如何在 scikit-learn 管道的步骤之间传递值? [英] How to pass values between steps of a scikit-learn pipeline?

问题描述

我想使用一个管道，它使用一个向量化器，然后是一个 LDA 预处理步骤.LDA 预处理步骤需要 Vectorizer 的词汇表.

I would like to use a pipeline, that uses a Vectorizer, followed by an LDA preprocessing step. The LDA preprocessing step needs the vocabulary_ of the Vectorizer.

我怎样才能将拟合的 Vectorizer 步骤的词汇表传递给下一个 LDA 步骤?我试图将管道本身传递给 LDA 步骤，但不幸的是这不起作用.

How can I pass thus the vocabulary_ of the fitted Vectorizer step to the next LDA step? I tried to pass the pipeline itself to the LDA step, but this unfortunately does not work.

    pipe_full = Pipeline(
        [('vect', StemmedCountVectorizer(strip_accents='unicode', analyzer='word')),
         ('lda', SklLdaModel_mod()),
         ('clf', SGDClassifier(loss='log', penalty='elasticnet', n_iter=5, random_state=42, class_weight={0: 1, 1: 2}))])
    param_grid_full = [{'vect__ngram_range': ((1, 1), (1, 2)), 'vect__stop_words': (None, 'english'),
                        'vect__token_pattern': (r'(?u)\b\w\w+\b', r'(?u)\b([a-zA-Z]{3,})\b'),
                        'vect__stemmer': (None, SnowCastleStemmer(mode='NLTK_EXTENSIONS')),
                        'lda': (None, SklLdaModel_mod(id2word=pipe_full, num_topics=10), SklLdaModel_mod(id2word=pipe_full, num_topics=20)),
                        # 'lda__topics': (100, 200),
                        # 'lda__topics': (10, 20),  # for testing purposes only
                        'clf__alpha': (1e-4, 5e-4)}]

...并且在 SklLdaModel_mod 的 fit 方法中我有:

... and in the fit method of SklLdaModel_mod I have:

    if isinstance(self.id2word, Pipeline):
        try:
            self.id2word = {v: k for k, v in self.id2word.named_steps['vect'].vocabulary_.items()}

任何建议如何做到这一点?

Any suggestions how to do this?

推荐答案

@Vivek,

不幸的是，这不起作用，因为 Vectorizer 也应该在管道内进行优化.查看不同的参数.

unfortunately this does not work, since the Vectorizer should also be optimized within the pipeline. See the different parameters.

我想出的解决方案有点笨拙:

The solution I came up with is a little hacky:

class XAmplifierForLDA(TransformerMixin, BaseEstimator):
    """
    This class amplifies the return value of the transform method of a model to include the vocab information for the 
    id2word parameter of the LDA model
    """
    def __init__(self, model=None):
        self.model = model

    def fit(self, *args, **kwargs):
        self.model.fit(*args, **kwargs)
        return self

    def transform(self, X, **transform_params):
        """
        This assumes model has a vocabulary
        :param X: 
        :param transform_params: 
        :return: 
        """
        return {'transformed': self.model.transform(X), 'vocab': self.model.vocabulary_}

    def set_params(self, **parameters):
        self.model.set_params(**parameters)
        return self

    def get_params(self, deep=True):
        """ return the parameters of the inner model """
        return {'model': self.model}

然后我将 CountVectorizer 包装在这个 XAmplifierLDA 中，然后除了词汇表之外，它还将返回一个包含转换后的 X 的字典！

and then I wrap the CountVectorizer inside this XAmplifierLDA, which will then return a dictionary with the transformed X in addition to the vocabulary!

 pipe_full = Pipeline(
            [('vect', XAmplifierForLDA(model=StemmedCountVectorizer(strip_accents='unicode', analyzer='word'))),
             ('lda', SklLdaModel_mod()),
             ('clf', SGDClassifier(loss='log', penalty='elasticnet', n_iter=5, random_state=42, class_weight={0: 1, 1: 2}))])

然后 SklLdaModel_mod 类负责正确解释字典.

The SklLdaModel_mod class then takes care to interpret the dictionary correctly.

关于如何更干净地实现这一点的任何其他想法?

Any other ideas of how to implement this more cleanly perhaps?

这篇关于如何在 scikit-learn 管道的步骤之间传递值?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！