import abc
import copy
import decimal
import json
from numbers import Real
from typing import Any
from typing import cast
from typing import Dict
from typing import Sequence
from typing import Union
import warnings
import numpy as np
from optuna._deprecated import deprecated_class
CategoricalChoiceType = Union[None, bool, int, float, str]
_float_distribution_deprecated_msg = (
"Use :class:`~optuna.distributions.FloatDistribution` instead."
)
_int_distribution_deprecated_msg = "Use :class:`~optuna.distributions.IntDistribution` instead."
class BaseDistribution(object, metaclass=abc.ABCMeta):
"""Base class for distributions.
Note that distribution classes are not supposed to be called by library users.
They are used by :class:`~optuna.trial.Trial` and :class:`~optuna.samplers` internally.
"""
def to_external_repr(self, param_value_in_internal_repr: float) -> Any:
"""Convert internal representation of a parameter value into external representation.
Args:
param_value_in_internal_repr:
Optuna's internal representation of a parameter value.
Returns:
Optuna's external representation of a parameter value.
"""
return param_value_in_internal_repr
def to_internal_repr(self, param_value_in_external_repr: Any) -> float:
"""Convert external representation of a parameter value into internal representation.
Args:
param_value_in_external_repr:
Optuna's external representation of a parameter value.
Returns:
Optuna's internal representation of a parameter value.
"""
return param_value_in_external_repr
@abc.abstractmethod
def single(self) -> bool:
"""Test whether the range of this distribution contains just a single value.
Returns:
:obj:`True` if the range of this distribution contains just a single value,
otherwise :obj:`False`.
"""
raise NotImplementedError
@abc.abstractmethod
def _contains(self, param_value_in_internal_repr: float) -> bool:
"""Test if a parameter value is contained in the range of this distribution.
Args:
param_value_in_internal_repr:
Optuna's internal representation of a parameter value.
Returns:
:obj:`True` if the parameter value is contained in the range of this distribution,
otherwise :obj:`False`.
"""
raise NotImplementedError
def _asdict(self) -> Dict:
return self.__dict__
def __eq__(self, other: Any) -> bool:
if not isinstance(other, BaseDistribution):
return NotImplemented
if not type(self) is type(other):
return False
return self.__dict__ == other.__dict__
def __hash__(self) -> int:
return hash((self.__class__,) + tuple(sorted(self.__dict__.items())))
def __repr__(self) -> str:
kwargs = ", ".join("{}={}".format(k, v) for k, v in sorted(self._asdict().items()))
return "{}({})".format(self.__class__.__name__, kwargs)
[docs]class FloatDistribution(BaseDistribution):
"""A distribution on floats.
This object is instantiated by :func:`~optuna.trial.Trial.suggest_float`, and passed to
:mod:`~optuna.samplers` in general.
.. note::
When ``step`` is not :obj:`None`, if the range :math:`[\\mathsf{low}, \\mathsf{high}]`
is not divisible by :math:`\\mathsf{step}`, :math:`\\mathsf{high}` will be replaced
with the maximum of :math:`k \\times \\mathsf{step} + \\mathsf{low} < \\mathsf{high}`,
where :math:`k` is an integer.
Attributes:
low:
Lower endpoint of the range of the distribution. ``low`` is included in the range.
``low`` must be less than or equal to ``high``. If ``log`` is :obj:`True`,
``low`` must be larger than 0.
high:
Upper endpoint of the range of the distribution. ``high`` is included in the range.
``high`` must be greater than or equal to ``low``.
log:
If ``log`` is :obj:`True`, this distribution is in log-scaled domain.
This parameter must be :obj:`False` when the parameter ``step`` is not :obj:`None`.
step:
A discretization step. ``step`` must be larger than 0.
This parameter must be :obj:`None` when the parameter ``log`` is :obj:`True`.
