CODEWITHSUNDEEP
pythonscikit-learnpipeline
Arnold Klein
Arnold Klein

Reputation: 3086

repeated FeatureUnion in scikit-learn

I am learning Pipelines and FeatureUnions in scikit-learn and thus wondering whether it is possible to repeated apply 'make_union' on a class?

Consider the following code:

import numpy as np
import pandas as pd
from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.linear_model import LogisticRegression
import sklearn.datasets as d

class IrisDataManupulation(BaseEstimator, TransformerMixin):
    """
       Raise the matrix of feature in power
    """
    def __init__(self, power=2):
        self.power = power

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

    def transform(self, X):
        return np.power(X, self.power)

iris_data = d.load_iris()

X, y = iris_data.data, iris_data.target


# feature union:
fu = FeatureUnion(transformer_list=[('squared', IrisDataManupulation(power=2)),
                               ('third', IrisDataManupulation(power=3))])

QUESTION Any neat way to create the FeatureUnion without repeating the same transformer, but rather passing a list of parameters?

For example:

fu_new = FeatureUnion(transformer_list=[('raise_power', IrisDataManupulation(), 
                      param_grid = {'raise_power__power':[2,3]})

Upvotes: 1

Views: 169

Answers (1)

Vivek Kumar
Vivek Kumar

Reputation: 36619

You can move all the powers work inside a single custom Transformer. We can change your IrisDataManupulation to handle the list of powers inside it:

class IrisDataManupulation(BaseEstimator, TransformerMixin):

    def __init__(self, powers=[2]):
        self.powers = powers

    def transform(self, X):
        powered_arrays = []
        for power in self.powers:
            powered_arrays.append(np.power(X, power))

        return np.hstack(powered_arrays)

Then you can just use this new transformer instead of FeatureUnion:

fu = IrisDataManupulation(powers=[2,3])

Note: If you want to generate polynomial features from your original features, I would recommend to see PolynomialFeatures, which can generate the powers you want in addition to other interactions between features.

Upvotes: 2

Related Questions