在运行期间计算python中GradientBoostingClassifier的损失

Question

我有以下用于创建和训练sklearn.ensemble.GradientBoostingClassifier的代码

class myMonitor:
    def __call__(self, i, estimator, locals):
        proba = estimator.predict_proba(Xp2)
        myloss = calculateMyLoss(proba, yp2) # calculateMyLoss is defined 
                                             # further on
        print("Calculated MYLOSS: ",myloss)
        return False

... #some more code

model = GradientBoostingClassifier(verbose=2, learning_rate = learningRate, n_estimators=numberOfIterations, max_depth=maxDepth, subsample = theSubsample, min_samples_leaf = minLeafSamples, max_features=maxFeatures)
model.fit(Xp1, yIntegersp1, monitor = myMonitor())

然而，当我运行这段代码时，我得到了错误：

    model.fit(Xp1, yIntegersp1, monitor = myMonitor())
  File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/sklearn/ensemble/gradient_boosting.py", line 980, in fit
    begin_at_stage, monitor)
  File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/sklearn/ensemble/gradient_boosting.py", line 1058, in _fit_stages
    early_stopping = monitor(i, self, locals())
  File "OTTOSolverGBM.py", line 44, in __call__
    proba = estimator.predict_proba(Xp2)
  File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/sklearn/ensemble/gradient_boosting.py", line 1376, in predict_proba
    score = self.decision_function(X)
  File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/sklearn/ensemble/gradient_boosting.py", line 1102, in decision_function
    score = self._decision_function(X)
  File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/sklearn/ensemble/gradient_boosting.py", line 1082, in _decision_function
    predict_stages(self.estimators_, X, self.learning_rate, score)
  File "sklearn/ensemble/_gradient_boosting.pyx", line 115, in sklearn.ensemble._gradient_boosting.predict_stages (sklearn/ensemble/_gradient_boosting.c:2502)
AttributeError: 'NoneType' object has no attribute 'tree_'

为什么我不能使用相同的估计器(不是None，我检查过了)来计算运行期间的类概率？有没有办法实现我想要的(例如，在拟合过程的每一次迭代中检查模型的验证数据)？

Answer 1

您的estimator是self。试一试

def __call__(self, i, locals)
    proba = self.predict_proba(Xp2)

Answer 2

你也许可以做一些类似于this example on forests的基于partial_fit的事情。要在培训后进行分析，请查看this example on gradient boosting.