【机器学习】太香啦!只需一行Python代码就可以自动完成模型训练!

共 28099字,需浏览 57分钟

 ·

2021-05-24 16:34

自动化机器学习(Auto-ML)是指数据科学模型开发的管道组件自动化。AutoML 减少了数据科学家的工作量并加快了工作流程。AutoML 可用于自动化各种管道组件,包括数据理解,EDA,数据处理,模型训练,超参数调整等。

对于端到端机器学习项目,每个组件的复杂性取决于项目。我们知道市面上有很多的 AutoML 开源库可加快开发的速度。在本文中,我将分享一个非常棒的python工具库「LazyPredict」

什么是LazyPredict?

LazyPredict是一个开源Python库,可自动执行模型训练管道并加快工作流程。LazyPredict可以为分类数据集训练约30个分类模型,为回归数据集训练约40个回归模型。

LazyPredict将返回经过训练的模型以及其性能指标,而无需编写太多代码。可以轻松比较每个模型的性能指标,并调整最佳模型以进一步提高性能。

安装

可以使用以下方法从PyPl库中安装LazyPredict:

pip install lazypredict

安装后,可以导入库以执行分类和回归模型的自动训练。

from lazypredict.Supervised import LazyRegressor, LazyClassifier

用法

LazyPredict 同时支持分类和回归问题,因此我将利用案例说明:波士顿住房(回归)和泰坦尼克号(分类)数据集用于LazyPredict库的演示。

分类任务

LazyPredict 的用法非常直观,类似于scikit-learn。首先为分类任务创建一个估计器 LazyClassifier 的实例,可以通过自定义指标进行评估,默认情况下,每个模型都将根据准确性,ROC、AUC得分, F1-score进行评估。

在进行 lazypredict 模型训练之前,必须先读取数据集并进行处理,以使其适合训练。在进行特征工程并将数据拆分为训练测试数据之后,我们可以使用 LazyPredict 进行模型训练。

# LazyClassifier Instance and fiting data
cls= LazyClassifier(ignore_warnings=False, custom_metric=None)
models, predictions = cls.fit(X_train, X_test, y_train, y_test)

回归任务

与分类模型训练相似,LazyPredict附带了针对回归数据集的自动模型训练。该实现类似于分类任务,只是实例LazyRegressor有所更改。

reg = LazyRegressor(ignore_warnings=False, custom_metric=None)
models, predictions = reg.fit(X_train, X_test, y_train, y_test)

观察以上性能指标,AdaBoost分类器是分类任务的最佳表现模型,而GradientBoostingRegressor模型是回归任务的最佳表现模型。

完整版案例

分类

from lazypredict.Supervised import LazyClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split

data = load_breast_cancer()
X = data.data
y= data.target

X_train, X_test, y_train, y_test = train_test_split(X, y,test_size=.5,random_state =123)

clf = LazyClassifier(verbose=0,ignore_warnings=True, custom_metric=None)
models,predictions = clf.fit(X_train, X_test, y_train, y_test)

print(models)


