sklearn GridSearchCV(评分函数错误)

时间:2015-07-24 16:12:02

标签: python machine-learning scikit-learn grid-search

我想知道你是否可以帮助我解决我在运行网格搜索时收到的错误。我认为这可能是由于对网格搜索实际如何运作的误解。

我现在正在运行一个应用程序,我需要使用网格搜索来评估使用不同评分函数的最佳参数。我使用RandomForestClassifier将大X数据集拟合到特征向量Y,该特征向量Y是0和1的列表。 (完全二进制)。我的评分函数(MCC)要求预测输入和实际输入完全是二进制的。但是,出于某种原因,我一直得到ValueError:不支持多类。

我的理解是,网格搜索对数据集进行交叉验证,提出基于交叉验证的预测输入,然后将特征向量和预测插入到函数中。由于我的特征向量是完全二元的,我的预测向量也应该是二进制的,并且在评估分数时不会产生任何问题。 当我使用单个定义的参数(不使用网格搜索)运行随机森林时,将预测数据和特征向量插入到MCC评分函数中运行完全正常。所以我对运行网格搜索会导致任何错误感到有点迷失。

数据快照:

        print len(X)
        print X[0]
        print len(Y)
        print Y[2990:3000]
17463699
[38.110903683955435, 38.110903683955435, 38.110903683955435, 9.899495124816895, 294.7808837890625, 292.3835754394531, 293.81494140625, 291.11065673828125, 293.51739501953125, 283.6424865722656, 13.580912590026855, 4.976086616516113, 1.1271398067474365, 0.9465181231498718, 0.5066819190979004, 0.1808401197195053, 0.0]
17463699
[0.0, 0.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]

代码:

def overall_average_score(actual,prediction):
    precision = precision_recall_fscore_support(actual, prediction, average = 'binary')[0]
    recall = precision_recall_fscore_support(actual, prediction, average = 'binary')[1]
    f1_score = precision_recall_fscore_support(actual, prediction, average = 'binary')[2]
    total_score = matthews_corrcoef(actual, prediction)+accuracy_score(actual, prediction)+precision+recall+f1_score
    return total_score/5

grid_scorer = make_scorer(overall_average_score, greater_is_better=True)
parameters = {'n_estimators': [10,20,30], 'max_features': ['auto','sqrt','log2',0.5,0.3], }
random  = RandomForestClassifier()
clf = grid_search.GridSearchCV(random, parameters, cv = 5, scoring = grid_scorer)
clf.fit(X,Y)

错误:

ValueError                                Traceback (most recent call last)
<ipython-input-39-a8686eb798b2> in <module>()
     18 random  = RandomForestClassifier()
     19 clf = grid_search.GridSearchCV(random, parameters, cv = 5, scoring = grid_scorer)
---> 20 clf.fit(X,Y)
     21 
     22 

/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/grid_search.pyc in fit(self, X, y)
    730 
    731         """
--> 732         return self._fit(X, y, ParameterGrid(self.param_grid))
    733 
    734 

/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/grid_search.pyc in _fit(self, X, y, parameter_iterable)
    503                                     self.fit_params, return_parameters=True,
    504                                     error_score=self.error_score)
--> 505                 for parameters in parameter_iterable
    506                 for train, test in cv)
    507 

/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.pyc in __call__(self, iterable)
    657             self._iterating = True
    658             for function, args, kwargs in iterable:
--> 659                 self.dispatch(function, args, kwargs)
    660 
    661             if pre_dispatch == "all" or n_jobs == 1:

/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.pyc in dispatch(self, func, args, kwargs)
    404         """
    405         if self._pool is None:
--> 406             job = ImmediateApply(func, args, kwargs)
    407             index = len(self._jobs)
    408             if not _verbosity_filter(index, self.verbose):

/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.pyc in __init__(self, func, args, kwargs)
    138         # Don't delay the application, to avoid keeping the input
    139         # arguments in memory
--> 140         self.results = func(*args, **kwargs)
    141 
    142     def get(self):

/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/cross_validation.pyc in _fit_and_score(estimator, X, y, scorer, train, test, verbose, parameters, fit_params, return_train_score, return_parameters, error_score)
   1476 
   1477     else:
-> 1478         test_score = _score(estimator, X_test, y_test, scorer)
   1479         if return_train_score:
   1480             train_score = _score(estimator, X_train, y_train, scorer)

/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/cross_validation.pyc in _score(estimator, X_test, y_test, scorer)
   1532         score = scorer(estimator, X_test)
   1533     else:
-> 1534         score = scorer(estimator, X_test, y_test)
   1535     if not isinstance(score, numbers.Number):
   1536         raise ValueError("scoring must return a number, got %s (%s) instead."

