XGB-25:Callback函数

本文档提供了XGBoost Python包中使用的回调API的基本概述。在XGBoost 1.3中,为Python包设计了一个新的回调接口,它为设计各种扩展提供了灵活性,用于训练。此外,XGBoost还预定义了许多回调函数,用于支持提前停止early stopping、检查点checkpoints等。

使用内置回调函数

默认情况下,XGBoost 中的训练方法具有参数,如 early_stopping_roundsverbose/verbose_eval,当指定这些参数时,训练过程将在内部定义相应的回调函数。例如,当指定了 early_stopping_rounds 时,EarlyStopping 回调将在迭代循环内调用。也可以直接将此回调函数传递给 XGBoost:

python 复制代码
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split

import xgboost as xgb
import numpy as np

X, y = load_breast_cancer(return_X_y=True)
X_train, X_valid, y_train, y_valid = train_test_split(X, y, stratify=y, random_state=94)

D_train = xgb.DMatrix(X_train, y_train)
D_valid = xgb.DMatrix(X_valid, y_valid)

# Define a custom evaluation metric used for early stopping.
def eval_error_metric(predt, dtrain: xgb.DMatrix):
    label = dtrain.get_label()
    r = np.zeros(predt.shape)
    gt = predt > 0.5
    r[gt] = 1 - label[gt]
    le = predt <= 0.5
    r[le] = label[le]
    return 'CustomErr', np.sum(r)

# Specify which dataset and which metric should be used for early stopping.
early_stop = xgb.callback.EarlyStopping(rounds=early_stopping_rounds,
                                        metric_name='CustomErr',
                                        data_name='Train')

booster = xgb.train(
    {'objective': 'binary:logistic',
     'eval_metric': ['error', 'rmse'],
     'tree_method': 'hist'}, D_train,
    evals=[(D_train, 'Train'), (D_valid, 'Valid')],
    feval=eval_error_metric,
    num_boost_round=1000,
    callbacks=[early_stop],
    verbose_eval=False)

dump = booster.get_dump(dump_format='json')
assert len(early_stop.stopping_history['Train']['CustomErr']) == len(dump)

定义自己的回调函数

XGBoost提供了一个回调接口类:TrainingCallback,用户定义的回调应该继承这个类并覆盖相应的方法。在示例中有使用和定义回调函数的工作示例。

python 复制代码
import argparse
import os
import tempfile
from typing import Dict

import numpy as np
from matplotlib import pyplot as plt
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split

import xgboost as xgb


class Plotting(xgb.callback.TrainingCallback):
    """Plot evaluation result during training.  Only for demonstration purpose as it's
    quite slow to draw using matplotlib.

    """

    def __init__(self, rounds: int) -> None:
        self.fig = plt.figure()
        self.ax = self.fig.add_subplot(111)
        self.rounds = rounds
        self.lines: Dict[str, plt.Line2D] = {}
        self.fig.show()
        self.x = np.linspace(0, self.rounds, self.rounds)
        plt.ion()

    def _get_key(self, data: str, metric: str) -> str:
        return f"{data}-{metric}"

    def after_iteration(
        self, model: xgb.Booster, epoch: int, evals_log: Dict[str, dict]
    ) -> bool:
        """Update the plot."""
        if not self.lines:
            for data, metric in evals_log.items():
                for metric_name, log in metric.items():
                    key = self._get_key(data, metric_name)
                    expanded = log + [0] * (self.rounds - len(log))
                    (self.lines[key],) = self.ax.plot(self.x, expanded, label=key)
                    self.ax.legend()
        else:
            # https://pythonspot.com/matplotlib-update-plot/
            for data, metric in evals_log.items():
                for metric_name, log in metric.items():
                    key = self._get_key(data, metric_name)
                    expanded = log + [0] * (self.rounds - len(log))
                    self.lines[key].set_ydata(expanded)
            self.fig.canvas.draw()
        # False to indicate training should not stop.
        return False


def custom_callback() -> None:
    """Demo for defining a custom callback function that plots evaluation result during
    training."""
    X, y = load_breast_cancer(return_X_y=True)
    X_train, X_valid, y_train, y_valid = train_test_split(X, y, random_state=0)

