The fact that you cannot print the de-serialized preprocessor is not necessarily an issue I think? It seems to follow the same behaviour before/after (de)serialization.

I was looking into the issue of why you cannot transform after de-serializing, and I think some information is lost in the (de-)serialization process.

To give an example:

• After creating and fitting a preprocessor object (called preprocessor) on sample data (using only continuous variables), there is some information stored on the bins per column which becomes visible by running:
  preprocessor._discretizer._bins_by_column.

The _bins_by_column element is not visible when just looking at the _discretizer, but it is still there.

• After serializing and de-serializing the same preprocessor object, this information is lost. When calling the transform method from the kbins_discretizer class, the following test is ran (line 272):

    if len(self._bins_by_column) == 0:
        msg = ("{} instance is not fitted yet. Call 'fit' with "
                   "appropriate arguments before using this method.")
        raise NotFittedError(msg.format(self.__class__.__name__))

This is why (for me) I cannot directly transform new data after de-serializing. I already wanted to leave some information here, but will investigate this further. I can imagine this is also happening in the categorical data processor.

A way forward would probably be to make sure the full information gets (de-)serialized.

Patrick's logged issue #176 might be a duplicate, @patrickleonardy to check if yours was on the model and Joost's on the preprocessor.
If questions on how to fix, you can also ask Benoît, who has struggled with the issue +- a month ago.

#176 relates to the serialization/de-serialization of the LinearRegressionModel and maybe the LogisticRegressionModel,
Where as this Issue is about the PreProcessing serialization/de-serialization

So no duplicates here

@patrickleonardy @sandervh14

For me the issue is solved now. The main issue was in target_encoder.py.

At line 126, there is a check on a parameter (_global_mean) of the target encoder. This is a floating number, in my case of type np.float64. In the if statement was only a check if type==float. This check failed, and hence the variable is left empty in the deserialization process. Therefore Cobra suspects the target_encoder was not fitted.

I extended the check to take into account different kinds of floating numbers using:
isinstance(params["_global_mean"], (np.floating, float))

I have tested the entire flow with continuous and categorical variables and everything seems to work fine now. The debugging is documented in a notebook which at the moment pushed to git as well.