#!/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, Tuple, TYPE_CHECKING
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 TParamValue
from ax.modelbridge.transforms.base import Transform
from ax.modelbridge.transforms.utils import (
ClosestLookupDict,
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 ChoiceEncode(Transform):
"""Convert general ChoiceParameters to Float ChoiceParameters.
ChoiceParameters will be transformed to ChoiceParameters of float or int type.
The resulting choice parameter will be considered ordered iff the original
parameter is.
If the parameter type is numeric (int, float) and the parameter is orderd,
then the values are normalized to the unit interval while retaining relative
spacing. If the parameter type is unordered (categorical) or ordered but
non-numeric, this transform uses an integer encoding.
In the inverse transform, parameters will be mapped back onto the original domain.
This transform does not transform task parameters (use TaskEncode for this).
Note that this behavior is different from that of OrderedChoiceEncode, which
transforms (ordered) ChoiceParameters to integer RangeParameters (rather than
ChoiceParameters).
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, "ChoiceEncode requires search space"
# Identify parameters that should be transformed
self.encoded_parameters: Dict[str, Dict[TParamValue, TParamValue]] = {}
self.encoded_parameters_inverse: Dict[str, ClosestLookupDict] = {}
for p in search_space.parameters.values():
if isinstance(p, ChoiceParameter) and not p.is_task:
transformed_values, _ = transform_choice_values(p)
self.encoded_parameters[p.name] = dict(
zip(p.values, transformed_values)
)
self.encoded_parameters_inverse[p.name] = ClosestLookupDict(
zip(transformed_values, p.values)
)
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 and isinstance(p, ChoiceParameter):
if p.is_fidelity:
raise ValueError( # pragma: no cover
f"Cannot choice-encode fidelity parameter {p_name}"
)
tvals, ptype = transform_choice_values(p)
transformed_parameters[p_name] = ChoiceParameter(
name=p_name,
parameter_type=ptype,
values=tvals.tolist(),
is_ordered=p.is_ordered,
)
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]class OrderedChoiceEncode(ChoiceEncode):
"""Convert ordered ChoiceParameters to integer RangeParameters.
Parameters will be transformed to an integer RangeParameters, mapped from the
original choice domain to a contiguous range `0, 1, ..., n_choices - 1`
of integers. Does not transform task parameters.
In the inverse transform, parameters will be mapped back onto the original domain.
In order to encode all ChoiceParameters (not just ordered ChoiceParameters),
use ChoiceEncode instead.
Transform is done in-place.
"""
def __init__(
self,
search_space: SearchSpace,
observations: List[Observation],
modelbridge: Optional["modelbridge_module.base.ModelBridge"] = None,
config: Optional[TConfig] = None,
) -> None:
# Identify parameters that should be transformed
self.encoded_parameters: Dict[str, Dict[TParamValue, int]] = {}
for p in search_space.parameters.values():
if isinstance(p, ChoiceParameter) and p.is_ordered and not p.is_task:
self.encoded_parameters[p.name] = {
original_value: transformed_value
for transformed_value, original_value in enumerate(p.values)
}
self.encoded_parameters_inverse: Dict[str, Dict[int, TParamValue]] = {
p_name: {
transformed_value: original_value
for original_value, transformed_value in transforms.items()
}
for p_name, transforms in self.encoded_parameters.items()
}
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 and isinstance(p, ChoiceParameter):
if p.is_fidelity:
raise ValueError(
f"Cannot choice-encode fidelity parameter {p_name}"
)
# Choice(|K|) => Range(0, K-1)
transformed_parameters[p_name] = RangeParameter(
name=p_name,
parameter_type=ParameterType.INT,
lower=0,
upper=len(p.values) - 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
],
)