| Model                          |   Accuracy |   Balanced Accuracy |   ROC AUC |   F1 Score |   Time Taken |
|:-------------------------------|-----------:|--------------------:|----------:|-----------:|-------------:|
| LinearSVC                      |   0.989474 |            0.987544 |  0.987544 |   0.989462 |    0.0150008 |
| SGDClassifier                  |   0.989474 |            0.987544 |  0.987544 |   0.989462 |    0.0109992 |
| MLPClassifier                  |   0.985965 |            0.986904 |  0.986904 |   0.985994 |    0.426     |
| Perceptron                     |   0.985965 |            0.984797 |  0.984797 |   0.985965 |    0.0120046 |
| LogisticRegression             |   0.985965 |            0.98269  |  0.98269  |   0.985934 |    0.0200036 |
| LogisticRegressionCV           |   0.985965 |            0.98269  |  0.98269  |   0.985934 |    0.262997  |
| SVC                            |   0.982456 |            0.979942 |  0.979942 |   0.982437 |    0.0140011 |
| CalibratedClassifierCV         |   0.982456 |            0.975728 |  0.975728 |   0.982357 |    0.0350015 |
| PassiveAggressiveClassifier    |   0.975439 |            0.974448 |  0.974448 |   0.975464 |    0.0130005 |
| LabelPropagation               |   0.975439 |            0.974448 |  0.974448 |   0.975464 |    0.0429988 |
| LabelSpreading                 |   0.975439 |            0.974448 |  0.974448 |   0.975464 |    0.0310006 |
| RandomForestClassifier         |   0.97193  |            0.969594 |  0.969594 |   0.97193  |    0.033     |
| GradientBoostingClassifier     |   0.97193  |            0.967486 |  0.967486 |   0.971869 |    0.166998  |
| QuadraticDiscriminantAnalysis  |   0.964912 |            0.966206 |  0.966206 |   0.965052 |    0.0119994 |
| HistGradientBoostingClassifier |   0.968421 |            0.964739 |  0.964739 |   0.968387 |    0.682003  |
| RidgeClassifierCV              |   0.97193  |            0.963272 |  0.963272 |   0.971736 |    0.0130029 |
| RidgeClassifier                |   0.968421 |            0.960525 |  0.960525 |   0.968242 |    0.0119977 |
| AdaBoostClassifier             |   0.961404 |            0.959245 |  0.959245 |   0.961444 |    0.204998  |
| ExtraTreesClassifier           |   0.961404 |            0.957138 |  0.957138 |   0.961362 |    0.0270066 |
| KNeighborsClassifier           |   0.961404 |            0.95503  |  0.95503  |   0.961276 |    0.0560005 |
| BaggingClassifier              |   0.947368 |            0.954577 |  0.954577 |   0.947882 |    0.0559971 |
| BernoulliNB                    |   0.950877 |            0.951003 |  0.951003 |   0.951072 |    0.0169988 |
| LinearDiscriminantAnalysis     |   0.961404 |            0.950816 |  0.950816 |   0.961089 |    0.0199995 |
| GaussianNB                     |   0.954386 |            0.949536 |  0.949536 |   0.954337 |    0.0139935 |
| NuSVC                          |   0.954386 |            0.943215 |  0.943215 |   0.954014 |    0.019989  |
| DecisionTreeClassifier         |   0.936842 |            0.933693 |  0.933693 |   0.936971 |    0.0170023 |
| NearestCentroid                |   0.947368 |            0.933506 |  0.933506 |   0.946801 |    0.0160074 |
| ExtraTreeClassifier            |   0.922807 |            0.912168 |  0.912168 |   0.922462 |    0.0109999 |
| CheckingClassifier             |   0.361404 |            0.5      |  0.5      |   0.191879 |    0.0170043 |
| DummyClassifier                |   0.512281 |            0.489598 |  0.489598 |   0.518924 |    0.0119965 |

回归

from lazypredict.Supervised import LazyRegressor
from sklearn import datasets
from sklearn.utils import shuffle
import numpy as np

boston = datasets.load_boston()
X, y = shuffle(boston.data, boston.target, random_state=13)
X = X.astype(np.float32)

offset = int(X.shape[0] * 0.9)

X_train, y_train = X[:offset], y[:offset]
X_test, y_test = X[offset:], y[offset:]

reg = LazyRegressor(verbose=0, ignore_warnings=False, custom_metric=None)
models, predictions = reg.fit(X_train, X_test, y_train, y_test)

print(models)


