instantiating a ColumnTransformer with already fitted transformers

[sebastiangonsal]

Can I instantiate a ColumnTransformer with already fitted transformers? I get an error right now saying the fit wasn't called on ColumnTransformer. I don't want to call fit on the transformer since I am using already fitted transformers and just want to use ColumnTransformer to aggregate the columns into a matrix.

[jnothman]

No you can't. What use case do you have, seeing as the applicable columns of the input are only identified when fitting the column transformer?

[thomasjpfan]

Recently spoke to someone that is running into this issue. He fitted a transformer on unlabeled data and they he created a metaestimator to wrap this transformer in order to freeze it (by overwriting fit to do nothing). Next, he placed it into a ColumnTransformer as one of the transformers. Since ColumnTransformer will clone its transformers, the metaestimator that he wanted to freeze lost its trained state.

He also mentioned how this worked in Pipeline and that its weird it didn't work in ColumnTransformer. (No cloning in Pipeline)

[jnothman]

Yes, cloning is precisely why freezing needs a specialised solution. It's very hard for a user unfamiliar with scikit-learn internals to fix.

[ribonucleic]

I have another use case where I want to fit two target encoders not on the target label but on other labels (e.g. email domains we get many clicks from, whereas the final target is a user buying). So basically I have 3 y vectors. Fitting has to happen outside of column transformer, then the fitted transformers would be passed to column transformer.

Great hint about "clone"! Helps to find a workaround.

[jnothman]

Why not pass categories to OneHotEncoder?

[sophieatalt]

was anyone able to find a solution to this?