"""
def __init__(
self, low: float, high: float, log: bool = False, step: Union[None, float] = None
) -> None:
if log and step is not None:
raise ValueError("The parameter `step` is not supported when `log` is true.")
if low > high:
raise ValueError(
"The `low` value must be smaller than or equal to the `high` value "
"(low={}, high={}).".format(low, high)
)
if log and low <= 0.0:
raise ValueError(
"The `low` value must be larger than 0 for a log distribution "
"(low={}, high={}).".format(low, high)
)
if step is not None and step <= 0:
raise ValueError(
"The `step` value must be non-zero positive value, " "but step={}.".format(step)
)
self.step = None
if step is not None:
high = _adjust_discrete_uniform_high(low, high, step)
self.step = float(step)
self.low = float(low)
self.high = float(high)
self.log = log
[docs] def single(self) -> bool:
if self.step is None:
return self.low == self.high
else:
if self.low == self.high:
return True
high = decimal.Decimal(str(self.high))
low = decimal.Decimal(str(self.low))
step = decimal.Decimal(str(self.step))
return (high - low) < step
def _contains(self, param_value_in_internal_repr: float) -> bool:
value = param_value_in_internal_repr
if self.step is None:
return self.low <= value <= self.high
else:
k = (value - self.low) / self.step
return self.low <= value <= self.high and abs(k - round(k)) < 1.0e-8
[docs]class IntDistribution(BaseDistribution):
"""A distribution on integers.
This object is instantiated by :func:`~optuna.trial.Trial.suggest_int`, and passed to
:mod:`~optuna.samplers` in general.
.. note::
When ``step`` is not :obj:`None`, if the range :math:`[\\mathsf{low}, \\mathsf{high}]`
is not divisible by :math:`\\mathsf{step}`, :math:`\\mathsf{high}` will be replaced
with the maximum of :math:`k \\times \\mathsf{step} + \\mathsf{low} < \\mathsf{high}`,
where :math:`k` is an integer.
Attributes:
low:
Lower endpoint of the range of the distribution. ``low`` is included in the range.
``low`` must be less than or equal to ``high``. If ``log`` is :obj:`True`,
``low`` must be larger than or equal to 1.
high:
Upper endpoint of the range of the distribution. ``high`` is included in the range.
``high`` must be greater than or equal to ``low``.
log:
If ``log`` is :obj:`True`, this distribution is in log-scaled domain.
This parameter must be :obj:`False` when the parameter ``step`` is not 1.
step:
A discretization step. ``step`` must be a positive integer. This parameter must be 1
when the parameter ``log`` is :obj:`True`.
"""
def __init__(self, low: int, high: int, log: bool = False, step: int = 1) -> None:
if log and step != 1:
raise ValueError(
"Samplers and other components in Optuna only accept step is 1 "
"when `log` argument is True."
)
if low > high:
raise ValueError(
"The `low` value must be smaller than or equal to the `high` value "
"(low={}, high={}).".format(low, high)
)
if log and low < 1:
raise ValueError(
"The `low` value must be equal to or greater than 1 for a log distribution "
"(low={}, high={}).".format(low, high)
)
if step <= 0:
raise ValueError(
"The `step` value must be non-zero positive value, but step={}.".format(step)
)
self.log = log
self.step = int(step)
self.low = int(low)
high = int(high)
self.high = _adjust_int_uniform_high(self.low, high, self.step)
[docs] def to_external_repr(self, param_value_in_internal_repr: float) -> int:
return int(param_value_in_internal_repr)
[docs] def to_internal_repr(self, param_value_in_external_repr: int) -> float:
return float(param_value_in_external_repr)
[docs] def single(self) -> bool:
if self.log:
return self.low == self.high
if self.low == self.high:
return True
return (self.high - self.low) < self.step
def _contains(self, param_value_in_internal_repr: float) -> bool:
value = param_value_in_internal_repr
return self.low <= value <= self.high and (value - self.low) % self.step == 0
def _categorical_choice_equal(
value1: CategoricalChoiceType, value2: CategoricalChoiceType
) -> bool:
"""A function to check two choices equal considering NaN.
This function can handle NaNs like np.float32("nan") other than float.
"""
value1_is_nan = isinstance(value1, Real) and np.isnan(float(value1))
value2_is_nan = isinstance(value2, Real) and np.isnan(float(value2))
return (value1 == value2) or (value1_is_nan and value2_is_nan)
[docs]class CategoricalDistribution(BaseDistribution):
"""A categorical distribution.