/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/metrics/scorer.pyc in __call__(self, estimator, X, y_true, sample_weight)
     87         else:
     88             return self._sign * self._score_func(y_true, y_pred,
---> 89                                                  **self._kwargs)
     90 
     91 

<ipython-input-39-a8686eb798b2> in overall_average_score(actual, prediction)
      3     recall = precision_recall_fscore_support(actual, prediction, average = 'binary')[1]
      4     f1_score = precision_recall_fscore_support(actual, prediction, average = 'binary')[2]
----> 5     total_score = matthews_corrcoef(actual, prediction)+accuracy_score(actual, prediction)+precision+recall+f1_score
      6     return total_score/5
      7 def show_score(actual,prediction):

/shared/studies/nonregulated/neurostream/neurostream/local/lib/python2.7/site-packages/sklearn/metrics/classification.pyc in matthews_corrcoef(y_true, y_pred)
    395 
    396     if y_type != "binary":
--> 397         raise ValueError("%s is not supported" % y_type)
    398 
    399     lb = LabelEncoder()

ValueError: multiclass is not supported

2 个答案:

答案 0 :(得分:1)

我复制了你的实验,但我没有收到任何错误。 该错误表示您的某个向量predictionGridSearchCV 包含两个以上的离散值

你能够在from sklearn.datasets import make_classification from sklearn.grid_search import GridSearchCV from sklearn.metrics import precision_recall_fscore_support, accuracy_score, \ matthews_corrcoef, make_scorer from sklearn.ensemble import RandomForestClassifier from sklearn.cross_validation import train_test_split def overall_average_score(actual,prediction): precision, recall, f1_score, _ = precision_recall_fscore_support( actual, prediction, average='binary') total_score = (matthews_corrcoef(actual, prediction) + accuracy_score(actual, prediction) + precision + recall + f1_score) return total_score / 5 grid_scorer = make_scorer(overall_average_score, greater_is_better=True) print("Without GridSearchCV") X, y = make_classification(n_samples=500, n_informative=10, n_classes=2) X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5, random_state=0) rf = RandomForestClassifier() rf.fit(X_train, y_train) y_pred = rf.predict(X_test) print("Overall average score: ", overall_average_score(y_test, y_pred)) print("-" * 30) print("With GridSearchCV:") parameters = {'n_estimators': [10,20,30], 'max_features': ['auto','sqrt','log2',0.5,0.3], } gs_rf = GridSearchCV(rf, parameters, cv=5, scoring=grid_scorer) gs_rf.fit(X_train,y_train) print("Best score with grid search: ", gs_rf.best_score_) 之外训练一个随机森林,这确实很奇怪 你能提供你运行的确切代码吗?

以下是我用来尝试重现错误的代码:

random

现在,我想就您提供的代码发表一些意见:

  • 使用变量名称(例如f1_score(通常是模块)或sklearn.metrics.f1_score(这与precision方法冲突)并不是一种很好的做法。
  • 您可以直接解包recallf1_scoreprecision_recall_fscore_support,而不是拨打n_estimators 3次。
  • max_depth上的网格搜索没有意义:更多的树总是更好。如果您担心过度拟合,可以使用min_samples_split或{{1}}等其他参数来降低单个模型的复杂程度。

答案 1 :(得分:1)

马修斯相关系数是-1和1之间的一个分数。因此,计算f1_score,precision,recall,precision_score和MCC之间的平均值是不正确的。

MCC值指示:    1是总正相关    0无相关   -1是总负相关

虽然上面提到的其他评估指标介于0和1之间(从最差到最佳准确度指标)。范围和意义不同。