67 lines
2.5 KiB
Python
67 lines
2.5 KiB
Python
|
from numbers import Number
|
||
|
|
import math
|
||
|
|
import torch
|
||
|
|
from torch.distributions import constraints
|
||
|
|
from torch.distributions.uniform import Uniform
|
||
|
|
from torch.distributions.transformed_distribution import TransformedDistribution
|
||
|
|
from torch.distributions.transforms import AffineTransform, ExpTransform
|
||
|
|
from torch.distributions.utils import broadcast_all, euler_constant
|
||
|
|
|
||
|
|
|
||
|
|
class Gumbel(TransformedDistribution):
|
||
|
|
r"""
|
||
|
|
Samples from a Gumbel Distribution.
|
||
|
|
|
||
|
|
Examples::
|
||
|
|
|
||
|
|
>>> m = Gumbel(torch.tensor([1.0]), torch.tensor([2.0]))
|
||
|
|
>>> m.sample() # sample from Gumbel distribution with loc=1, scale=2
|
||
|
|
tensor([ 1.0124])
|
||
|
|
|
||
|
|
Args:
|
||
|
|
loc (float or Tensor): Location parameter of the distribution
|
||
|
|
scale (float or Tensor): Scale parameter of the distribution
|
||
|
|
"""
|
||
|
|
arg_constraints = {'loc': constraints.real, 'scale': constraints.positive}
|
||
|
|
support = constraints.real
|
||
|
|
|
||
|
|
def __init__(self, loc, scale, validate_args=None):
|
||
|
|
self.loc, self.scale = broadcast_all(loc, scale)
|
||
|
|
finfo = torch.finfo(self.loc.dtype)
|
||
|
|
if isinstance(loc, Number) and isinstance(scale, Number):
|
||
|
|
base_dist = Uniform(finfo.tiny, 1 - finfo.eps)
|
||
|
|
else:
|
||
|
|
base_dist = Uniform(torch.full_like(self.loc, finfo.tiny),
|
||
|
|
torch.full_like(self.loc, 1 - finfo.eps))
|
||
|
|
transforms = [ExpTransform().inv, AffineTransform(loc=0, scale=-torch.ones_like(self.scale)),
|
||
|
|
ExpTransform().inv, AffineTransform(loc=loc, scale=-self.scale)]
|
||
|
|
super(Gumbel, self).__init__(base_dist, transforms, validate_args=validate_args)
|
||
|
|
|
||
|
|
def expand(self, batch_shape, _instance=None):
|
||
|
|
new = self._get_checked_instance(Gumbel, _instance)
|
||
|
|
new.loc = self.loc.expand(batch_shape)
|
||
|
|
new.scale = self.scale.expand(batch_shape)
|
||
|
|
return super(Gumbel, self).expand(batch_shape, _instance=new)
|
||
|
|
|
||
|
|
# Explicitly defining the log probability function for Gumbel due to precision issues
|
||
|
|
def log_prob(self, value):
|
||
|
|
if self._validate_args:
|
||
|
|
self._validate_sample(value)
|
||
|
|
y = (self.loc - value) / self.scale
|
||
|
|
return (y - y.exp()) - self.scale.log()
|
||
|
|
|
||
|
|
@property
|
||
|
|
def mean(self):
|
||
|
|
return self.loc + self.scale * euler_constant
|
||
|
|
|
||
|
|
@property
|
||
|
|
def stddev(self):
|
||
|
|
return (math.pi / math.sqrt(6)) * self.scale
|
||
|
|
|
||
|
|
@property
|
||
|
|
def variance(self):
|
||
|
|
return self.stddev.pow(2)
|
||
|
|
|
||
|
|
def entropy(self):
|
||
|
|
return self.scale.log() + (1 + euler_constant)
|