This object is instantiated by :func:`~optuna.trial.Trial.suggest_categorical`, and
passed to :mod:`~optuna.samplers` in general.
Args:
choices:
Parameter value candidates. ``choices`` must have one element at least.
.. note::
Not all types are guaranteed to be compatible with all storages. It is recommended to
restrict the types of the choices to :obj:`None`, :class:`bool`, :class:`int`,
:class:`float` and :class:`str`.
Attributes:
choices:
Parameter value candidates.
"""
def __init__(self, choices: Sequence[CategoricalChoiceType]) -> None:
if len(choices) == 0:
raise ValueError("The `choices` must contain one or more elements.")
for choice in choices:
if choice is not None and not isinstance(choice, (bool, int, float, str)):
message = (
"Choices for a categorical distribution should be a tuple of None, bool, "
"int, float and str for persistent storage but contains {} which is of type "
"{}.".format(choice, type(choice).__name__)
)
warnings.warn(message)
self.choices = tuple(choices)
[docs] def to_external_repr(self, param_value_in_internal_repr: float) -> CategoricalChoiceType:
return self.choices[int(param_value_in_internal_repr)]
[docs] def to_internal_repr(self, param_value_in_external_repr: CategoricalChoiceType) -> float:
for index, choice in enumerate(self.choices):
if _categorical_choice_equal(param_value_in_external_repr, choice):
return index
raise ValueError(f"'{param_value_in_external_repr}' not in {self.choices}.")
[docs] def single(self) -> bool:
return len(self.choices) == 1
def _contains(self, param_value_in_internal_repr: float) -> bool:
index = int(param_value_in_internal_repr)
return 0 <= index < len(self.choices)
def __eq__(self, other: Any) -> bool:
if not isinstance(other, BaseDistribution):
return NotImplemented
if not isinstance(other, self.__class__):
return False
if self.__dict__.keys() != other.__dict__.keys():
return False
for key, value in self.__dict__.items():
if key == "choices":
if len(value) != len(getattr(other, key)):
return False
for choice, other_choice in zip(value, getattr(other, key)):
if not _categorical_choice_equal(choice, other_choice):
return False
else:
if value != getattr(other, key):
return False
return True
__hash__ = BaseDistribution.__hash__
DISTRIBUTION_CLASSES = (
IntDistribution,
IntLogUniformDistribution,
IntUniformDistribution,
FloatDistribution,
UniformDistribution,
LogUniformDistribution,
DiscreteUniformDistribution,
CategoricalDistribution,
)
[docs]def json_to_distribution(json_str: str) -> BaseDistribution:
"""Deserialize a distribution in JSON format.
Args:
json_str: A JSON-serialized distribution.
Returns:
A deserialized distribution.
"""
json_dict = json.loads(json_str)
if "name" in json_dict:
if json_dict["name"] == CategoricalDistribution.__name__:
json_dict["attributes"]["choices"] = tuple(json_dict["attributes"]["choices"])
for cls in DISTRIBUTION_CLASSES:
if json_dict["name"] == cls.__name__:
return cls(**json_dict["attributes"])
raise ValueError("Unknown distribution class: {}".format(json_dict["name"]))
else:
# Deserialize a distribution from an abbreviated format.
if json_dict["type"] == "categorical":
return CategoricalDistribution(json_dict["choices"])
elif json_dict["type"] in ("float", "int"):
low = json_dict["low"]
high = json_dict["high"]
step = json_dict.get("step")
log = json_dict.get("log", False)
if json_dict["type"] == "float":
return FloatDistribution(low, high, log=log, step=step)
else:
if step is None:
step = 1
return IntDistribution(low=low, high=high, log=log, step=step)
raise ValueError("Unknown distribution type: {}".format(json_dict["type"]))
[docs]def distribution_to_json(dist: BaseDistribution) -> str:
"""Serialize a distribution to JSON format.
Args:
dist: A distribution to be serialized.
Returns:
A JSON string of a given distribution.
