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
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prediction
或GridSearchCV
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:
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。 recall
,f1_score
和precision_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。