python - stratifiedkfold - train_test_split
Scikit-learn TypeError: si no se especifica ningĂșn puntaje, el estimador aprobado debe tener un mĂ©todo de ''puntaje'' (1)
Creé un modelo personalizado en python usando scikit-learn, y quiero usar la validación cruzada.
La clase para el modelo se define de la siguiente manera:
class MultiLabelEnsemble:
'''''' MultiLabelEnsemble(predictorInstance, balance=False)
Like OneVsRestClassifier: Wrapping class to train multiple models when
several objectives are given as target values. Its predictor may be an ensemble.
This class can be used to create a one-vs-rest classifier from multiple 0/1 labels
to treat a multi-label problem or to create a one-vs-rest classifier from
a categorical target variable.
Arguments:
predictorInstance -- A predictor instance is passed as argument (be careful, you must instantiate
the predictor class before passing the argument, i.e. end with (),
e.g. LogisticRegression().
balance -- True/False. If True, attempts to re-balance classes in training data
by including a random sample (without replacement) s.t. the largest class has at most 2 times
the number of elements of the smallest one.
Example Usage: mymodel = MultiLabelEnsemble (GradientBoostingClassifier(), True)''''''
def __init__(self, predictorInstance, balance=False):
self.predictors = [predictorInstance]
self.n_label = 1
self.n_target = 1
self.n_estimators = 1 # for predictors that are ensembles of estimators
self.balance=balance
def __repr__(self):
return "MultiLabelEnsemble"
def __str__(self):
return "MultiLabelEnsemble : /n" + "/tn_label={}/n".format(self.n_label) + "/tn_target={}/n".format(self.n_target) + "/tn_estimators={}/n".format(self.n_estimators) + str(self.predictors[0])
def fit(self, Xtrain, Ytrain):
if len(Ytrain.shape)==1:
Ytrain = np.array([Ytrain]).transpose() # Transform vector into column matrix
# This is NOT what we want: Y = Y.reshape( -1, 1 ), because Y.shape[1] out of range
self.n_target = Ytrain.shape[1] # Num target values = num col of Y
self.n_label = len(set(Ytrain.ravel())) # Num labels = num classes (categories of categorical var if n_target=1 or n_target if labels are binary )
# Create the right number of copies of the predictor instance
if len(self.predictors)!=self.n_target:
predictorInstance = self.predictors[0]
self.predictors = [predictorInstance]
for i in range(1,self.n_target):
self.predictors.append(copy.copy(predictorInstance))
# Fit all predictors
for i in range(self.n_target):
# Update the number of desired prodictos
if hasattr(self.predictors[i], ''n_estimators''):
self.predictors[i].n_estimators=self.n_estimators
# Subsample if desired
if self.balance:
pos = Ytrain[:,i]>0
neg = Ytrain[:,i]<=0
if sum(pos)<sum(neg):
chosen = pos
not_chosen = neg
else:
chosen = neg
not_chosen = pos
num = sum(chosen)
idx=filter(lambda(x): x[1]==True, enumerate(not_chosen))
idx=np.array(zip(*idx)[0])
np.random.shuffle(idx)
chosen[idx[0:min(num, len(idx))]]=True
# Train with chosen samples
self.predictors[i].fit(Xtrain[chosen,:],Ytrain[chosen,i])
else:
self.predictors[i].fit(Xtrain,Ytrain[:,i])
return
def predict_proba(self, Xtrain):
if len(Xtrain.shape)==1: # IG modif Feb3 2015
X = np.reshape(Xtrain,(-1,1))
prediction = self.predictors[0].predict_proba(Xtrain)
if self.n_label==2: # Keep only 1 prediction, 1st column = (1 - 2nd column)
prediction = prediction[:,1]
for i in range(1,self.n_target): # More than 1 target, we assume that labels are binary
new_prediction = self.predictors[i].predict_proba(Xtrain)[:,1]
prediction = np.column_stack((prediction, new_prediction))
return prediction
Cuando llamo a esta clase para validación cruzada como esta:
kf = cross_validation.KFold(len(Xtrain), n_folds=10)
score = cross_val_score(self.model, Xtrain, Ytrain, cv=kf, n_jobs=-1).mean()
Obtuve el siguiente error:
TypeError: si no se especifica ningún puntaje, el estimador aprobado debe tener un método de "puntaje". El estimador MultiLabelEnsemble no lo hace.
¿Cómo creo un método de puntaje?
La forma más fácil de hacer desaparecer el error es pasar la scoring="accuracy"
o la scoring="hamming"
a cross_val_score
. La función cross_val_score
sí no sabe qué tipo de problema está tratando de resolver, por lo que no sabe qué medida es la adecuada. Parece que estás tratando de hacer una clasificación de etiqueta múltiple, ¿entonces quizás quieras usar la pérdida de Hamming?
También puede implementar un método de score
como se explica en los documentos "Roll your own estimator", que tiene como def score(self, X, y_true)
signature def score(self, X, y_true)
. Ver http://scikit-learn.org/stable/developers/#different-objects
Por cierto, usted sabe sobre el OneVsRestClassifier
, ¿verdad? Parece un poco como si lo estuvieras reimplantando.