解决为GaussianProcessRegressor调参多优化器时的getattr类型错误

我基于GaussianProcessRegressor实现了自定义优化器，用于时间序列目标预测，同时用RandomizedSearchCV进行超参数寻优，运行时出现TypeError: getattr(): attribute name must be string错误，相关代码及报错信息如下：

关键代码片段（HelperFunctions.py） #

模型与搜索参数定义 #

def intermediate_models(kernel):
    dtr_dic = dict(
        ccp_alpha=uniform(loc=0.0, scale=10.0),
        max_features=randint(low=1, high=100),
        max_depth=randint(low=1, high=100),
        criterion=["squared_error", "friedman_mse", "absolute_error", "poisson"],
    )
    optimizer_names = [
        "minimize_wrapper",
        "least_squares_wrapper",
        "differential_evolution_wrapper",
        "basinhopping_wrapper",
        "dual_annealing_wrapper",
    ]
    model_dist_pairs = []
    for optimizer_name in optimizer_names:
        gpr = GPRWithCustomOptimizer(kernel=kernel, optimizer=optimizer_name)

        gpr_dic = dict(
            optimizer=optimizer_names,  # 错误根源
            n_restarts_optimizer=np.arange(0, 20 + 1),
            normalize_y=[False, True],
            copy_X_train=[True, False],
            random_state=np.arange(0, 10 + 1),
        )
        model_dist_pairs.append((gpr, gpr_dic))
    return [(DecisionTreeRegressor(), dtr_dic)] + model_dist_pairs

自定义GPR类 #

class GPRWithCustomOptimizer(GaussianProcessRegressor):
    def __init__(
        self,
        optimizer="minimize",
        initial_theta=None,
        bounds=None,
        random_state=None,
        normalize_y=True,
        n_restarts_optimizer=0,
        copy_X_train=True,
        **kwargs,
    ):
        self.initial_theta = initial_theta
        self.bounds = bounds
        self.custom_optimizers = CustomOptimizers(self, self.initial_theta, self.bounds)
        self.optimizer_func = getattr(self.custom_optimizers, optimizer)  # 触发错误的位置
        super().__init__(
            optimizer=self.optimizer_func,
            random_state=random_state,
            normalize_y=normalize_y,
            n_restarts_optimizer=n_restarts_optimizer,
            copy_X_train=copy_X_train,
            **kwargs,
        )

核心错误信息 #

TypeError: getattr(): attribute name must be string

1. 参数类型不匹配：在intermediate_models函数中，gpr_dic的optimizer字段被赋值为整个列表optimizer_names，而非单个字符串。当RandomizedSearchCV克隆GPR实例时，会将这个列表传入GPRWithCustomOptimizer的构造函数，而getattr要求第二个参数必须是字符串，因此抛出TypeError。
2. 逻辑冗余：循环遍历每个优化器名称创建GPR实例，却给每个实例的搜索参数设置为遍历所有优化器，导致重复搜索，完全没必要。
3. 默认值不匹配：GPRWithCustomOptimizer的默认optimizer值是"minimize"，但自定义优化器类中对应的方法名是minimize_wrapper，默认初始化也会报错。

1. 修正搜索参数的optimizer字段 #

将gpr_dic中的optimizer=optimizer_names改为optimizer=[optimizer_name]，确保每个GPR实例仅搜索当前指定的优化器，同时符合RandomizedSearchCV的参数格式要求。

修改后的intermediate_models函数（二选一即可）：

方案一：每个优化器单独搜索

def intermediate_models(kernel):
    dtr_dic = dict(
        ccp_alpha=uniform(loc=0.0, scale=10.0),
        max_features=randint(low=1, high=100),
        max_depth=randint(low=1, high=100),
        criterion=["squared_error", "friedman_mse", "absolute_error", "poisson"],
    )
    optimizer_names = [
        "minimize_wrapper",
        "least_squares_wrapper",
        "differential_evolution_wrapper",
        "basinhopping_wrapper",
        "dual_annealing_wrapper",
    ]
    model_dist_pairs = []
    for optimizer_name in optimizer_names:
        gpr = GPRWithCustomOptimizer(kernel=kernel, optimizer=optimizer_name)

        gpr_dic = dict(
            optimizer=[optimizer_name],  # 改为当前优化器名称的单元素列表
            n_restarts_optimizer=np.arange(0, 20 + 1),
            normalize_y=[False, True],
            copy_X_train=[True, False],
            random_state=np.arange(0, 10 + 1),
        )
        model_dist_pairs.append((gpr, gpr_dic))
    return [(DecisionTreeRegressor(), dtr_dic)] + model_dist_pairs

方案二：单实例搜索所有优化器（更高效）

def intermediate_models(kernel):
    dtr_dic = dict(
        ccp_alpha=uniform(loc=0.0, scale=10.0),
        max_features=randint(low=1, high=100),
        max_depth=randint(low=1, high=100),
        criterion=["squared_error", "friedman_mse", "absolute_error", "poisson"],
    )
    optimizer_names = [
        "minimize_wrapper",
        "least_squares_wrapper",
        "differential_evolution_wrapper",
        "basinhopping_wrapper",
        "dual_annealing_wrapper",
    ]
    model_dist_pairs = []
    # 仅创建一个GPR实例，一次性搜索所有优化器
    gpr = GPRWithCustomOptimizer(kernel=kernel)
    gpr_dic = dict(
        optimizer=optimizer_names,
        n_restarts_optimizer=np.arange(0, 20 + 1),
        normalize_y=[False, True],
        copy_X_train=[True, False],
        random_state=np.arange(0, 10 + 1),
    )
    model_dist_pairs.append((gpr, gpr_dic))
    return [(DecisionTreeRegressor(), dtr_dic)] + model_dist_pairs

2. 修复自定义GPR类的默认值与鲁棒性 #

class GPRWithCustomOptimizer(GaussianProcessRegressor):
    def __init__(
        self,
        optimizer="minimize_wrapper",  # 修正默认值为存在的方法名
        initial_theta=None,
        bounds=None,
        random_state=None,
        normalize_y=True,
        n_restarts_optimizer=0,
        copy_X_train=True,
        **kwargs,
    ):
        self.initial_theta = initial_theta
        self.bounds = bounds
        # 添加参数校验，避免非字符串参数传入
        if not isinstance(optimizer, str):
            raise ValueError(f"optimizer必须是字符串类型，当前传入类型为{type(optimizer)}")
        self.custom_optimizers = CustomOptimizers(self, self.initial_theta, self.bounds)
        self.optimizer_func = getattr(self.custom_optimizers, optimizer)
        super().__init__(
            optimizer=self.optimizer_func,
            random_state=random_state,
            normalize_y=normalize_y,
            n_restarts_optimizer=n_restarts_optimizer,
            copy_X_train=copy_X_train,
            **kwargs,
        )

内容的提问来源于stack exchange，提问作者Rebel