Source code for ax.modelbridge.transforms.remove_fixed

#!/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, Union

from ax.core.observation import Observation, ObservationFeatures
from ax.core.parameter import ChoiceParameter, FixedParameter, RangeParameter
from ax.core.search_space import SearchSpace
from ax.modelbridge.transforms.base import Transform
from ax.modelbridge.transforms.utils import construct_new_search_space
from ax.models.types import TConfig

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


[docs]class RemoveFixed(Transform):
    """Remove fixed parameters.

    Fixed parameters should not be included in the SearchSpace.
    This transform removes these parameters, leaving only tunable parameters.

    Transform is done in-place for observation features.
    """

    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, "RemoveFixed requires search space"
        # Identify parameters that should be transformed
        self.fixed_parameters: Dict[str, FixedParameter] = {
            p_name: p
            for p_name, p in search_space.parameters.items()
            if isinstance(p, FixedParameter)
        }

[docs]    def transform_observation_features(
        self, observation_features: List[ObservationFeatures]
    ) -> List[ObservationFeatures]:
        for obsf in observation_features:
            for p_name, fixed_p in self.fixed_parameters.items():
                if p_name in obsf.parameters:
                    if obsf.parameters[p_name] != fixed_p.value:
                        raise ValueError(
                            f"Fixed parameter {p_name} with out of design value: "
                            f"{obsf.parameters[p_name]} passed to `RemoveFixed`."
                        )
                    obsf.parameters.pop(p_name)
        return observation_features

    def _transform_search_space(self, search_space: SearchSpace) -> SearchSpace:
        tunable_parameters: List[Union[ChoiceParameter, RangeParameter]] = []
        for p in search_space.parameters.values():
            if p.name not in self.fixed_parameters:
                # If it's not in fixed_parameters, it must be a tunable param.
                # pyre: p_ is declared to have type `Union[ChoiceParameter,
                # pyre: RangeParameter]` but is used as type `ax.core.
                # pyre-fixme[9]: parameter.Parameter`.
                p_: Union[ChoiceParameter, RangeParameter] = p
                tunable_parameters.append(p_)
        return construct_new_search_space(
            search_space=search_space,
            # pyre-ignore Incompatible parameter type [6]
            parameters=tunable_parameters,
            parameter_constraints=[
                pc.clone() 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, p in self.fixed_parameters.items():
                obsf.parameters[p_name] = p.value
        return observation_features