sklearn是利用python進(jìn)行機(jī)器學(xué)習(xí)中一個(gè)非常全面和好用的第三方庫，用過的都說好。今天主要記錄一下sklearn中關(guān)于交叉驗(yàn)證的各種用法，主要是對(duì)sklearn官方文檔 Cross-validation: evaluating estimator performance進(jìn)行講解，英文水平好的建議讀官方文檔，里面的知識(shí)點(diǎn)很詳細(xì)。
先導(dǎo)入需要的庫及數(shù)據(jù)集
In [1]: import numpy as np
In [2]: from sklearn.model_selection import train_test_split
In [3]: from sklearn.datasets import load_iris
In [4]: from sklearn import svm
In [5]: iris = load_iris()
In [6]: iris.data.shape, iris.target.shape
Out[6]: ((150, 4), (150,))
1.train_test_split
對(duì)數(shù)據(jù)集進(jìn)行快速打亂（分為訓(xùn)練集和測試集）
這里相當(dāng)于對(duì)數(shù)據(jù)集進(jìn)行了shuffle后按照給定的test_size 進(jìn)行數(shù)據(jù)集劃分。
In [7]: X_train, X_test, y_train, y_test = train_test_split(
...: iris.data, iris.target, test_size=.4, random_state=0)
#這里是按照6:4對(duì)訓(xùn)練集測試集進(jìn)行劃分
In [8]: X_train.shape, y_train.shape
Out[8]: ((90, 4), (90,))
In [9]: X_test.shape, y_test.shape
Out[9]: ((60, 4), (60,))
In [10]: iris.data[:5]
Out[10]:
array([[ 5.1, 3.5, 1.4, 0.2],
[ 4.9, 3. , 1.4, 0.2],
[ 4.7, 3.2, 1.3, 0.2],
[ 4.6, 3.1, 1.5, 0.2],
[ 5. , 3.6, 1.4, 0.2]])
In [11]: X_train[:5]
Out[11]:
array([[ 6. , 3.4, 4.5, 1.6],
[ 4.8, 3.1, 1.6, 0.2],
[ 5.8, 2.7, 5.1, 1.9],
[ 5.6, 2.7, 4.2, 1.3],
[ 5.6, 2.9, 3.6, 1.3]])
In [12]: clf = svm.SVC(kernel='linear', C=1).fit(X_train, y_train)
In [13]: clf.score(X_test, y_test)
Out[13]: 0.96666666666666667
2.cross_val_score
對(duì)數(shù)據(jù)集進(jìn)行指定次數(shù)的交叉驗(yàn)證并為每次驗(yàn)證效果評(píng)測
其中，score 默認(rèn)是以 scoring='f1_macro'進(jìn)行評(píng)測的，余外針對(duì)分類或回歸還有：
這需要from sklearn import metrics ,通過在cross_val_score 指定參數(shù)來設(shè)定評(píng)測標(biāo)準(zhǔn)；
當(dāng)cv 指定為int 類型時(shí)，默認(rèn)使用KFold 或StratifiedKFold 進(jìn)行數(shù)據(jù)集打亂，下面會(huì)對(duì)KFold 和StratifiedKFold 進(jìn)行介紹。
In [15]: from sklearn.model_selection import cross_val_score
In [16]: clf = svm.SVC(kernel='linear', C=1)
In [17]: scores = cross_val_score(clf, iris.data, iris.target, cv=5)
In [18]: scores
Out[18]: array([ 0.96666667, 1. , 0.96666667, 0.96666667, 1. ])
In [19]: scores.mean()
Out[19]: 0.98000000000000009
除使用默認(rèn)交叉驗(yàn)證方式外，可以對(duì)交叉驗(yàn)證方式進(jìn)行指定，如驗(yàn)證次數(shù)，訓(xùn)練集測試集劃分比例等
In [20]: from sklearn.model_selection import ShuffleSplit
In [21]: n_samples = iris.data.shape[0]
In [22]: cv = ShuffleSplit(n_splits=3, test_size=.3, random_state=0)
In [23]: cross_val_score(clf, iris.data, iris.target, cv=cv)
Out[23]: array([ 0.97777778, 0.97777778, 1. ])
在cross_val_score 中同樣可使用pipeline 進(jìn)行流水線操作
In [24]: from sklearn import preprocessing
In [25]: from sklearn.pipeline import make_pipeline
In [26]: clf = make_pipeline(preprocessing.StandardScaler(), svm.SVC(C=1))
In [27]: cross_val_score(clf, iris.data, iris.target, cv=cv)
