1 # mypy: allow-untyped-defs 2 from torch.distributions import constraints 3 from torch.distributions.normal import Normal 4 from torch.distributions.transformed_distribution import TransformedDistribution 5 from torch.distributions.transforms import StickBreakingTransform 6 7 8 __all__ = ["LogisticNormal"] 9 10 11 class LogisticNormal(TransformedDistribution): 12 r""" 13 Creates a logistic-normal distribution parameterized by :attr:`loc` and :attr:`scale` 14 that define the base `Normal` distribution transformed with the 15 `StickBreakingTransform` such that:: 16 17 X ~ LogisticNormal(loc, scale) 18 Y = log(X / (1 - X.cumsum(-1)))[..., :-1] ~ Normal(loc, scale) 19 20 Args: 21 loc (float or Tensor): mean of the base distribution 22 scale (float or Tensor): standard deviation of the base distribution 23 24 Example:: 25 26 >>> # logistic-normal distributed with mean=(0, 0, 0) and stddev=(1, 1, 1) 27 >>> # of the base Normal distribution 28 >>> # xdoctest: +IGNORE_WANT("non-deterministic") 29 >>> m = LogisticNormal(torch.tensor([0.0] * 3), torch.tensor([1.0] * 3)) 30 >>> m.sample() 31 tensor([ 0.7653, 0.0341, 0.0579, 0.1427]) 32 33 """ 34 arg_constraints = {"loc": constraints.real, "scale": constraints.positive} 35 support = constraints.simplex 36 has_rsample = True 37 38 def __init__(self, loc, scale, validate_args=None): 39 base_dist = Normal(loc, scale, validate_args=validate_args) 40 if not base_dist.batch_shape: 41 base_dist = base_dist.expand([1]) 42 super().__init__( 43 base_dist, StickBreakingTransform(), validate_args=validate_args 44 ) 45 46 def expand(self, batch_shape, _instance=None): 47 new = self._get_checked_instance(LogisticNormal, _instance) 48 return super().expand(batch_shape, _instance=new) 49 50 @property 51 def loc(self): 52 return self.base_dist.base_dist.loc 53 54 @property 55 def scale(self): 56 return self.base_dist.base_dist.scale 57