如何pickle sklearn管道对象？

Question

我想保存一个管道。我做不到。这是我的类对象，我试着将其pickling。

class SentimentModel():

    def __init__(self,model_instance,x_train,x_test,y_train,y_test):
        import string
        from nltk import ngrams
        self.ngrams = ngrams
        self.string = string
        self.model = model_instance
        self.x_train = x_train
        self.x_test = x_test
        self.y_train = y_train
        self.y_test = y_test       
        self._fit()


    def _fit(self):
        from sklearn.pipeline import Pipeline
        from sklearn.feature_extraction.text import TfidfTransformer
        from sklearn.feature_extraction.text import CountVectorizer      

        self.pipeline = Pipeline([
            ('bow', CountVectorizer(analyzer=self._text_process)), 
            ('tfidf', TfidfTransformer()), 
            ('classifier', self.model), 
        ])
        self.pipeline.fit(self.x_train,self.y_train)
        self.preds = self.pipeline.predict(self.x_test)     

    def _text_process(self,text):
        def remove_non_ascii(text):
            return ''.join(i for i in text if ord(i)<128)        

        text = remove_non_ascii(text)
        text = [char.lower() for char in text if char not in self.string.punctuation]
        text = ''.join(text)
        unigrams = [word for word in text.split()]
        bigrams = [' '.join(g) for g in self.ngrams(unigrams,2)]
        trigrams = [' '.join(g) for g in self.ngrams(unigrams,3)]
        tokens = []
        tokens.extend(unigrams+bigrams+trigrams)
        return tokens        

    def predict(self,observation):
        return self.pipeline.predict(observation)

我得到这些错误。

from sklearn.naive_bayes import MultinomialNB
nb = MultinomialNB()
nb_model = SentimentModel(nb,X_train,X_test,y_train,y_test)

import pickle
with open('nb_model1.pkl','wb') as f:
    pickle.dump(nb_model,f)

>>>
TypeError: can't pickle module objects

同样的：

with open('nb_model1.pkl','wb') as f:
    pickle.dump(nb_model.pipeline,f)

TypeError: can't pickle module objects

然而，我可以保存 nb_model.model. 但不是管道对象。有什么解释？如何让我的整个管道持久化？

我见过 如何在sklearn的Pipeline中挑选单个步骤？但问题是，它不能拾取的是 bow 属性。

joblib.dump(nb_model.pipeline.get_params()['tfidf'], 'nb_tfidf.pkl') # pass
joblib.dump(nb_model.pipeline.get_params()['bow'], 'nb_bow.pkl') # fail
joblib.dump(nb_model.pipeline.get_params()['classifier'], 'nb_classifier.pkl') #pass

>>>
TypeError: can't pickle module objects

我应该怎么做？

Answer 1

再试一次，不要在你的类定义中导入模块。这不是一个好的做法，因为当你导入一些东西，例如 import string，你给你的代码带来了一整套的第三方代码，甚至可能在其他想要使用这个皮卡的机器上都没有安装；这不是一个好的做法。也许 pickle 是保护你做这种事情。