Using multiple custom classes with Pipeline sklearn (Python)

问题

I try to do a tutorial on Pipeline for students but I block. I'm not an expert but I'm trying to improve. So thank you for your indulgence. In fact, I try in a pipeline to execute several steps in preparing a dataframe for a classifier:

• Step 1: Description of the dataframe
• Step 2: Fill NaN Values
• Step 3: Transforming Categorical Values into Numbers

Here is my code:

class Descr_df(object):

    def transform (self, X):
        print ("Structure of the data: \n {}".format(X.head(5)))
        print ("Features names: \n {}".format(X.columns))
        print ("Target: \n {}".format(X.columns[0]))
        print ("Shape of the data: \n {}".format(X.shape))

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

class Fillna(object):

    def transform(self, X):
        non_numerics_columns = X.columns.difference(X._get_numeric_data().columns)
        for column in X.columns:
            if column in non_numerics_columns:
                X[column] = X[column].fillna(df[column].value_counts().idxmax())
            else:
                 X[column] = X[column].fillna(X[column].mean())            
        return X

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

class Categorical_to_numerical(object):

    def transform(self, X):
        non_numerics_columns = X.columns.difference(X._get_numeric_data().columns)
        le = LabelEncoder()
        for column in non_numerics_columns:
            X[column] = X[column].fillna(X[column].value_counts().idxmax())
            le.fit(X[column])
            X[column] = le.transform(X[column]).astype(int)
        return X

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

If I execute step 1 and 2 or step 1 and 3 it works but if I execute step 1, 2 and 3 at the same time. I have this error:

pipeline = Pipeline([('df_intropesction', Descr_df()), ('fillna',Fillna()), ('Categorical_to_numerical', Categorical_to_numerical())])
pipeline.fit(X, y)
AttributeError: 'NoneType' object has no attribute 'columns'

回答1:

This error arises because in the Pipeline the output of first estimator goes to the second, then the output of second estimator goes to third and so on...

From the documentation of Pipeline:

Fit all the transforms one after the other and transform the data, then fit the transformed data using the final estimator.

So for your pipeline, the steps of execution are following:

1. Descr_df.fit(X) -> doesn't do anything and returns self
2. newX = Descr_df.transform(X) -> should return some value to assign to newX that should be passed on to next estimator, but your definition does not return anything (only prints). So None is returned implicitly
3. Fillna.fit(newX) -> doesn't do anything and returns self
4. Fillna.transform(newX) -> Calls newX.columns. But newX=None from step2. Hence the error.

Solution: Change the transform method of Descr_df to return the dataframe as it is:

def transform (self, X):
    print ("Structure of the data: \n {}".format(X.head(5)))
    print ("Features names: \n {}".format(X.columns))
    print ("Target: \n {}".format(X.columns[0]))
    print ("Shape of the data: \n {}".format(X.shape))
    return X

Suggestion : Make your classes inherit from Base Estimator and Transformer classes in scikit to confirm to the good practice.

i.e change the class Descr_df(object) to class Descr_df(BaseEstimator, TransformerMixin), Fillna(object) to Fillna(BaseEstimator, TransformerMixin) and so on.

See this example for more details on custom classes in Pipeline:

• http://scikit-learn.org/stable/auto_examples/hetero_feature_union.html#sphx-glr-auto-examples-hetero-feature-union-py

来源：https://stackoverflow.com/questions/43499342/using-multiple-custom-classes-with-pipeline-sklearn-python