加载已保存PyCaret模型调用tune_model触发ValueError报错求助

你通过save_model保存的是包含预处理流程的完整Pipeline对象，而非PyCaret内置的模型实例。当用load_model加载后，tune_model无法从这个Pipeline中识别出对应的内置模型ID（比如xgboost），因此将其判定为自定义模型，触发必须提供custom_grid的报错。

方法一：保存模型名称而非完整模型（推荐） #

不需要保存整个模型文件，直接记录compare_models返回的模型对应的名称，后续调参时直接使用模型名称调用tune_model，PyCaret会自动识别内置模型并使用默认参数网格：

# 对比模型时同时记录模型名称
models = exp.compare_models(include=models_to_ignore, n_select=4)
# 获取模型名称列表（比如['xgboost', 'lightgbm', ...]）
model_names = [exp.models_dict[model.__class__.__name__]['name'] for model in models]

# 后续调参直接循环模型名称
for model_name in model_names:
    print(f'Started tuning {model_name} model')
    tuned_model = tune_model(
        model_name, 
        n_iter=20, 
        optimize='RMSLE',
        search_library="tune-sklearn",
        search_algorithm="optuna",
        early_stopping='asha',
        return_tuner=True,
        return_train_score=True
    )
    save_model(tuned_model, os.path.join(model_dir, f'tuned_{model_name}'))

方法二：从加载的Pipeline中提取模型并指定默认参数网格 #

如果必须加载保存的Pipeline，可以从中提取底层模型，然后获取PyCaret内置的参数网格传给tune_model：

from pycaret.regression import get_config

model_dir = Path('../models')
for model_path in glob(os.path.join(model_dir, '*.pkl')):
    tuned_model_file = os.path.splitext(os.path.basename(model_path))[0]
    print(f'Started tuning {tuned_model_file} model')
    
    # 加载Pipeline
    loaded_pipeline = load_model(os.path.splitext(model_path)[0])
    # 提取底层模型（Pipeline的最后一步是模型）
    model = loaded_pipeline.steps[-1][1]
    
    # 获取对应模型的默认参数网格
    model_id = [k for k, v in exp.models_dict.items() if v['class'] == model.__class__][0]
    custom_grid = get_config(f'{model_id}_grid')
    
    # 调参时传入custom_grid
    tuned_model = tune_model(
        model, 
        n_iter=20, 
        optimize='RMSLE',
        search_library="tune-sklearn",
        search_algorithm="optuna",
        early_stopping='asha',
        return_tuner=True,
        return_train_score=True,
        custom_grid=custom_grid
    )
    
    save_model(tuned_model, os.path.join(model_dir, f'tuned_{tuned_model_file}'))

方法三：复用实验上下文并直接使用模型标识符 #

确保调参时的实验上下文和之前compare_models的实验一致，直接使用模型名称而非加载的Pipeline：

# 确保当前实验是之前的exp
exp = get_current_experiment()

model_dir = Path('../models')
# 假设保存的模型文件名就是模型名称（比如xgboost.pkl）
for model_path in glob(os.path.join(model_dir, '*.pkl')):
    model_name = os.path.splitext(os.path.basename(model_path))[0]
    print(f'Started tuning {model_name} model')
    
    tuned_model = tune_model(
        model_name, 
        n_iter=20, 
        optimize='RMSLE',
        search_library="tune-sklearn",
        search_algorithm="optuna",
        early_stopping='asha',
        return_tuner=True,
        return_train_score=True
    )
    
    save_model(tuned_model, os.path.join(model_dir, f'tuned_{model_name}'))

内容的提问来源于stack exchange，提问作者Bappa