import numpy as np
from syne_tune.config_space import Domain, is_log_space, is_reverse_log_space
[docs]
class Scaling:
[docs]
def to_internal(self, value: float) -> float:
raise NotImplementedError
[docs]
def from_internal(self, value: float) -> float:
raise NotImplementedError
def __repr__(self):
return "{}()".format(self.__class__.__name__)
def __eq__(self, other):
# For usage in tests. Make sure to edit if parameters are added.
return self.__class__ == other.__class__
[docs]
class LinearScaling(Scaling):
[docs]
def to_internal(self, value: float) -> float:
return value
[docs]
def from_internal(self, value: float) -> float:
return value
[docs]
class LogScaling(Scaling):
[docs]
def to_internal(self, value: float) -> float:
assert value > 0, "Value must be strictly positive to be log-scaled."
return np.log(value)
[docs]
def from_internal(self, value: float) -> float:
return np.exp(value)
[docs]
class ReverseLogScaling(Scaling):
[docs]
def to_internal(self, value: float) -> float:
assert (
0 <= value < 1
), "Value must be between 0 (inclusive) and 1 (exclusive) to be reverse-log-scaled."
return -np.log(1.0 - value)
[docs]
def from_internal(self, value: float) -> float:
return 1.0 - np.exp(-value)
[docs]
def get_scaling(hp_range: Domain) -> Scaling:
if is_log_space(hp_range):
return LogScaling()
elif is_reverse_log_space(hp_range):
return ReverseLogScaling()
else:
return LinearScaling()