Moving towards custom pipeline steps

Description

The current implementation of our model tuner supports custom steps in two ways: by using predefined steps like impute or by specifying a scaler type via the scaler_type variable, and by defining custom steps in the pipeline_steps. Managing compatibility between these two approaches is leading to complex code and interaction issues between these variables. We have decided to fully adopt the pipeline_steps approach.

This issue will focus on ensuring that all functionalities (such as ADASYN and SMOTE) work smoothly with pipeline_steps and cleaning up legacy code that still uses the older variables like impute or scaler_type.

Problem

Complexity and Compatibility: The mix of custom steps in pipeline_steps and predefined variables (impute, scaler_type, etc.) creates complexity and potential for bugs, making the code harder to maintain and extend.
Interaction Issues: The way these variables interact leads to compatibility issues, especially when combining custom steps with predefined configurations.

Proposed Solution

Full Migration to pipeline_steps: Ensure that all other code paths, including ADASYN and SMOTE implementations, work with custom steps defined in pipeline_steps.
Code Cleanup: Once compatibility is confirmed or fixed, remove the legacy support for predefined variables like impute and scaler_type. The entire process should be managed via pipeline_steps.

Code

This code is a fix for the SMOTE and ADASYN sections but still contains support for the previous implementation.

        if self.impute:
            preproc_test = clone(self.estimator.named_steps["imputer"])
        elif self.pipeline:
            ### Need to detect what the name of a column transformer has been called
            ### if we are using custom pipeline steps
            preproc_test = clone(self.estimator.named_steps['Preprocessor'])
        else:
            pass

        resampler_test = clone(self.estimator.named_steps["Resampler"])

uclamii / model_tuner