pytorch

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

4
import torch
5
from torch.distributions import constraints
6
from torch.distributions.exp_family import ExponentialFamily
7
from torch.distributions.utils import broadcast_all
8
from torch.types import _size
9

10

11
__all__ = ["Exponential"]
12

13

14
class Exponential(ExponentialFamily):
15
    r"""
16
    Creates a Exponential distribution parameterized by :attr:`rate`.
17

18
    Example::
19

20
        >>> # xdoctest: +IGNORE_WANT("non-deterministic")
21
        >>> m = Exponential(torch.tensor([1.0]))
22
        >>> m.sample()  # Exponential distributed with rate=1
23
        tensor([ 0.1046])
24

25
    Args:
26
        rate (float or Tensor): rate = 1 / scale of the distribution
27
    """
28
    arg_constraints = {"rate": constraints.positive}
29
    support = constraints.nonnegative
30
    has_rsample = True
31
    _mean_carrier_measure = 0
32

33
    @property
34
    def mean(self):
35
        return self.rate.reciprocal()
36

37
    @property
38
    def mode(self):
39
        return torch.zeros_like(self.rate)
40

41
    @property
42
    def stddev(self):
43
        return self.rate.reciprocal()
44

45
    @property
46
    def variance(self):
47
        return self.rate.pow(-2)
48

49
    def __init__(self, rate, validate_args=None):
50
        (self.rate,) = broadcast_all(rate)
51
        batch_shape = torch.Size() if isinstance(rate, Number) else self.rate.size()
52
        super().__init__(batch_shape, validate_args=validate_args)
53

54
    def expand(self, batch_shape, _instance=None):
55
        new = self._get_checked_instance(Exponential, _instance)
56
        batch_shape = torch.Size(batch_shape)
57
        new.rate = self.rate.expand(batch_shape)
58
        super(Exponential, new).__init__(batch_shape, validate_args=False)
59
        new._validate_args = self._validate_args
60
        return new
61

62
    def rsample(self, sample_shape: _size = torch.Size()) -> torch.Tensor:
63
        shape = self._extended_shape(sample_shape)
64
        return self.rate.new(shape).exponential_() / self.rate
65

66
    def log_prob(self, value):
67
        if self._validate_args:
68
            self._validate_sample(value)
69
        return self.rate.log() - self.rate * value
70

71
    def cdf(self, value):
72
        if self._validate_args:
73
            self._validate_sample(value)
74
        return 1 - torch.exp(-self.rate * value)
75

76
    def icdf(self, value):
77
        return -torch.log1p(-value) / self.rate
78

79
    def entropy(self):
80
        return 1.0 - torch.log(self.rate)
81

82
    @property
83
    def _natural_params(self):
84
        return (-self.rate,)
85

86
    def _log_normalizer(self, x):
87
        return -torch.log(-x)
88