Sklearn：分组数据的交叉验证

Question

我正在尝试对分组数据实施交叉验证方案。我希望使用 GroupKFold 方法，但我一直收到错误消息。我究竟做错了什么？ 代码（与我使用的代码略有不同——我有不同的数据，所以我有一个更大的 n_splits，但其他一切都是一样的）

from sklearn import metrics
import matplotlib.pyplot as plt
import numpy as np
from sklearn.model_selection import GroupKFold
from sklearn.grid_search import GridSearchCV
from xgboost import XGBRegressor
#generate data
x=np.array([0,1,2,3,4,5,6,7,8,9,10,11,12,13])
y= np.array([1,2,3,4,5,6,7,1,2,3,4,5,6,7])
group=np.array([1,0,1,1,2,2,2,1,1,1,2,0,0,2)]
#grid search
gkf = GroupKFold( n_splits=3).split(x,y,group)
subsample = np.arange(0.3,0.5,0.1)
param_grid = dict( subsample=subsample)
rgr_xgb = XGBRegressor(n_estimators=50)
grid_search = GridSearchCV(rgr_xgb, param_grid, cv=gkf, n_jobs=-1)
result = grid_search.fit(x, y)

错误：

Traceback (most recent call last):

File "<ipython-input-143-11d785056a08>", line 8, in <module>
result = grid_search.fit(x, y)

File "/home/student/anaconda/lib/python3.5/site-packages/sklearn/grid_search.py", line 813, in fit
return self._fit(X, y, ParameterGrid(self.param_grid))

 File "/home/student/anaconda/lib/python3.5/site-packages/sklearn/grid_search.py", line 566, in _fit
n_folds = len(cv)

TypeError: object of type 'generator' has no len()

换线

gkf = GroupKFold( n_splits=3).split(x,y,group)

到

gkf = GroupKFold( n_splits=3)

也不行。然后错误信息是：

'GroupKFold' object is not iterable

Answer 1

split

的

GroupKFold

函数产生训练和测试指标一次一对。您应该在拆分值上调用

list

以将它们全部放入列表中，以便可以计算长度：

gkf = list(GroupKFold( n_splits=3).split(x,y,group))

Answer 2

这里是摩西答案的优化。同时存储所有拆分可能会限制内存，因此我们可以绕过原始的 yield 机制一次只返回一个训练/测试拆分

class KFoldHelper:
    def __init__(self, kfold: sklearn.model_selection._split._BaseKFold, x: np.ndarray,
                 classes: np.ndarray = None, groups: np.ndarray = None):
        self.iter = kfold.split(x, y = classes, groups=groups)

    def __iter__(self):
        for idxsTrain, idxsTest in self.iter:
            yield idxsTrain, idxsTest

现在我们可以打电话了

kfold = KFoldHelper(GroupKFold(n_splits=3), x, classes=y, groups=group)

和

GridSearchCV(rgr_xgb, param_grid, cv=kfold, n_jobs=-1)