from __future__ import annotations
from typing import Any
import warnings
import numpy as np
from smac.model.gaussian_process.priors.abstract_prior import AbstractPrior
__copyright__ = "Copyright 2022, automl.org"
__license__ = "3-clause BSD"
[docs]class TophatPrior(AbstractPrior):
"""Tophat prior as it used in the original spearmint code.
Parameters
----------
lower_bound : float
Lower bound of the prior. In original scale.
upper_bound : float
Upper bound of the prior. In original scale.
seed : int, defaults to 0
"""
def __init__(
self,
lower_bound: float,
upper_bound: float,
seed: int = 0,
):
super().__init__(seed=seed)
self._min = lower_bound
self._log_min = np.log(lower_bound)
self._max = upper_bound
self._log_max = np.log(upper_bound)
if not (self._max > self._min):
raise Exception("Upper bound of Tophat prior must be greater than the lower bound.")
@property
def meta(self) -> dict[str, Any]: # noqa: D102
meta = super().meta
meta.update({"lower_bound": self._min, "upper_bound": self._max})
return meta
def _get_log_probability(self, theta: float) -> float:
if theta < self._min or theta > self._max:
return -np.inf
else:
return 0
def _sample_from_prior(self, n_samples: int) -> np.ndarray:
if np.ndim(n_samples) != 0:
raise ValueError("The argument `n_samples` needs to be a scalar (is %s)." % n_samples)
if n_samples <= 0:
raise ValueError("The argument `n_samples` needs to be positive (is %d)." % n_samples)
p0 = np.exp(self._rng.uniform(low=self._log_min, high=self._log_max, size=(n_samples,)))
return p0
def _get_gradient(self, theta: float) -> float:
return 0
[docs] def get_gradient(self, theta: float) -> float: # noqa: D102
return 0
[docs]class SoftTopHatPrior(AbstractPrior):
"""Soft Tophat prior as it used in the original spearmint code.
Parameters
----------
lower_bound : float
Lower bound of the prior. In original scale.
upper_bound : float
Upper bound of the prior. In original scale.
exponent : float
Exponent of the prior.
seed : int, defaults to 0
"""
def __init__(
self,
lower_bound: float,
upper_bound: float,
exponent: float,
seed: int = 0,
) -> None:
super().__init__(seed=seed)
with warnings.catch_warnings():
warnings.simplefilter("error")
self._lower_bound = lower_bound
try:
self._log_lower_bound = np.log(lower_bound)
except RuntimeWarning as w:
if "invalid value encountered in log" in w.args[0]:
raise ValueError("Invalid lower bound %f (cannot compute log)" % lower_bound)
raise w
self._upper_bound = upper_bound
try:
self._log_upper_bound = np.log(upper_bound)
except RuntimeWarning as w:
if "invalid value encountered in log" in w.args[0]:
raise ValueError("Invalid lower bound %f (cannot compute log)" % lower_bound)
raise w
if exponent <= 0:
raise ValueError("Exponent cannot be less or equal than zero (but is %f)" % exponent)
self._exponent = exponent
def __repr__(self) -> str:
return "SoftTopHatPrior(lower_bound=%f, upper_bound=%f)" % (
self._lower_bound,
self._upper_bound,
)
@property
def meta(self) -> dict[str, Any]: # noqa: D102
meta = super().meta
meta.update({"lower_bound": self._lower_bound, "upper_bound": self._upper_bound, "exponent": self._exponent})
return meta
[docs] def get_log_probability(self, theta: float) -> float: # noqa: D102
# We need to use lnprob here instead of _lnprob to have the squared function work
# in the logarithmic space, too.
if np.ndim(theta) == 0:
if theta < self._log_lower_bound:
return -((theta - self._log_lower_bound) ** self._exponent)
elif theta > self._log_upper_bound:
return -((self._log_upper_bound - theta) ** self._exponent)
else:
return 0
else:
raise NotImplementedError()
[docs] def get_gradient(self, theta: float) -> float: # noqa: D102
if np.ndim(theta) == 0:
if theta < self._log_lower_bound:
return -self._exponent * (theta - self._log_lower_bound)
elif theta > self._log_upper_bound:
return self._exponent * (self._log_upper_bound - theta)
else:
return 0
else:
raise NotImplementedError()
def _get_log_probability(self, theta: float) -> float:
return 0
def _get_gradient(self, theta: float) -> float:
return 0
def _sample_from_prior(self, n_samples: int) -> np.ndarray:
return np.exp(self._rng.uniform(self._log_lower_bound, self._log_upper_bound, size=(n_samples,)))