def rsample(self, sample_shape=torch.Size(())):
shape = self._extended_shape(sample_shape)
# X1 ~ Gamma(df1 / 2, 1 / df1), X2 ~ Gamma(df2 / 2, 1 / df2)
# Y = df2 * df1 * X1 / (df1 * df2 * X2) = X1 / X2 ~ F(df1, df2)
X1 = self._gamma1.rsample(sample_shape).view(shape)
X2 = self._gamma2.rsample(sample_shape).view(shape)
X2.clamp_(min=_finfo(X2).tiny)
Y = X1 / X2
Y.clamp_(min=_finfo(X2).tiny)
return Y
def rsample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
if torch._C._get_tracing_state():
# [JIT WORKAROUND] lack of support for .uniform_()
u = torch.rand(shape, dtype=self.loc.dtype, device=self.loc.device) * 2 - 1
return self.loc - self.scale * u.sign() * torch.log1p(-u.abs().clamp(min=_finfo(self.loc).tiny))
u = self.loc.new(shape).uniform_(_finfo(self.loc).eps - 1, 1)
# TODO: If we ever implement tensor.nextafter, below is what we want ideally.
# u = self.loc.new(shape).uniform_(self.loc.nextafter(-.5, 0), .5)
return self.loc - self.scale * u.sign() * torch.log1p(-u.abs())
def rsample(self, sample_shape=torch.Size(())):
shape = self._extended_shape(sample_shape)
# X1 ~ Gamma(df1 / 2, 1 / df1), X2 ~ Gamma(df2 / 2, 1 / df2)
# Y = df2 * df1 * X1 / (df1 * df2 * X2) = X1 / X2 ~ F(df1, df2)
X1 = self._gamma1.rsample(sample_shape).view(shape)
X2 = self._gamma2.rsample(sample_shape).view(shape)
X2.clamp_(min=_finfo(X2).tiny)
Y = X1 / X2
Y.clamp_(min=_finfo(X2).tiny)
return Y
def sample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
with torch.no_grad():
if torch._C._get_tracing_state():
# [JIT WORKAROUND] lack of support for .uniform_()
u = torch.rand(shape,
dtype=self.probs.dtype,
device=self.probs.device)
u = u.clamp(min=_finfo(self.probs).tiny)
else:
u = self.probs.new(shape).uniform_(_finfo(self.probs).tiny, 1)
return (u.log() / (-self.probs).log1p()).floor()
def rsample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
value = _standard_gamma(
self.concentration.expand(shape)) / self.rate.expand(shape)
data = value.data if isinstance(value, Variable) else value
data.clamp_(min=_finfo(value).tiny) # do not record in autograd graph
return value
def rsample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
value = _standard_gamma(
self.concentration.expand(shape)) / self.rate.expand(shape)
value.detach().clamp_(
min=_finfo(value).tiny) # do not record in autograd graph
return value
def __init__(self, loc, scale, validate_args=None):
self.loc, self.scale = broadcast_all(loc, scale)
finfo = _finfo(self.loc)
if isinstance(loc, Number) and isinstance(scale, Number):
base_dist = Uniform(finfo.tiny, 1 - finfo.eps)
else:
base_dist = Uniform(self.loc.new(self.loc.size()).fill_(finfo.tiny), 1 - finfo.eps)
transforms = [ExpTransform().inv, AffineTransform(loc=0, scale=-torch.ones_like(self.scale)),
ExpTransform().inv, AffineTransform(loc=loc, scale=-self.scale)]
super(Gumbel, self).__init__(base_dist, transforms, validate_args=validate_args)
def __init__(self, loc, scale, validate_args=None):
self.loc, self.scale = broadcast_all(loc, scale)
finfo = _finfo(self.loc)
if isinstance(loc, Number) and isinstance(scale, Number):
batch_shape = torch.Size()
base_dist = Uniform(finfo.tiny, 1 - finfo.eps)
else:
batch_shape = self.scale.size()
base_dist = Uniform(self.loc.new(self.loc.size()).fill_(finfo.tiny), 1 - finfo.eps)
transforms = [ExpTransform().inv, AffineTransform(loc=0, scale=-torch.ones_like(self.scale)),
ExpTransform().inv, AffineTransform(loc=loc, scale=-self.scale)]
super(Gumbel, self).__init__(base_dist, transforms, validate_args=validate_args)
def rsample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
uni_dist = self.scale.new(shape).uniform_(_finfo(self.scale).eps, 1)
# X ~ Uniform(0, 1)
# Y = loc - scale * ln (-ln (X)) ~ Gumbel(loc, scale)
return self.loc - self.scale * torch.log(-uni_dist.log())
def sample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
with torch.no_grad():
u = self.probs.new(shape).uniform_(_finfo(self.probs).tiny, 1)
return (u.log() / (-self.probs).log1p()).floor()
def rsample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
value = _standard_gamma(self.alpha.expand(shape)) / self.beta.expand(shape)
data = value.data if isinstance(value, Variable) else value
data.clamp_(min=_finfo(value).tiny) # do not record in autograd graph
return value
def rsample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
u = self.loc.new(shape).uniform_(_finfo(self.loc).eps - 1, 1)
# TODO: If we ever implement tensor.nextafter, below is what we want ideally.
# u = self.loc.new(shape).uniform_(self.loc.nextafter(-.5, 0), .5)
return self.loc - self.scale * u.sign() * torch.log1p(-u.abs())
def rsample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
u = self.loc.new(shape).uniform_(_finfo(self.loc).eps - 1, 1)
# TODO: If we ever implement tensor.nextafter, below is what we want ideally.
# u = self.loc.new(shape).uniform_(self.loc.nextafter(-.5, 0), .5)
return self.loc - self.scale * u.sign() * torch.log1p(-u.abs())
def _dirichlet_sample_nograd(alpha):
probs = torch._C._standard_gamma(alpha)
probs /= probs.sum(-1, True)
eps = _finfo(probs).eps
return probs.clamp_(min=eps, max=1 - eps)
def sample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
with torch.no_grad():
u = self.probs.new(shape).uniform_(_finfo(self.probs).tiny, 1)
return (u.log() / (-self.probs).log1p()).floor()
def rsample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
value = _standard_gamma(self.concentration.expand(shape)) / self.rate.expand(shape)
value.data.clamp_(min=_finfo(value).tiny) # do not record in autograd graph
return value
def rsample(self, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
uni_dist = self.scale.new(shape).uniform_(_finfo(self.scale).eps, 1)
# X ~ Uniform(0, 1)
# Y = loc - scale * ln (-ln (X)) ~ Gumbel(loc, scale)
return self.loc - self.scale * torch.log(-uni_dist.log())
