crossentropy', metrics=['accuracy'])
    return model

# 加载并预处理数据
data = pd.read_csv('data.csv', encoding='latin-1', on_bad_lines='skip')
# 删除不必要的列
data = data.drop(['filename'], axis=1)
# 编码标签
encoder = LabelEncoder()
y = encoder.fit_transform(data['label'])
# 标准化特征
scaler = StandardScaler()
X = scaler.fit_transform(np.array(data.iloc[:, :-1], dtype=float))
# 划分训练集和测试集
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

# 初始化KerasClassifier，指定模型保存格式为keras
model = KerasClassifier(
    build_fn=create_model,
    input_dim=X_train.shape[1],  # 传入输入维度参数
    model__save_format='keras'   # 解决扩展名报错的关键参数
)

param_grid = {
    'epochs':[10, 50, 100],
    'batch_size': [32, 64, 128],
    'activation': ['relu', 'softmax', 'sigmoid'], 
    'optimizer': ['adam', 'adamw', 'rmsprop']
}
    
grid_result = GridSearchCV(estimator=model, param_grid=param_grid)
    
grid_result.fit(X_train, y_train, verbose=0)
    
best_model = grid_result.best_estimator_
# 用keras格式保存最佳模型
best_model.model.save('best_model.keras')  
print("最佳得分: %f，使用参数: %s" % (grid_result.best_score_, grid_result.best_params_))

关键修改说明 #

• 在KerasClassifier初始化时添加model__save_format='keras'，强制Scikeras使用新的模型保存格式，解决扩展名报错；
• 重构create_model函数，新增input_dim参数，替换原全局变量X_train.shape[1]，提升代码健壮性；
• 为create_model添加model.compile()步骤，补全模型训练必要流程；
• 将模型保存格式从.h5改为.keras，符合Keras新版本的推荐规范。

内容的提问来源于stack exchange，提问作者Bird