Problem with custom Transformers for ColumnTransformer in scikit-learn

Question

I want to create a stable pipeline in scikit-learn for preprocessing the data. The first step that I am trying to complete is the imputation of None values applied with different strategies (i.e. replacing with mean, median or other descriptive statistics) for different columns in the dataframe. However I

I started using SimpleImputer transformer together with ColumnTransformer. Because SimpleImputer returns numpy array instead of pandas dataframe, I wrote a new transformer which uses SimpleImputer under the hood but adds pandas columns and indices back to the numpy array. Why I need exactly pandas dataframe back? Because I see my pipeline like this:

pipeline = Pipeline([
    ('imputation', ImputationColumnTransformer),
    ('feature_encoding', EncodingColumnTransformer),
    ('model', MLModel)
])

Without column access, the second step of feature encoding simply wouldn't be able to proceed.

The problem is that when I am using customized transformers I always get some errors from internal scikit-learn validation code.

I created a simple example to show the type of errors I get:

# Creating a toy dataset
m = np.random.randn(3, 3)
m[0, 1] = np.nan
m[2, 2] = np.nan
df = pd.DataFrame(m, columns=['a', 'b', 'c'])


class Imputer(BaseEstimator, TransformerMixin):
    # This transformer returns dataframe instead of default ndarray
    def __init__(self, ImputerCls, strategy):
        self.imputer = ImputerCls(strategy=strategy)

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

    def transform(self, X):
        res = self.imputer.transform(X)
        res = pd.DataFrame(res)
        res.columns = X.columns
        res.index = X.index
        return res


imputation = ColumnTransformer([
    ('categorial_imputer', Imputer(SimpleImputer, strategy='most_frequent'), ['a']),
    ('numeric_imputer', Imputer(SimpleImputer, strategy='mean'), ['b', 'c'])
])
imputation.fit_transform(df)

I expect pandas dataframe with all the columns preserved, however I am getting a long traceback log which I can not fully understand to find the problem. It seems that at some stage ImputerCls is None.

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-70-0ea27e638c36> in <module>
      3     ('numeric_imputer', Imputer(SimpleImputer, strategy='most_frequent'), ['b', 'c'])
      4 ])
----> 5 imputation.fit_transform(df)

~/anaconda3/lib/python3.7/site-packages/sklearn/compose/_column_transformer.py in fit_transform(self, X, y)
    466         self._validate_remainder(X)
    467 
--> 468         result = self._fit_transform(X, y, _fit_transform_one)
    469 
    470         if not result:

~/anaconda3/lib/python3.7/site-packages/sklearn/compose/_column_transformer.py in _fit_transform(self, X, y, func, fitted)
    410                     message=self._log_message(name, idx, len(transformers)))
    411                 for idx, (name, trans, column, weight) in enumerate(
--> 412                         self._iter(fitted=fitted, replace_strings=True), 1))
    413         except ValueError as e:
    414             if "Expected 2D array, got 1D array instead" in str(e):

~/anaconda3/lib/python3.7/site-packages/joblib/parallel.py in __call__(self, iterable)
    919             # remaining jobs.
    920             self._iterating = False
--> 921             if self.dispatch_one_batch(iterator):
    922                 self._iterating = self._original_iterator is not None
    923 

~/anaconda3/lib/python3.7/site-packages/joblib/parallel.py in dispatch_one_batch(self, iterator)
    752             tasks = BatchedCalls(itertools.islice(iterator, batch_size),
    753                                  self._backend.get_nested_backend(),
--> 754                                  self._pickle_cache)
    755             if len(tasks) == 0:
    756                 # No more tasks available in the iterator: tell caller to stop.

~/anaconda3/lib/python3.7/site-packages/joblib/parallel.py in __init__(self, iterator_slice, backend_and_jobs, pickle_cache)
    208 
    209     def __init__(self, iterator_slice, backend_and_jobs, pickle_cache=None):
--> 210         self.items = list(iterator_slice)
    211         self._size = len(self.items)
    212         if isinstance(backend_and_jobs, tuple):

~/anaconda3/lib/python3.7/site-packages/sklearn/compose/_column_transformer.py in <genexpr>(.0)
    409                     message_clsname='ColumnTransformer',
    410                     message=self._log_message(name, idx, len(transformers)))
--> 411                 for idx, (name, trans, column, weight) in enumerate(
    412                         self._iter(fitted=fitted, replace_strings=True), 1))
    413         except ValueError as e:

~/anaconda3/lib/python3.7/site-packages/sklearn/base.py in clone(estimator, safe)
     63     for name, param in new_object_params.items():
     64         new_object_params[name] = clone(param, safe=False)
---> 65     new_object = klass(**new_object_params)
     66     params_set = new_object.get_params(deep=False)
     67 

<ipython-input-57-a319579eaf68> in __init__(self, ImputerCls, strategy)
      2     # This class returns dataframe instead of default ndarray
      3     def __init__(self, ImputerCls, strategy):
----> 4         self.imputer = ImputerCls(strategy=strategy)
      5 
      6     def fit(self, X, y=None):

TypeError: 'NoneType' object is not callable

Answer 1

I got it working this way. I think the Imputer was not being instantiated:

from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.compose import ColumnTransformer
from sklearn.impute import SimpleImputer

import pandas as pd 
import numpy as np 

# Creating a toy dataset
m = np.random.randn(3, 3)
m[0, 1] = np.nan
m[2, 2] = np.nan
df = pd.DataFrame(m, columns=['a', 'b', 'c'])


class Imputer(BaseEstimator, TransformerMixin):
    # This transformer returns dataframe instead of default ndarray
    def __init__(self, imputer, strategy):
        self.imputer = imputer
        self.strategy = strategy

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

    def transform(self, X, *_):
        return self.imputer.transform(X)


imputation = ColumnTransformer([
    ('categorial_imputer', Imputer(SimpleImputer, strategy='most_frequent'), ['a']),
    ('numeric_imputer', Imputer(SimpleImputer, strategy='mean'), ['b', 'c'])
])
df = pd.DataFrame(imputation.fit_transform(df), columns=df.columns, index=df.index)

And that's it!

Answer 2

I think the problem is that ColumnTransformer returns a numpy darray. The above solution still converts the ColumnTransformer result to pandas dataframe outside the pipeline. Thus, the solution cannot be used as a step in sklearn Pipeline as the original poster desires.

I used inheritance to create a solution which can be used in a Pipeline. It works if the ColumnTransformer does not change the shape or order of columns or rows of input.

class PandasColumnTransformer(ColumnTransformer):
    def transform(self, X: pd.DataFrame) -> pd.DataFrame:
        return pd.DataFrame(super().transform(X), columns=X.columns, index=X.index)

    def fit_transform(self, X: pd.DataFrame, y=None) -> pd.DataFrame:
        return pd.DataFrame(super().fit_transform(X), columns=X.columns, index=X.index)


imputation = PandasColumnTransformer([
    ('categorial_imputer', SimpleImputer(strategy='most_frequent'), ['a']),
    ('numeric_imputer', SimpleImputer(strategy='mean'), ['b', 'c'])
])

df = imputation.fit_transform(df)