"""
return json.dumps({"name": dist.__class__.__name__, "attributes": dist._asdict()})
[docs]def check_distribution_compatibility(
dist_old: BaseDistribution, dist_new: BaseDistribution
) -> None:
"""A function to check compatibility of two distributions.
It checks whether ``dist_old`` and ``dist_new`` are the same kind of distributions.
If ``dist_old`` is :class:`~optuna.distributions.CategoricalDistribution`,
it further checks ``choices`` are the same between ``dist_old`` and ``dist_new``.
Note that this method is not supposed to be called by library users.
Args:
dist_old:
A distribution previously recorded in storage.
dist_new:
A distribution newly added to storage.
"""
if dist_old.__class__ != dist_new.__class__:
raise ValueError("Cannot set different distribution kind to the same parameter name.")
if isinstance(dist_old, (FloatDistribution, IntDistribution)):
# For mypy.
assert isinstance(dist_new, (FloatDistribution, IntDistribution))
if dist_old.log != dist_new.log:
raise ValueError("Cannot set different log configuration to the same parameter name.")
if not isinstance(dist_old, CategoricalDistribution):
return
if not isinstance(dist_new, CategoricalDistribution):
return
if dist_old != dist_new:
raise ValueError(
CategoricalDistribution.__name__ + " does not support dynamic value space."
)
def _adjust_discrete_uniform_high(low: float, high: float, q: float) -> float:
d_high = decimal.Decimal(str(high))
d_low = decimal.Decimal(str(low))
d_q = decimal.Decimal(str(q))
d_r = d_high - d_low
if d_r % d_q != decimal.Decimal("0"):
old_high = high
high = float((d_r // d_q) * d_q + d_low)
warnings.warn(
"The distribution is specified by [{low}, {old_high}] and q={step}, but the range "
"is not divisible by `q`. It will be replaced by [{low}, {high}].".format(
low=low, old_high=old_high, high=high, step=q
)
)
return high
def _adjust_int_uniform_high(low: int, high: int, step: int) -> int:
r = high - low
if r % step != 0:
old_high = high
high = r // step * step + low
warnings.warn(
"The distribution is specified by [{low}, {old_high}] and step={step}, but the range "
"is not divisible by `step`. It will be replaced by [{low}, {high}].".format(
low=low, old_high=old_high, high=high, step=step
)
)
return high
def _get_single_value(distribution: BaseDistribution) -> Union[int, float, CategoricalChoiceType]:
assert distribution.single()
if isinstance(
distribution,
(
FloatDistribution,
IntDistribution,
),
):
return distribution.low
elif isinstance(distribution, CategoricalDistribution):
return distribution.choices[0]
assert False
# TODO(himkt): Remove this method with the deletion of deprecated distributions.
# https://github.com/optuna/optuna/issues/2941
def _convert_old_distribution_to_new_distribution(
distribution: BaseDistribution,
suppress_warning: bool = False,
) -> BaseDistribution:
new_distribution: BaseDistribution
# Float distributions.
if isinstance(distribution, UniformDistribution):
new_distribution = FloatDistribution(
low=distribution.low,
high=distribution.high,
log=False,
step=None,
)
elif isinstance(distribution, LogUniformDistribution):
new_distribution = FloatDistribution(
low=distribution.low,
high=distribution.high,
log=True,
step=None,
)
elif isinstance(distribution, DiscreteUniformDistribution):
new_distribution = FloatDistribution(
low=distribution.low,
high=distribution.high,
log=False,
step=distribution.q,
)
# Integer distributions.
elif isinstance(distribution, IntUniformDistribution):
new_distribution = IntDistribution(
low=distribution.low,
high=distribution.high,
log=False,
step=distribution.step,
)
elif isinstance(distribution, IntLogUniformDistribution):
new_distribution = IntDistribution(
low=distribution.low,
high=distribution.high,
log=True,
step=distribution.step,
)
# Categorical distribution.
else:
new_distribution = distribution
if new_distribution != distribution and not suppress_warning:
message = (
f"{distribution} is deprecated and internally converted to"
f" {new_distribution}. See https://github.com/optuna/optuna/issues/2941."
)
warnings.warn(message, FutureWarning)
return new_distribution