Source code for zhusuan.distributions.logistic

import torch
import numpy as np
from zhusuan.distributions import Distribution
from zhusuan.distributions.utils import (
    assert_same_log_float_dtype,
    check_broadcast
)


class Logistic(Distribution):
    """
    The class of univariate Logistic distribution, always using the reparametrization trick from (Kingma, 2013).
    See :class:`~zhusuan.distributions.base.Distribution` for details.

    :param loc: A 'float' Var. The location term acting on standard Logistic distribution.
    :param scale: A 'float' Var. The scale term acting on standard Logistic distribution.
    """

    def __init__(self,
                 loc,
                 scale,
                 dtype=None,
                 is_continuous=True,
                 group_ndims=0,
                 device=torch.device('cpu'),
                 **kwargs):

        self._loc = torch.as_tensor(loc, dtype=dtype).to(device)
        self._scale = torch.as_tensor(scale, dtype=dtype).to(device)
        if torch.less_equal(self.scale, 0.).any():
            raise ValueError("scale less than zero")
        check_broadcast(self._loc, self.scale)
        dtype = assert_same_log_float_dtype([(self._loc, "Logistic.loc"), (self._scale, "Logistic.scale")])
        super(Logistic, self).__init__(dtype,
                                       is_continuous,
                                       is_reparameterized=True,
                                       group_ndims=group_ndims,
                                       device=device,
                                       **kwargs)

    @property
    def loc(self):
        return self._loc

    @property
    def scale(self):
        return self._scale

    def _batch_shape(self):
        return torch.broadcast_shapes(self._loc.shape, self._scale.shape)

    def _sample(self, n_samples=1, **kwargs):
        if n_samples > 1:
            _shape = self._loc.shape
            _shape = torch.Size([n_samples]) + _shape
            _len = len(self._loc.shape)
            _loc = self._loc.repeat([n_samples, *_len * [1]])
            _scale = self._scale.repeat([n_samples, *_len * [1]])
        else:
            _shape = self._loc.shape
            _loc = self._loc
            _scale = self._scale

        uniform = torch.nn.init.uniform_(torch.empty(_shape, dtype=self._dtype), 0., 1.)  # !check efficiency
        epsilon = (torch.log(uniform) - torch.log(1 - uniform)).to(self._device)
        _sample = _loc + _scale * epsilon
        self.sample_cache = _sample
        return _sample

    def _log_prob(self, sample=None, **kwargs):
        if sample is None:
            sample = self.sample_cache
        if len(sample.shape) > len(self._loc.shape):
            n_samples = sample.shape[0]
            _len = len(self._loc.shape)
            _loc = self._loc.repeat([n_samples, *_len * [1]])
            _scale = self._scale.repeat([n_samples, *_len * [1]])
        else:
            _loc = self._loc
            _scale = self._scale

        z = (sample - _loc) / _scale
        return -z - 2. * torch.nn.Softplus()(-z) - torch.log(_scale)

    def _prob(self, given):
        return torch.exp(self._log_prob(given))