如何将不同的输入拟合到 sklearn 管道中？

Posted 2023-03-12

【中文标题】如何将不同的输入拟合到 sklearn 管道中？

【英文标题】：How to fit different inputs into an sklearn Pipeline?

【发布时间】：2016-07-15 04:13:44

【问题描述】：

我正在使用 sklearn 中的 Pipeline 对文本进行分类。

在这个管道示例中，我有一个 TfIDF 矢量化器和一些用 FeatureUnion 和分类器包装的自定义特征作为管道步骤，然后我拟合训练数据并进行预测：

from sklearn.pipeline import FeatureUnion, Pipeline
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.svm import LinearSVC

X = ['I am a sentence', 'an example']
Y = [1, 2]
X_dev = ['another sentence']

# load custom features and FeatureUnion with Vectorizer
features = []
measure_features = MeasureFeatures() # this class includes my custom features
features.append(('measure_features', measure_features))

countVecWord = TfidfVectorizer(ngram_range=(1, 3), max_features= 4000)
features.append(('ngram', countVecWord))

all_features = FeatureUnion(features)

# classifier
LinearSVC1 = LinearSVC(tol=1e-4,  C = 0.10000000000000001)

pipeline = Pipeline(
    [('all', all_features ),
    ('clf', LinearSVC1),
    ])

pipeline.fit(X, Y)
y_pred = pipeline.predict(X_dev)

# etc.

上面的代码工作得很好，但有一个转折点。我想对文本进行部分语音标记，并在 tagget 文本上使用不同的 Vectorizer。

X = ['I am a sentence', 'an example']
X_tagged = do_tagging(X) 
# X_tagged = ['PP AUX DET NN', 'DET NN']
Y = [1, 2]
X_dev = ['another sentence']
X_dev_tagged = do_tagging(X_dev)

# load custom featues and FeatureUnion with Vectorizer
features = []
measure_features = MeasureFeatures() # this class includes my custom features
features.append(('measure_features', measure_features))

countVecWord = TfidfVectorizer(ngram_range=(1, 3), max_features= 4000)
# new POS Vectorizer
countVecPOS = TfidfVectorizer(ngram_range=(1, 4), max_features= 2000)

features.append(('ngram', countVecWord))
features.append(('pos_ngram', countVecWord))

all_features = FeatureUnion(features)

# classifier
LinearSVC1 = LinearSVC(tol=1e-4,  C = 0.10000000000000001)

pipeline = Pipeline(
    [('all', all_features ),
    ('clf', LinearSVC1),
    ])

# how do I fit both X and X_tagged here
# how can the different vectorizers get either X or X_tagged?
pipeline.fit(X, Y)
y_pred = pipeline.predict(X_dev)

# etc.

如何正确拟合此类数据？两个矢量化器如何区分原始文本和 pos 文本？我有哪些选择？

我也有自定义功能，其中一些会采用原始​​文本，而另一些会采用 POS 文本。

编辑：添加 MeasureFeatures()

from sklearn.base import BaseEstimator
import numpy as np

class MeasureFeatures(BaseEstimator):

    def __init__(self):
        pass

    def get_feature_names(self):
        return np.array(['type_token', 'count_nouns'])

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

    def transform(self, x_dataset):


        X_type_token = list()
        X_count_nouns = list()

        for sentence in x_dataset:

            # takes raw text and calculates type token ratio
            X_type_token.append(type_token_ratio(sentence))

            # takes pos tag text and counts number of noun pos tags (NN, NNS etc.)
            X_count_nouns.append(count_nouns(sentence))

        X = np.array([X_type_token, X_count_nouns]).T

        print X
        print X.shape

        if not hasattr(self, 'scalar'):
            self.scalar = StandardScaler().fit(X)
        return self.scalar.transform(X)

然后，此特征转换器需要为 count_nouns() 函数获取标记文本或为 type_token_ratio() 获取原始文本

【参考方案1】：

我认为您必须在 2 个 Transformer（TfidfTransformer 和 POSTransformer）上执行 FeatureUnion。当然你需要定义那个 POSTransformer。 也许这个article 会对你有所帮助。

也许你的管道会是这样的。

pipeline = Pipeline([
  ('features', FeatureUnion([
    ('ngram_tf_idf', Pipeline([
      ('counts_ngram', CountVectorizer()),
      ('tf_idf_ngram', TfidfTransformer())
    ])),
    ('pos_tf_idf', Pipeline([
      ('pos', POSTransformer()),          
      ('counts_pos', CountVectorizer()),
      ('tf_idf_pos', TfidfTransformer())
    ])),
    ('measure_features', MeasureFeatures())
  ])),
  ('classifier', LinearSVC())
])

这假设 MeasureFeatures 和 POSTransformer 是符合 sklearn API 的 Transformer。

【讨论】：

我在最新的编辑中添加了 MeasureFeatures()。基本上，它需要为一组功能获取原始文本，为另一组功能获取 pos 标签集。有两个 MeasureFeature 类会有帮助吗？一种用于原始文本功能，一种用于 pos 标签功能？

我在这里看不到您的工作流程。看看我向你提议的那个，链接和这个例子（scikit-learn.org/stable/auto_examples/hetero_feature_union.html）。之后，您只需要考虑您的工作流程，您的数据会发生什么。