pytorch

1
# mypy: allow-untyped-defs
2
import math
3

4
import torch
5
from torch import inf, nan
6
from torch.distributions import Chi2, constraints
7
from torch.distributions.distribution import Distribution
8
from torch.distributions.utils import _standard_normal, broadcast_all
9
from torch.types import _size
10

11

12
__all__ = ["StudentT"]
13

14

15
class StudentT(Distribution):
16
    r"""
17
    Creates a Student's t-distribution parameterized by degree of
18
    freedom :attr:`df`, mean :attr:`loc` and scale :attr:`scale`.
19

20
    Example::
21

22
        >>> # xdoctest: +IGNORE_WANT("non-deterministic")
23
        >>> m = StudentT(torch.tensor([2.0]))
24
        >>> m.sample()  # Student's t-distributed with degrees of freedom=2
25
        tensor([ 0.1046])
26

27
    Args:
28
        df (float or Tensor): degrees of freedom
29
        loc (float or Tensor): mean of the distribution
30
        scale (float or Tensor): scale of the distribution
31
    """
32
    arg_constraints = {
33
        "df": constraints.positive,
34
        "loc": constraints.real,
35
        "scale": constraints.positive,
36
    }
37
    support = constraints.real
38
    has_rsample = True
39

40
    @property
41
    def mean(self):
42
        m = self.loc.clone(memory_format=torch.contiguous_format)
43
        m[self.df <= 1] = nan
44
        return m
45

46
    @property
47
    def mode(self):
48
        return self.loc
49

50
    @property
51
    def variance(self):
52
        m = self.df.clone(memory_format=torch.contiguous_format)
53
        m[self.df > 2] = (
54
            self.scale[self.df > 2].pow(2)
55
            * self.df[self.df > 2]
56
            / (self.df[self.df > 2] - 2)
57
        )
58
        m[(self.df <= 2) & (self.df > 1)] = inf
59
        m[self.df <= 1] = nan
60
        return m
61

62
    def __init__(self, df, loc=0.0, scale=1.0, validate_args=None):
63
        self.df, self.loc, self.scale = broadcast_all(df, loc, scale)
64
        self._chi2 = Chi2(self.df)
65
        batch_shape = self.df.size()
66
        super().__init__(batch_shape, validate_args=validate_args)
67

68
    def expand(self, batch_shape, _instance=None):
69
        new = self._get_checked_instance(StudentT, _instance)
70
        batch_shape = torch.Size(batch_shape)
71
        new.df = self.df.expand(batch_shape)
72
        new.loc = self.loc.expand(batch_shape)
73
        new.scale = self.scale.expand(batch_shape)
74
        new._chi2 = self._chi2.expand(batch_shape)
75
        super(StudentT, new).__init__(batch_shape, validate_args=False)
76
        new._validate_args = self._validate_args
77
        return new
78

79
    def rsample(self, sample_shape: _size = torch.Size()) -> torch.Tensor:
80
        # NOTE: This does not agree with scipy implementation as much as other distributions.
81
        # (see https://github.com/fritzo/notebooks/blob/master/debug-student-t.ipynb). Using DoubleTensor
82
        # parameters seems to help.
83

84
        #   X ~ Normal(0, 1)
85
        #   Z ~ Chi2(df)
86
        #   Y = X / sqrt(Z / df) ~ StudentT(df)
87
        shape = self._extended_shape(sample_shape)
88
        X = _standard_normal(shape, dtype=self.df.dtype, device=self.df.device)
89
        Z = self._chi2.rsample(sample_shape)
90
        Y = X * torch.rsqrt(Z / self.df)
91
        return self.loc + self.scale * Y
92

93
    def log_prob(self, value):
94
        if self._validate_args:
95
            self._validate_sample(value)
96
        y = (value - self.loc) / self.scale
97
        Z = (
98
            self.scale.log()
99
            + 0.5 * self.df.log()
100
            + 0.5 * math.log(math.pi)
101
            + torch.lgamma(0.5 * self.df)
102
            - torch.lgamma(0.5 * (self.df + 1.0))
103
        )
104
        return -0.5 * (self.df + 1.0) * torch.log1p(y**2.0 / self.df) - Z
105

106
    def entropy(self):
107
        lbeta = (
108
            torch.lgamma(0.5 * self.df)
109
            + math.lgamma(0.5)
110
            - torch.lgamma(0.5 * (self.df + 1))
111
        )
112
        return (
113
            self.scale.log()
114
            + 0.5
115
            * (self.df + 1)
116
            * (torch.digamma(0.5 * (self.df + 1)) - torch.digamma(0.5 * self.df))
117
            + 0.5 * self.df.log()
118
            + lbeta
119
        )
120