    D_train = xgb.DMatrix(X_train, y_train)
    D_valid = xgb.DMatrix(X_valid, y_valid)

    num_boost_round = 100
    plotting = Plotting(num_boost_round)

    # Pass it to the `callbacks` parameter as a list.
    xgb.train(
        {
            "objective": "binary:logistic",
            "eval_metric": ["error", "rmse"],
            "tree_method": "hist",
            "device": "cuda",
        },
        D_train,
        evals=[(D_train, "Train"), (D_valid, "Valid")],
        num_boost_round=num_boost_round,
        callbacks=[plotting],
    )


def check_point_callback() -> None:
    """Demo for using the checkpoint callback. Custom logic for handling output is
    usually required and users are encouraged to define their own callback for
    checkpointing operations. The builtin one can be used as a starting point.

    """
    # Only for demo, set a larger value (like 100) in practice as checkpointing is quite
    # slow.
    rounds = 2

    def check(as_pickle: bool) -> None:
        for i in range(0, 10, rounds):
            if i == 0:
                continue
            if as_pickle:
                path = os.path.join(tmpdir, "model_" + str(i) + ".pkl")
            else:
                path = os.path.join(
                    tmpdir,
                    f"model_{i}.{xgb.callback.TrainingCheckPoint.default_format}",
                )
            assert os.path.exists(path)

    X, y = load_breast_cancer(return_X_y=True)
    m = xgb.DMatrix(X, y)
    # Check point to a temporary directory for demo
    with tempfile.TemporaryDirectory() as tmpdir:
        # Use callback class from xgboost.callback
        # Feel free to subclass/customize it to suit your need.
        check_point = xgb.callback.TrainingCheckPoint(
            directory=tmpdir, interval=rounds, name="model"
        )
        xgb.train(
            {"objective": "binary:logistic"},
            m,
            num_boost_round=10,
            verbose_eval=False,
            callbacks=[check_point],
        )
        check(False)

        # This version of checkpoint saves everything including parameters and
        # model.  See: doc/tutorials/saving_model.rst
        check_point = xgb.callback.TrainingCheckPoint(
            directory=tmpdir, interval=rounds, as_pickle=True, name="model"
        )
        xgb.train(
            {"objective": "binary:logistic"},
            m,
            num_boost_round=10,
            verbose_eval=False,
            callbacks=[check_point],
        )
        check(True)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--plot", default=1, type=int)
    args = parser.parse_args()

    check_point_callback()

    if args.plot:
        custom_callback()

参考

相关推荐
honeysuckle_luo1 小时前
RandLA-net-pytorch 复现
人工智能·pytorch·python
SunnyDays10112 小时前
Python 高效实现 Excel 与 TXT 文本文件之间的数据转换
python·excel转txt·文本转excel·excel转文本·txt转excel
硬件人某某某3 小时前
python基于卷积神经网络的桥梁裂缝检测系统(django),附可视化界面,源码
python·cnn·django
java1234_小锋3 小时前
PyTorch2 Python深度学习 - 自动微分(Autograd)与梯度优化
开发语言·python·深度学习·pytorch2
java1234_小锋4 小时前
PyTorch2 Python深度学习 - 简介以及入门
python·深度学习·pytorch2
tao3556675 小时前
【Python刷力扣hot100】42. Trapping Rain Water
开发语言·python·leetcode
一只安5 小时前
从零开发AI(不依赖任何模型)
人工智能·python
2501_938782096 小时前
实战 Python NLP:处理 PDF 文档批量提取文本并进行主题建模
python·自然语言处理·pdf
成长痕迹6 小时前
【Python与Matlab数据分析对比】
python·matlab·数据分析
11年老程序猿在线搬砖6 小时前
如何搭建自己的量化交易平台
大数据·人工智能·python·自动交易·量化交易系统