| Model                         | Adjusted R-Squared | R-Squared |  RMSE | Time Taken |
|:------------------------------|-------------------:|----------:|------:|-----------:|
| SVR                           |               0.83 |      0.88 |  2.62 |       0.01 |
| BaggingRegressor              |               0.83 |      0.88 |  2.63 |       0.03 |
| NuSVR                         |               0.82 |      0.86 |  2.76 |       0.03 |
| RandomForestRegressor         |               0.81 |      0.86 |  2.78 |       0.21 |
| XGBRegressor                  |               0.81 |      0.86 |  2.79 |       0.06 |
| GradientBoostingRegressor     |               0.81 |      0.86 |  2.84 |       0.11 |
| ExtraTreesRegressor           |               0.79 |      0.84 |  2.98 |       0.12 |
| AdaBoostRegressor             |               0.78 |      0.83 |  3.04 |       0.07 |
| HistGradientBoostingRegressor |               0.77 |      0.83 |  3.06 |       0.17 |
| PoissonRegressor              |               0.77 |      0.83 |  3.11 |       0.01 |
| LGBMRegressor                 |               0.77 |      0.83 |  3.11 |       0.07 |
| KNeighborsRegressor           |               0.77 |      0.83 |  3.12 |       0.01 |
| DecisionTreeRegressor         |               0.65 |      0.74 |  3.79 |       0.01 |
| MLPRegressor                  |               0.65 |      0.74 |  3.80 |       1.63 |
| HuberRegressor                |               0.64 |      0.74 |  3.84 |       0.01 |
| GammaRegressor                |               0.64 |      0.73 |  3.88 |       0.01 |
| LinearSVR                     |               0.62 |      0.72 |  3.96 |       0.01 |
| RidgeCV                       |               0.62 |      0.72 |  3.97 |       0.01 |
| BayesianRidge                 |               0.62 |      0.72 |  3.97 |       0.01 |
| Ridge                         |               0.62 |      0.72 |  3.97 |       0.01 |
| TransformedTargetRegressor    |               0.62 |      0.72 |  3.97 |       0.01 |
| LinearRegression              |               0.62 |      0.72 |  3.97 |       0.01 |
| ElasticNetCV                  |               0.62 |      0.72 |  3.98 |       0.04 |
| LassoCV                       |               0.62 |      0.72 |  3.98 |       0.06 |
| LassoLarsIC                   |               0.62 |      0.72 |  3.98 |       0.01 |
| LassoLarsCV                   |               0.62 |      0.72 |  3.98 |       0.02 |
| Lars                          |               0.61 |      0.72 |  3.99 |       0.01 |
| LarsCV                        |               0.61 |      0.71 |  4.02 |       0.04 |
| SGDRegressor                  |               0.60 |      0.70 |  4.07 |       0.01 |
| TweedieRegressor              |               0.59 |      0.70 |  4.12 |       0.01 |
| GeneralizedLinearRegressor    |               0.59 |      0.70 |  4.12 |       0.01 |
| ElasticNet                    |               0.58 |      0.69 |  4.16 |       0.01 |
| Lasso                         |               0.54 |      0.66 |  4.35 |       0.02 |
| RANSACRegressor               |               0.53 |      0.65 |  4.41 |       0.04 |
| OrthogonalMatchingPursuitCV   |               0.45 |      0.59 |  4.78 |       0.02 |
| PassiveAggressiveRegressor    |               0.37 |      0.54 |  5.09 |       0.01 |
| GaussianProcessRegressor      |               0.23 |      0.43 |  5.65 |       0.03 |
| OrthogonalMatchingPursuit     |               0.16 |      0.38 |  5.89 |       0.01 |
| ExtraTreeRegressor            |               0.08 |      0.32 |  6.17 |       0.01 |
| DummyRegressor                |              -0.38 |     -0.02 |  7.56 |       0.01 |
| LassoLars                     |              -0.38 |     -0.02 |  7.56 |       0.01 |
| KernelRidge                   |             -11.50 |     -8.25 | 22.74 |       0.01 |

结论

在本文中,我们讨论了LazyPredict库的实现,该库可以在几行Python代码中训练大约70个分类和回归模型。这是一个非常方便的工具,因为它提供了模型执行的总体情况,并且可以比较每个模型的性能。

每个模型都使用其默认参数进行训练,因为它不执行超参数调整。选择性能最佳的模型后,开发人员可以调整模型以进一步提高性能。

往期精彩回顾





本站qq群851320808,加入微信群请扫码:

浏览 80
点赞
评论
收藏
分享

手机扫一扫分享

分享
举报
评论
图片
表情
推荐
点赞
评论
收藏
分享

手机扫一扫分享

分享
举报