cdplib/cdplib/hyperparameter_space_composer/SpaceComposer.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Wed Sep 30 13:54:04 2020

@author: tanya
@description: a class that from a given list of pipeline steps
 composes a space to be passed in the GridsearchPipelineSelector
 or HyperoptPipelineSelector classes.
 A classic list of steps would be: [encoders, transformers, selectors, models]
"""
from sklearn.pipeline import Pipeline
from hyperopt import hp
from itertools import product


class SpaceComposer:
    """
    A class that from a given list of pipeline steps
    composes a space to be passed to GridsearchPipelineSelector
    or HyperoptPipelineSelector.
    """
    def compose_gridsearch_space(self, step_list: list) -> list:
        """
        Composes a hyperparameter space for input to the
        GridsearchPipelineSelector class.

        :param step_list: a classic list of steps would be
        [encoders, transformers, selectors, models],
        where, for example, selectors is a list
        of sklearn feature selectors, each selector given as a dict:
        for example {"name": "kbest",
                     "object": SelectPercentile(),
                     "params": {
                             "percentile":
                                 [5, 10, 20],
                             "score_func":
                                 [f_classif, chi2, mutual_info_classif]}}

        :return: a list of dictionaries of form
            {"name": NAME, "pipeline": PIPELINE, "params": PARAMS}
        """
        space = []

        step_combinations = product(*[step for step in
                                      step_list if len(step) > 0])

        for step_combination in step_combinations:

            space_element = {}

            space_element["name"] = "_".join([step["name"]
                                              for step in step_combination])

            space_element["pipeline"] = Pipeline(
                    [(step["name"], step["object"])
                     for step in step_combination])

            space_element["params"] =\
                {step["name"] + "__" + param_name: param_dist
                 for step in step_combination
                 for param_name, param_dist
                 in step["params"].items()}

            space.append(space_element)

        return space

    def compose_hyperopt_space(self, step_list: list) -> hp.choice:
        """
        Composes a hyperopt space from a list of steps.
        A classic list of steps would be
        [encoders, transformers, selectors, models],
        where, for example, selectors is a list
        of sklearn feature selectors, each selector given as a dict:
        for example {"name": "kbest",
                     "object": SelectPercentile(),
                     "params": {
                             "percentile":
                                 3 + hp.randint("kbest__percentile", 200),
                             "score_func":
                                 hp.choice("kbest__score_func",
                                    [f_classif, chi2, mutual_info_classif])}}
        """
        return hp.choice("pipelines", self.compose_gridsearch_space(step_list))