pytorch

1
from numbers import Number, Real
2

3
import torch
4
from torch.distributions import constraints
5
from torch.distributions.dirichlet import Dirichlet
6
from torch.distributions.exp_family import ExponentialFamily
7
from torch.distributions.utils import broadcast_all
8

9
__all__ = ["Beta"]
10

11

12
class Beta(ExponentialFamily):
13
    r"""
14
    Beta distribution parameterized by :attr:`concentration1` and :attr:`concentration0`.
15

16
    Example::
17

18
        >>> # xdoctest: +IGNORE_WANT("non-deterministic")
19
        >>> m = Beta(torch.tensor([0.5]), torch.tensor([0.5]))
20
        >>> m.sample()  # Beta distributed with concentration concentration1 and concentration0
21
        tensor([ 0.1046])
22

23
    Args:
24
        concentration1 (float or Tensor): 1st concentration parameter of the distribution
25
            (often referred to as alpha)
26
        concentration0 (float or Tensor): 2nd concentration parameter of the distribution
27
            (often referred to as beta)
28
    """
29
    arg_constraints = {
30
        "concentration1": constraints.positive,
31
        "concentration0": constraints.positive,
32
    }
33
    support = constraints.unit_interval
34
    has_rsample = True
35

36
    def __init__(self, concentration1, concentration0, validate_args=None):
37
        if isinstance(concentration1, Real) and isinstance(concentration0, Real):
38
            concentration1_concentration0 = torch.tensor(
39
                [float(concentration1), float(concentration0)]
40
            )
41
        else:
42
            concentration1, concentration0 = broadcast_all(
43
                concentration1, concentration0
44
            )
45
            concentration1_concentration0 = torch.stack(
46
                [concentration1, concentration0], -1
47
            )
48
        self._dirichlet = Dirichlet(
49
            concentration1_concentration0, validate_args=validate_args
50
        )
51
        super().__init__(self._dirichlet._batch_shape, validate_args=validate_args)
52

53
    def expand(self, batch_shape, _instance=None):
54
        new = self._get_checked_instance(Beta, _instance)
55
        batch_shape = torch.Size(batch_shape)
56
        new._dirichlet = self._dirichlet.expand(batch_shape)
57
        super(Beta, new).__init__(batch_shape, validate_args=False)
58
        new._validate_args = self._validate_args
59
        return new
60

61
    @property
62
    def mean(self):
63
        return self.concentration1 / (self.concentration1 + self.concentration0)
64

65
    @property
66
    def mode(self):
67
        return self._dirichlet.mode[..., 0]
68

69
    @property
70
    def variance(self):
71
        total = self.concentration1 + self.concentration0
72
        return self.concentration1 * self.concentration0 / (total.pow(2) * (total + 1))
73

74
    def rsample(self, sample_shape=()):
75
        return self._dirichlet.rsample(sample_shape).select(-1, 0)
76

77
    def log_prob(self, value):
78
        if self._validate_args:
79
            self._validate_sample(value)
80
        heads_tails = torch.stack([value, 1.0 - value], -1)
81
        return self._dirichlet.log_prob(heads_tails)
82

83
    def entropy(self):
84
        return self._dirichlet.entropy()
85

86
    @property
87
    def concentration1(self):
88
        result = self._dirichlet.concentration[..., 0]
89
        if isinstance(result, Number):
90
            return torch.tensor([result])
91
        else:
92
            return result
93

94
    @property
95
    def concentration0(self):
96
        result = self._dirichlet.concentration[..., 1]
97
        if isinstance(result, Number):
98
            return torch.tensor([result])
99
        else:
100
            return result
101

102
    @property
103
    def _natural_params(self):
104
        return (self.concentration1, self.concentration0)
105

106
    def _log_normalizer(self, x, y):
107
        return torch.lgamma(x) + torch.lgamma(y) - torch.lgamma(x + y)
108