pytorch

1
# mypy: allow-untyped-defs
2
import math
3
from numbers import Number
4

5
import torch
6
from torch.distributions import constraints
7
from torch.distributions.transformed_distribution import TransformedDistribution
8
from torch.distributions.transforms import AffineTransform, ExpTransform
9
from torch.distributions.uniform import Uniform
10
from torch.distributions.utils import broadcast_all, euler_constant
11

12

13
__all__ = ["Gumbel"]
14

15

16
class Gumbel(TransformedDistribution):
17
    r"""
18
    Samples from a Gumbel Distribution.
19

20
    Examples::
21

22
        >>> # xdoctest: +IGNORE_WANT("non-deterministic")
23
        >>> m = Gumbel(torch.tensor([1.0]), torch.tensor([2.0]))
24
        >>> m.sample()  # sample from Gumbel distribution with loc=1, scale=2
25
        tensor([ 1.0124])
26

27
    Args:
28
        loc (float or Tensor): Location parameter of the distribution
29
        scale (float or Tensor): Scale parameter of the distribution
30
    """
31
    arg_constraints = {"loc": constraints.real, "scale": constraints.positive}
32
    support = constraints.real
33

34
    def __init__(self, loc, scale, validate_args=None):
35
        self.loc, self.scale = broadcast_all(loc, scale)
36
        finfo = torch.finfo(self.loc.dtype)
37
        if isinstance(loc, Number) and isinstance(scale, Number):
38
            base_dist = Uniform(finfo.tiny, 1 - finfo.eps, validate_args=validate_args)
39
        else:
40
            base_dist = Uniform(
41
                torch.full_like(self.loc, finfo.tiny),
42
                torch.full_like(self.loc, 1 - finfo.eps),
43
                validate_args=validate_args,
44
            )
45
        transforms = [
46
            ExpTransform().inv,
47
            AffineTransform(loc=0, scale=-torch.ones_like(self.scale)),
48
            ExpTransform().inv,
49
            AffineTransform(loc=loc, scale=-self.scale),
50
        ]
51
        super().__init__(base_dist, transforms, validate_args=validate_args)
52

53
    def expand(self, batch_shape, _instance=None):
54
        new = self._get_checked_instance(Gumbel, _instance)
55
        new.loc = self.loc.expand(batch_shape)
56
        new.scale = self.scale.expand(batch_shape)
57
        return super().expand(batch_shape, _instance=new)
58

59
    # Explicitly defining the log probability function for Gumbel due to precision issues
60
    def log_prob(self, value):
61
        if self._validate_args:
62
            self._validate_sample(value)
63
        y = (self.loc - value) / self.scale
64
        return (y - y.exp()) - self.scale.log()
65

66
    @property
67
    def mean(self):
68
        return self.loc + self.scale * euler_constant
69

70
    @property
71
    def mode(self):
72
        return self.loc
73

74
    @property
75
    def stddev(self):
76
        return (math.pi / math.sqrt(6)) * self.scale
77

78
    @property
79
    def variance(self):
80
        return self.stddev.pow(2)
81

82
    def entropy(self):
83
        return self.scale.log() + (1 + euler_constant)
84