Out[27]: array([ 0.97777778, 0.93333333, 0.95555556])
3.cross_val_predict
cross_val_predict 與cross_val_score 很相像，不過不同于返回的是評(píng)測效果，cross_val_predict 返回的是estimator 的分類結(jié)果（或回歸值），這個(gè)對(duì)于后期模型的改善很重要，可以通過該預(yù)測輸出對(duì)比實(shí)際目標(biāo)值，準(zhǔn)確定位到預(yù)測出錯(cuò)的地方，為我們參數(shù)優(yōu)化及問題排查十分的重要。
In [28]: from sklearn.model_selection import cross_val_predict
In [29]: from sklearn import metrics
In [30]: predicted = cross_val_predict(clf, iris.data, iris.target, cv=10)
In [31]: predicted
Out[31]:
array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 2, 2,
2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2,
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2])
In [32]: metrics.accuracy_score(iris.target, predicted)
Out[32]: 0.96666666666666667
4.KFold
K折交叉驗(yàn)證，這是將數(shù)據(jù)集分成K份的官方給定方案，所謂K折就是將數(shù)據(jù)集通過K次分割，使得所有數(shù)據(jù)既在訓(xùn)練集出現(xiàn)過，又在測試集出現(xiàn)過，當(dāng)然，每次分割中不會(huì)有重疊。相當(dāng)于無放回抽樣。
In [33]: from sklearn.model_selection import KFold
In [34]: X = ['a','b','c','d']
In [35]: kf = KFold(n_splits=2)
In [36]: for train, test in kf.split(X):
...: print train, test
...: print np.array(X)[train], np.array(X)[test]
...: print '\n'
...:
[2 3] [0 1]
['c' 'd'] ['a' 'b']
[0 1] [2 3]
['a' 'b'] ['c' 'd']
5.LeaveOneOut
LeaveOneOut 其實(shí)就是KFold 的一個(gè)特例，因?yàn)槭褂么螖?shù)比較多，因此獨(dú)立的定義出來，完全可以通過KFold 實(shí)現(xiàn)。
In [37]: from sklearn.model_selection import LeaveOneOut
In [38]: X = [1,2,3,4]
In [39]: loo = LeaveOneOut()
In [41]: for train, test in loo.split(X):
...: print train, test
...:
[1 2 3] [0]
[0 2 3] [1]
[0 1 3] [2]
[0 1 2] [3]
#使用KFold實(shí)現(xiàn)LeaveOneOtut
In [42]: kf = KFold(n_splits=len(X))
In [43]: for train, test in kf.split(X):
...: print train, test
...:
[1 2 3] [0]
[0 2 3] [1]
[0 1 3] [2]
[0 1 2] [3]
6.LeavePOut
這個(gè)也是KFold 的一個(gè)特例，用KFold 實(shí)現(xiàn)起來稍麻煩些，跟LeaveOneOut 也很像。
In [44]: from sklearn.model_selection import LeavePOut
In [45]: X = np.ones(4)
In [46]: lpo = LeavePOut(p=2)
In [47]: for train, test in lpo.split(X):
...: print train, test
...:
[2 3] [0 1]
[1 3] [0 2]
[1 2] [0 3]
[0 3] [1 2]
[0 2] [1 3]
[0 1] [2 3]
7.ShuffleSplit
ShuffleSplit 咋一看用法跟LeavePOut 很像，其實(shí)兩者完全不一樣，LeavePOut 是使得數(shù)據(jù)集經(jīng)過數(shù)次分割后，所有的測試集出現(xiàn)的元素的集合即是完整的數(shù)據(jù)集，即無放回的抽樣，而ShuffleSplit 則是有放回的抽樣，只能說經(jīng)過一個(gè)足夠大的抽樣次數(shù)后，保證測試集出現(xiàn)了完成的數(shù)據(jù)集的倍數(shù)。
In [48]: from sklearn.model_selection import ShuffleSplit
In [49]: X = np.arange(5)
In [50]: ss = ShuffleSplit(n_splits=3, test_size=.25, random_state=0)
In [51]: for train_index, test_index in ss.split(X):
...: print train_index, test_index
...:
[1 3 4] [2 0]
[1 4 3] [0 2]
[4 0 2] [1 3]
8.StratifiedKFold
這個(gè)就比較好玩了，通過指定分組，對(duì)測試集進(jìn)行無放回抽樣。
