如何使用AutoSklearn避免根日志记录器被修改？

问题描述 #

调用AutoSklearn的.fit()方法后，项目的根日志记录器格式和行为被意外修改，即使在实例化AutoML对象时配置logging_configuration参数也无法解决问题。

最小复现代码 #

import logging
import autosklearn.regression
AUTOML_CONFIG = dict(time_left_for_this_task=30, per_run_time_limit=10, memory_limit=None, n_jobs=1)
x=[1, 2, 3, 4]
y=[1, 2, 3, 4]

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(name)s %(levelname)s:%(message)s")
logging.info("Initial")
automl = autosklearn.regression.AutoSklearnRegressor(**AUTOML_CONFIG)
logging.info("Create object")
automl.fit(x,y)
logging.info("Finished training")
logging.warning("Finished training")

预期输出 #

2022-08-10 14:06:13,635 root INFO:Initial
2022-08-10 14:06:13,635 root INFO:Create object
2022-08-10 14:06:13,635 root INFO:Finished training
2022-08-10 14:06:13,635 root WARNING:Finished training

实际输出 #

2022-08-10 14:06:13,635 root INFO:Initial
2022-08-10 14:06:13,635 root INFO:Create object
[WARNING] [2022-08-10 14:06:41,523:root] Finished training

解决方法 #

方法1：使用自定义项目日志器（推荐） #

避免直接使用根日志器，创建专属的项目日志记录器，关闭日志传递到根日志器的功能，彻底隔离AutoSklearn的日志修改。

import logging
import autosklearn.regression

# 创建并配置项目专属日志器
project_logger = logging.getLogger("my_project")
project_logger.setLevel(logging.INFO)

# 设置日志格式
formatter = logging.Formatter("%(asctime)s %(name)s %(levelname)s:%(message)s")
stream_handler = logging.StreamHandler()
stream_handler.setFormatter(formatter)
project_logger.addHandler(stream_handler)

# 禁止日志传递到根日志器，避免被AutoSklearn修改
project_logger.propagate = False

AUTOML_CONFIG = dict(time_left_for_this_task=30, per_run_time_limit=10, memory_limit=None, n_jobs=1)
x=[1, 2, 3, 4]
y=[1, 2, 3, 4]

project_logger.info("Initial")
automl = autosklearn.regression.AutoSklearnRegressor(**AUTOML_CONFIG)
project_logger.info("Create object")
automl.fit(x,y)
project_logger.info("Finished training")
project_logger.warning("Finished training")

方法2：保存并恢复根日志器配置 #

在调用.fit()前后，手动保存根日志器的原有配置（处理器、日志级别、格式），训练完成后恢复这些配置，抵消AutoSklearn的修改。

import logging
import autosklearn.regression

AUTOML_CONFIG = dict(time_left_for_this_task=30, per_run_time_limit=10, memory_limit=None, n_jobs=1)
x=[1, 2, 3, 4]
y=[1, 2, 3, 4]

# 初始化根日志器
logging.basicConfig(level=logging.INFO, format="%(asctime)s %(name)s %(levelname)s:%(message)s")
logging.info("Initial")
automl = autosklearn.regression.AutoSklearnRegressor(**AUTOML_CONFIG)
logging.info("Create object")

# 保存根日志器的原始配置
root_logger = logging.getLogger()
original_handlers = root_logger.handlers.copy()
original_level = root_logger.level
original_formatters = [handler.formatter for handler in original_handlers]

# 执行训练
automl.fit(x, y)

# 恢复根日志器的原始配置
root_logger.handlers = original_handlers
root_logger.setLevel(original_level)
for handler, formatter in zip(root_logger.handlers, original_formatters):
    handler.setFormatter(formatter)

# 输出日志验证
logging.info("Finished training")
logging.warning("Finished training")

方法3：配置AutoSklearn使用独立日志器 #

通过logging_configuration参数指定AutoSklearn使用自己的日志命名空间，避免影响根日志器。注意部分版本的AutoSklearn可能存在内部逻辑仍修改根日志器的情况，需结合前两种方法使用。

import logging
import autosklearn.regression

# 配置AutoSklearn的日志隔离
automl_log_config = {
    "version": 1,
    "disable_existing_loggers": False,
    "loggers": {
        "autosklearn": {
            "level": "INFO",
            "handlers": ["console"],
            "propagate": False
        }
    },
    "handlers": {
        "console": {
            "class": "logging.StreamHandler",
            "formatter": "autosklearn_formatter"
        }
    },
    "formatters": {
        "autosklearn_formatter": {
            "format": "[%(levelname)s] [%(asctime)s:%(name)s] %(message)s"
        }
    }
}

AUTOML_CONFIG = dict(
    time_left_for_this_task=30,
    per_run_time_limit=10,
    memory_limit=None,
    n_jobs=1,
    logging_configuration=automl_log_config
)

# 初始化项目根日志器
logging.basicConfig(level=logging.INFO, format="%(asctime)s %(name)s %(levelname)s:%(message)s")

x=[1, 2, 3, 4]
y=[1, 2, 3, 4]

logging.info("Initial")
automl = autosklearn.regression.AutoSklearnRegressor(**AUTOML_CONFIG)
logging.info("Create object")
automl.fit(x,y)
logging.info("Finished training")
logging.warning("Finished training")

内容的提问来源于stack exchange，提问作者Sergei I