Sklearn Pipeline：将参数传递给自定义变压器？

Question

我的 sklearn 管道中有一个自定义 Transformer，我想知道如何将参数传递给我的 Transformer：

在下面的代码中，您可以看到我在 Transformer 中使用了字典“权重”。我不想在我的 Transformer 中定义这个字典，而是从管道传递它，这样我就可以在网格搜索中包含这个字典。是否可以将字典作为参数传递给我的 Transformer ？

# My custom Transformer
  class TextExtractor(BaseEstimator, TransformerMixin):
        """Concat the 'title', 'body' and 'code' from the results of 
        Stackoverflow query
        Keys are 'title', 'body' and 'code'.
        """
        def fit(self, x, y=None):
            return self

        def transform(self, x):
            # here is the parameter  I want to pass to my transformer
            weight ={'title' : 10, 'body': 1, 'code' : 1}
            x['text'] = weight['title']*x['Title'] +  
            weight['body']*x['Body'] +  
            weight['code']*x['Code']

            return x['text']

param_grid = {
    'min_df' : [10],
    'max_df' : [0.01],
    'max_features': [200],
    'clf' : [sgd]
    # here is the parameter  I want to pass to my transformer
    'weigth' : [{'title' : 10, 'body': 1, 'code' : 1}, {'title' : 1, 'body': 
     1, 'code' : 1}]

}

for g in ParameterGrid(param_grid) :   

    classifier_pipe = Pipeline(

    steps=[    ('textextractor', TextExtractor()), #is it possible to pass 
                my parameter ?
               ('vectorizer', TfidfVectorizer(max_df=g['max_df'], 
                     min_df=g['min_df'], max_features=g['max_features'])),
               ('clf', g['clf']), 
            ],
    )

Answer 1

为此，您只需在类定义的开头添加一个__init__() 方法。在此步骤中，您将定义您的类 TextExtractor 为采用您称为 weight 的参数。

这是如何完成的：（为了可重复性，我之前添加了很多代码行 - 鉴于您没有指定任何内容，我编造了一些虚假数据。我还假设您正在尝试使用权重是乘以字符串？）

# import all the necessary packages
from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.pipeline import Pipeline
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.model_selection import ParameterGrid, GridSearchCV
from sklearn.linear_model import SGDClassifier

import pandas as pd
import numpy as np

#Sample data
X = pd.DataFrame({"Title" : ["T1","T2","T3","T4","T5"], "Body": ["B1","B2","B3","B4","B5"], "Code": ["C1","C2","C3","C4","C5"]})
y = np.array([0,0,1,1,1])

#Define the SGDClassifier
sgd = SGDClassifier()

下面，我只添加了init这一步：

# My custom Transformer

class TextExtractor(BaseEstimator, TransformerMixin):
    """Concat the 'title', 'body' and 'code' from the results of 
    Stackoverflow query
    Keys are 'title', 'body' and 'code'.


    """

    def __init__(self, weight = {'title' : 10, 'body': 1, 'code' : 1}):

        self.weight = weight

    def fit(self, x, y=None):
        return self

    def transform(self, x):

        x['text'] = self.weight['title']*x['Title'] + self.weight['body']*x['Body'] + self.weight['code']*x['Code']

        return x['text']

请注意，在您未指定的情况下，我默认传递了一个参数值。这取决于你。然后你可以调用你的转换器：

textextractor = TextExtractor(weight = {'title' : 5, 'body': 2, 'code' : 1})
textextractor.transform(X)

这应该返回：

0    T1T1T1T1T1B1B1C1
1    T2T2T2T2T2B2B2C2
2    T3T3T3T3T3B3B3C3
3    T4T4T4T4T4B4B4C4
4    T5T5T5T5T5B5B5C5

然后你可以定义你的参数网格：

param_grid = {
'vectorizer__min_df' : [0.1],
'vectorizer__max_df' : [0.9],
'vectorizer__max_features': [200],
# here is the parameter  I want to pass to my transformer
'textextractor__weight' : [{'title' : 10, 'body': 1, 'code' : 1}, {'title' : 1, 'body': 
 1, 'code' : 1}]
}

最后做：

for g in ParameterGrid(param_grid) :   

classifier_pipe = Pipeline(

steps=[    ('textextractor', TextExtractor(weight = g['textextractor__weight'])), 
           ('vectorizer', TfidfVectorizer(max_df=g['vectorizer__max_df'], 
                 min_df=g['vectorizer__min_df'], max_features=g['vectorizer__max_features'])),
           ('clf', sgd),  ] )

您可能想要做一个网格搜索，而不是这个，这需要您编写：

pipe = Pipeline( steps=[    ('textextractor', TextExtractor()), 
           ('vectorizer', TfidfVectorizer()),
           ('clf', sgd) ] )
grid = GridSearchCV(pipe, param_grid, cv = 3)
grid.fit(X,y)