In [52]: from sklearn.model_selection import StratifiedKFold
In [53]: X = np.ones(10)
In [54]: y = [0,0,0,0,1,1,1,1,1,1]
In [55]: skf = StratifiedKFold(n_splits=3)
In [56]: for train, test in skf.split(X,y):
...: print train, test
...:
[2 3 6 7 8 9] [0 1 4 5]
[0 1 3 4 5 8 9] [2 6 7]
[0 1 2 4 5 6 7] [3 8 9]
9.GroupKFold
這個(gè)跟StratifiedKFold 比較像，不過測試集是按照一定分組進(jìn)行打亂的，即先分堆，然后把這些堆打亂，每個(gè)堆里的順序還是固定不變的。
In [57]: from sklearn.model_selection import GroupKFold
In [58]: X = [.1, .2, 2.2, 2.4, 2.3, 4.55, 5.8, 8.8, 9, 10]
In [59]: y = ['a','b','b','b','c','c','c','d','d','d']
In [60]: groups = [1,1,1,2,2,2,3,3,3,3]
In [61]: gkf = GroupKFold(n_splits=3)
In [62]: for train, test in gkf.split(X,y,groups=groups):
...: print train, test
...:
[0 1 2 3 4 5] [6 7 8 9]
[0 1 2 6 7 8 9] [3 4 5]
[3 4 5 6 7 8 9] [0 1 2]
10.LeaveOneGroupOut
這個(gè)是在GroupKFold 上的基礎(chǔ)上混亂度又減小了，按照給定的分組方式將測試集分割下來。
In [63]: from sklearn.model_selection import LeaveOneGroupOut
In [64]: X = [1, 5, 10, 50, 60, 70, 80]
In [65]: y = [0, 1, 1, 2, 2, 2, 2]
In [66]: groups = [1, 1, 2, 2, 3, 3, 3]
In [67]: logo = LeaveOneGroupOut()
In [68]: for train, test in logo.split(X, y, groups=groups):
...: print train, test
...:
[2 3 4 5 6] [0 1]
[0 1 4 5 6] [2 3]
[0 1 2 3] [4 5 6]
11.LeavePGroupsOut
這個(gè)沒啥可說的，跟上面那個(gè)一樣，只是一個(gè)是單組，一個(gè)是多組
from sklearn.model_selection import LeavePGroupsOut
X = np.arange(6)
y = [1, 1, 1, 2, 2, 2]
groups = [1, 1, 2, 2, 3, 3]
lpgo = LeavePGroupsOut(n_groups=2)
for train, test in lpgo.split(X, y, groups=groups):
print train, test
[4 5] [0 1 2 3]
[2 3] [0 1 4 5]
[0 1] [2 3 4 5]
12.GroupShuffleSplit
這個(gè)是有放回抽樣
In [75]: from sklearn.model_selection import GroupShuffleSplit
In [76]: X = [.1, .2, 2.2, 2.4, 2.3, 4.55, 5.8, .001]
In [77]: y = ['a', 'b','b', 'b', 'c','c', 'c', 'a']
In [78]: groups = [1,1,2,2,3,3,4,4]
In [79]: gss = GroupShuffleSplit(n_splits=4, test_size=.5, random_state=0)
In [80]: for train, test in gss.split(X, y, groups=groups):
...: print train, test
...:
[0 1 2 3] [4 5 6 7]
[2 3 6 7] [0 1 4 5]
[2 3 4 5] [0 1 6 7]
[4 5 6 7] [0 1 2 3]
13.TimeSeriesSplit
針對(duì)時(shí)間序列的處理，防止未來數(shù)據(jù)的使用，分割時(shí)是將數(shù)據(jù)進(jìn)行從前到后切割（這個(gè)說法其實(shí)不太恰當(dāng)，因?yàn)榍懈钍茄永m(xù)性的。。）
In [81]: from sklearn.model_selection import TimeSeriesSplit
In [82]: X = np.array([[1,2],[3,4],[1,2],[3,4],[1,2],[3,4]])
In [83]: tscv = TimeSeriesSplit(n_splits=3)
In [84]: for train, test in tscv.split(X):
...: print train, test
...:
[0 1 2] [3]
[0 1 2 3] [4]
[0 1 2 3 4] [5]
這個(gè)repo 用來記錄一些python技巧、書籍、學(xué)習(xí)鏈接等，歡迎star github地址
您可能感興趣的文章:- sklearn和keras的數(shù)據(jù)切分與交叉驗(yàn)證的實(shí)例詳解
- 使用sklearn的cross_val_score進(jìn)行交叉驗(yàn)證實(shí)例
- Python sklearn KFold 生成交叉驗(yàn)證數(shù)據(jù)集的方法
