Source code for ax.modelbridge.transforms.one_hot

#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

from typing import Dict, List, Optional, TYPE_CHECKING, TypeVar

import numpy as np
from ax.core.observation import Observation, ObservationFeatures
from ax.core.parameter import ChoiceParameter, Parameter, ParameterType, RangeParameter
from ax.core.search_space import SearchSpace
from ax.core.types import TParameterization
from ax.modelbridge.transforms.base import Transform
from ax.modelbridge.transforms.rounding import (
    randomized_onehot_round,
    strict_onehot_round,
)
from ax.modelbridge.transforms.utils import construct_new_search_space
from ax.models.types import TConfig
from sklearn.preprocessing import LabelBinarizer, LabelEncoder

if TYPE_CHECKING:
    # import as module to make sphinx-autodoc-typehints happy
    from ax import modelbridge as modelbridge_module  # noqa F401  # pragma: no cover


OH_PARAM_INFIX = "_OH_PARAM_"
T = TypeVar("T")


[docs]class OneHotEncoder:
    """Joins the two encoders needed for OneHot transform."""

    int_encoder: LabelEncoder
    label_binarizer: LabelBinarizer

    def __init__(self, values: List[T]) -> None:
        self.int_encoder = LabelEncoder().fit(values)
        self.label_binarizer = LabelBinarizer().fit(self.int_encoder.transform(values))

[docs]    def transform(self, labels: List[T]) -> np.ndarray:
        """One hot encode a list of labels."""
        return self.label_binarizer.transform(self.int_encoder.transform(labels))

[docs]    def inverse_transform(self, encoded_labels: List[T]) -> List[T]:
        """Inverse transorm a list of one hot encoded labels."""
        return self.int_encoder.inverse_transform(
            self.label_binarizer.inverse_transform(encoded_labels)
        )

    @property
    def classes(self) -> np.ndarray:
        """Return number of classes discovered while fitting transform."""
        # pyre-fixme[16]: `LabelBinarizer` has no attribute `classes_`.
        return self.label_binarizer.classes_


[docs]class OneHot(Transform):
    """Convert categorical parameters (unordered ChoiceParameters) to
    one-hot-encoded parameters.

    Does not convert task parameters.

    Parameters will be one-hot-encoded, yielding a set of RangeParameters,
    of type float, on [0, 1]. If there are two values, one single RangeParameter
    will be yielded, otherwise there will be a new RangeParameter for each
    ChoiceParameter value.

    In the reverse transform, floats can be converted to a one-hot encoded vector
    using one of two methods:

    Strict rounding: Choose the maximum value. With levels ['a', 'b', 'c'] and
        float values [0.2, 0.4, 0.3], the restored parameter would be set to 'b'.
        Ties are broken randomly, so values [0.2, 0.4, 0.4] is randomly set to 'b'
        or 'c'.

    Randomized rounding: Sample from the distribution. Float values
        [0.2, 0.4, 0.3] are transformed to 'a' w.p.
        0.2/0.9, 'b' w.p. 0.4/0.9, or 'c' w.p. 0.3/0.9.

    Type of rounding can be set using transform_config['rounding'] to either
    'strict' or 'randomized'. Defaults to strict.

    Transform is done in-place.
    """

    def __init__(
        self,
        search_space: Optional[SearchSpace] = None,
        observations: Optional[List[Observation]] = None,
        modelbridge: Optional["modelbridge_module.base.ModelBridge"] = None,
        config: Optional[TConfig] = None,
    ) -> None:
        assert search_space is not None, "OneHot requires search space"
        # Identify parameters that should be transformed
        # pyre-fixme[4]: Attribute must be annotated.
        self.rounding = "strict"
        if config is not None:
            self.rounding = config.get("rounding", "strict")
        self.encoder: Dict[str, OneHotEncoder] = {}
        self.encoded_parameters: Dict[str, List[str]] = {}
        for p in search_space.parameters.values():
            if isinstance(p, ChoiceParameter) and not p.is_ordered and not p.is_task:
                self.encoder[p.name] = OneHotEncoder(p.values)
                nc = len(self.encoder[p.name].classes)
                if nc == 2:
                    # Two levels handled in one parameter
                    self.encoded_parameters[p.name] = [p.name + OH_PARAM_INFIX]
                else:
                    self.encoded_parameters[p.name] = [
                        "{}{}_{}".format(p.name, OH_PARAM_INFIX, i) for i in range(nc)
                    ]

[docs]    def transform_observation_features(
        self, observation_features: List[ObservationFeatures]
    ) -> List[ObservationFeatures]:
        for obsf in observation_features:
            for p_name, encoder in self.encoder.items():
                if p_name in obsf.parameters:
                    vals = encoder.transform(labels=[obsf.parameters.pop(p_name)])[0]
                    updated_parameters: TParameterization = {
                        self.encoded_parameters[p_name][i]: v
                        for i, v in enumerate(vals)
                    }
                    obsf.parameters.update(updated_parameters)
        return observation_features

    def _transform_search_space(self, search_space: SearchSpace) -> SearchSpace:
        transformed_parameters: Dict[str, Parameter] = {}
        for p_name, p in search_space.parameters.items():
            if p_name in self.encoded_parameters:
                if p.is_fidelity:
                    raise ValueError(
                        f"Cannot one-hot-encode fidelity parameter {p_name}"
                    )
                for new_p_name in self.encoded_parameters[p_name]:
                    transformed_parameters[new_p_name] = RangeParameter(
                        name=new_p_name,
                        parameter_type=ParameterType.FLOAT,
                        lower=0,
                        upper=1,
                    )
            else:
                transformed_parameters[p_name] = p
        return construct_new_search_space(
            search_space=search_space,
            parameters=list(transformed_parameters.values()),
            parameter_constraints=[
                pc.clone_with_transformed_parameters(
                    transformed_parameters=transformed_parameters
                )
                for pc in search_space.parameter_constraints
            ],
        )

[docs]    def untransform_observation_features(
        self, observation_features: List[ObservationFeatures]
    ) -> List[ObservationFeatures]:
        for obsf in observation_features:
            for p_name in self.encoder.keys():
                has_params = [
                    p in obsf.parameters for p in self.encoded_parameters[p_name]
                ]
                if not all(has_params):
                    if any(has_params):  # pragma: no cover
                        raise ValueError(  # pragma: no cover
                            f"Missing some parameters for {p_name}"
                        )
                    continue  # pragma: no cover
                x = np.array(
                    [obsf.parameters.pop(p) for p in self.encoded_parameters[p_name]]
                )
                if self.rounding == "strict":
                    x = strict_onehot_round(x)
                else:
                    x = randomized_onehot_round(x)
                val = self.encoder[p_name].inverse_transform(encoded_labels=x[None, :])[
                    0
                ]
                if isinstance(val, np.str_):
                    val = str(val)
                if isinstance(val, np.bool_):
                    val = bool(val)  # Numpy bools don't serialize
                obsf.parameters[p_name] = val
        return observation_features