xref: /aosp_15_r20/external/pytorch/torch/distributions/kumaraswamy.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45) 
1# mypy: allow-untyped-defs
2import torch
3from torch import nan
4from torch.distributions import constraints
5from torch.distributions.transformed_distribution import TransformedDistribution
6from torch.distributions.transforms import AffineTransform, PowerTransform
7from torch.distributions.uniform import Uniform
8from torch.distributions.utils import broadcast_all, euler_constant
9
10
11__all__ = ["Kumaraswamy"]
12
13
14def _moments(a, b, n):
15    """
16    Computes nth moment of Kumaraswamy using using torch.lgamma
17    """
18    arg1 = 1 + n / a
19    log_value = torch.lgamma(arg1) + torch.lgamma(b) - torch.lgamma(arg1 + b)
20    return b * torch.exp(log_value)
21
22
23class Kumaraswamy(TransformedDistribution):
24    r"""
25    Samples from a Kumaraswamy distribution.
26
27    Example::
28
29        >>> # xdoctest: +IGNORE_WANT("non-deterministic")
30        >>> m = Kumaraswamy(torch.tensor([1.0]), torch.tensor([1.0]))
31        >>> m.sample()  # sample from a Kumaraswamy distribution with concentration alpha=1 and beta=1
32        tensor([ 0.1729])
33
34    Args:
35        concentration1 (float or Tensor): 1st concentration parameter of the distribution
36            (often referred to as alpha)
37        concentration0 (float or Tensor): 2nd concentration parameter of the distribution
38            (often referred to as beta)
39    """
40    arg_constraints = {
41        "concentration1": constraints.positive,
42        "concentration0": constraints.positive,
43    }
44    support = constraints.unit_interval
45    has_rsample = True
46
47    def __init__(self, concentration1, concentration0, validate_args=None):
48        self.concentration1, self.concentration0 = broadcast_all(
49            concentration1, concentration0
50        )
51        finfo = torch.finfo(self.concentration0.dtype)
52        base_dist = Uniform(
53            torch.full_like(self.concentration0, 0),
54            torch.full_like(self.concentration0, 1),
55            validate_args=validate_args,
56        )
57        transforms = [
58            PowerTransform(exponent=self.concentration0.reciprocal()),
59            AffineTransform(loc=1.0, scale=-1.0),
60            PowerTransform(exponent=self.concentration1.reciprocal()),
61        ]
62        super().__init__(base_dist, transforms, validate_args=validate_args)
63
64    def expand(self, batch_shape, _instance=None):
65        new = self._get_checked_instance(Kumaraswamy, _instance)
66        new.concentration1 = self.concentration1.expand(batch_shape)
67        new.concentration0 = self.concentration0.expand(batch_shape)
68        return super().expand(batch_shape, _instance=new)
69
70    @property
71    def mean(self):
72        return _moments(self.concentration1, self.concentration0, 1)
73
74    @property
75    def mode(self):
76        # Evaluate in log-space for numerical stability.
77        log_mode = (
78            self.concentration0.reciprocal() * (-self.concentration0).log1p()
79            - (-self.concentration0 * self.concentration1).log1p()
80        )
81        log_mode[(self.concentration0 < 1) | (self.concentration1 < 1)] = nan
82        return log_mode.exp()
83
84    @property
85    def variance(self):
86        return _moments(self.concentration1, self.concentration0, 2) - torch.pow(
87            self.mean, 2
88        )
89
90    def entropy(self):
91        t1 = 1 - self.concentration1.reciprocal()
92        t0 = 1 - self.concentration0.reciprocal()
93        H0 = torch.digamma(self.concentration0 + 1) + euler_constant
94        return (
95            t0
96            + t1 * H0
97            - torch.log(self.concentration1)
98            - torch.log(self.concentration0)
99        )
100