feature request: only apply random noise in `RandomAdder` to training data about scikit-lego

[MBrouns]

So I got asked a similar type of thing in todays training where someone wanted to drop rows with too many missing values from train but not from test so I was toying around to see if I could find something that would work.

I might have figured out a way but I'm not sure I like it all that much:

import pandas as pd
import hashlib

class TrainOnlyMixin(BaseEstimator, TransformerMixin):
    
    def fit(self, X, y):
        self.df_hash_ = self.hash_df(X)
        return self
    
    
    @staticmethod
    def hash_df(df):
        return hashlib.sha256(pd.util.hash_pandas_object(df, index=True).values).hexdigest()
    
    def transform(self, X, y=None):
    
        if self.hash_df(X) == self.df_hash_:
            return self.transform_train(X)
        
        else:
            return self.transform_test(X)

I basically store a hash of the train dataframe and compare X with it in transform and then call transform_train or transform_test. I think this can be made quite generic and I can't think of a case where it wouldn't work. What do you think?

[koaning]

I think we agree this can't be done. Close issue?

[koaning]

My first thought: WE NEED TO BE CAREFUL there's methods that might do a partial fit and then this would totally break.

My second thought: but yeah ... this could totally work

Third thought: do we have other usecases for it?

Fourth thought: will it also work on numpy arrays?

Thought five: I'm not sure if I want to implement transform_test on each object. cant we just return X?

[MBrouns]

1. Pipelines don't implement partial_fit so that shouldn't be a problem i think

2. 😄

3. So far I've come up with the following use cases:
   a. dropping rows with many missing values from train but not from test
   b. adding noise to train but not to test
   c. removing outliers from training data

4. As is no, but it is possible to calculate a hash on a numpy array afaik

5. Agreed, probably have this as a mixin where transform_test has a default implementation to return X

[MBrouns]

#81