pytorch

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

4
import torch
5
from torch import nan
6
from torch.distributions import constraints
7
from torch.distributions.exp_family import ExponentialFamily
8
from torch.distributions.utils import (
9
    broadcast_all,
10
    lazy_property,
11
    logits_to_probs,
12
    probs_to_logits,
13
)
14
from torch.nn.functional import binary_cross_entropy_with_logits
15

16

17
__all__ = ["Bernoulli"]
18

19

20
class Bernoulli(ExponentialFamily):
21
    r"""
22
    Creates a Bernoulli distribution parameterized by :attr:`probs`
23
    or :attr:`logits` (but not both).
24

25
    Samples are binary (0 or 1). They take the value `1` with probability `p`
26
    and `0` with probability `1 - p`.
27

28
    Example::
29

30
        >>> # xdoctest: +IGNORE_WANT("non-deterministic")
31
        >>> m = Bernoulli(torch.tensor([0.3]))
32
        >>> m.sample()  # 30% chance 1; 70% chance 0
33
        tensor([ 0.])
34

35
    Args:
36
        probs (Number, Tensor): the probability of sampling `1`
37
        logits (Number, Tensor): the log-odds of sampling `1`
38
    """
39
    arg_constraints = {"probs": constraints.unit_interval, "logits": constraints.real}
40
    support = constraints.boolean
41
    has_enumerate_support = True
42
    _mean_carrier_measure = 0
43

44
    def __init__(self, probs=None, logits=None, validate_args=None):
45
        if (probs is None) == (logits is None):
46
            raise ValueError(
47
                "Either `probs` or `logits` must be specified, but not both."
48
            )
49
        if probs is not None:
50
            is_scalar = isinstance(probs, Number)
51
            (self.probs,) = broadcast_all(probs)
52
        else:
53
            is_scalar = isinstance(logits, Number)
54
            (self.logits,) = broadcast_all(logits)
55
        self._param = self.probs if probs is not None else self.logits
56
        if is_scalar:
57
            batch_shape = torch.Size()
58
        else:
59
            batch_shape = self._param.size()
60
        super().__init__(batch_shape, validate_args=validate_args)
61

62
    def expand(self, batch_shape, _instance=None):
63
        new = self._get_checked_instance(Bernoulli, _instance)
64
        batch_shape = torch.Size(batch_shape)
65
        if "probs" in self.__dict__:
66
            new.probs = self.probs.expand(batch_shape)
67
            new._param = new.probs
68
        if "logits" in self.__dict__:
69
            new.logits = self.logits.expand(batch_shape)
70
            new._param = new.logits
71
        super(Bernoulli, new).__init__(batch_shape, validate_args=False)
72
        new._validate_args = self._validate_args
73
        return new
74

75
    def _new(self, *args, **kwargs):
76
        return self._param.new(*args, **kwargs)
77

78
    @property
79
    def mean(self):
80
        return self.probs
81

82
    @property
83
    def mode(self):
84
        mode = (self.probs >= 0.5).to(self.probs)
85
        mode[self.probs == 0.5] = nan
86
        return mode
87

88
    @property
89
    def variance(self):
90
        return self.probs * (1 - self.probs)
91

92
    @lazy_property
93
    def logits(self):
94
        return probs_to_logits(self.probs, is_binary=True)
95

96
    @lazy_property
97
    def probs(self):
98
        return logits_to_probs(self.logits, is_binary=True)
99

100
    @property
101
    def param_shape(self):
102
        return self._param.size()
103

104
    def sample(self, sample_shape=torch.Size()):
105
        shape = self._extended_shape(sample_shape)
106
        with torch.no_grad():
107
            return torch.bernoulli(self.probs.expand(shape))
108

109
    def log_prob(self, value):
110
        if self._validate_args:
111
            self._validate_sample(value)
112
        logits, value = broadcast_all(self.logits, value)
113
        return -binary_cross_entropy_with_logits(logits, value, reduction="none")
114

115
    def entropy(self):
116
        return binary_cross_entropy_with_logits(
117
            self.logits, self.probs, reduction="none"
118
        )
119

120
    def enumerate_support(self, expand=True):
121
        values = torch.arange(2, dtype=self._param.dtype, device=self._param.device)
122
        values = values.view((-1,) + (1,) * len(self._batch_shape))
123
        if expand:
124
            values = values.expand((-1,) + self._batch_shape)
125
        return values
126

127
    @property
128
    def _natural_params(self):
129
        return (torch.logit(self.probs),)
130

131
    def _log_normalizer(self, x):
132
        return torch.log1p(torch.exp(x))
133