def _log_pdf(self, x):
return (
self.shape * log(self.rate)
+ (self.shape - 1) * log(x)
- self.rate * x
- log(gamma(self.shape))
)
def _log_pdf(self, x):
return (
log(gamma(self.alpha.sum(axis=self.axis)))
+ sum(
(self.alpha - 1) * log(x) - log(gamma(self.alpha)),
axis=self.axis
)
)
def __init__(self,
mean=None,
logit=None,
use_gumbel_softmax=True,
tau=0.1):
super().__init__()
if mean is not None:
self.mean = asarray(mean)
assert ((self.mean.value >= 0).all()
and np.allclose(self.mean.value.sum(axis=-1), 1))
self.logit = log(self.mean)
self._log_pdf = self._log_pdf_mean
elif logit is not None:
self.mean = softmax(logit)
self.logit = asarray(logit)
self._log_pdf = self._log_pdf_logit
else:
raise ValueError
self.n_category = self.mean.shape[-1]
if use_gumbel_softmax:
self.forward = self._forward_gumbel_softmax
else:
self.forward = self._forward
self.tau = tau
self.eye = np.eye(self.n_category)
def _log_pdf(self, x):
assert x.shape[-1] == self.mu.size
if x.ndim == 1:
squeeze = True
x = broadcast_to(x, (1, self.mu.size))
else:
squeeze = False
assert x.ndim == 2
d = x - self.mu
d = d.transpose()
logp = (-0.5 * (solve(self.cov, d) * d).sum(axis=0) -
(self.mu.size * 0.5) * log(2 * np.pi) - 0.5 * logdet(self.cov))
if squeeze:
logp = logp.sum()
return logp
def _log_pdf(self, x):
return (-np.log(np.pi) - log(self.scale) -
log(1 + square((x - self.loc) / self.scale)))
def _log_pdf(self, x):
return (self.shape * log(self.rate) + (self.shape - 1) * log(x) -
self.rate * x - log(gamma(self.shape)))
def _log_pdf_mu(self, x):
return x * log(self.mean) + (1 - x) * log(1 - self.mean)
def _log_pdf(self, x):
return ((self.ndim - 1) * log(x) - 0.5 * square(x / self.std) -
self.ndim * log(self.std) - np.log(sp.gamma(0.5 * self.ndim)))
def _log_pdf(self, x):
try:
return -SigmoidCrossEntropy().forward(self.logit, x)
except AttributeError:
return x * log(self.mu) + (1 - x) * log(1 - self.mu)
def _log_pdf(self, x):
try:
return -SoftmaxCrossEntropy(axis=self.axis).forward(self.logit, x)
except (KeyError, AttributeError):
return (x * log(self.mu)).sum(axis=self.axis)
def _log_pdf_mean(self, x):
return x * log(self.mean)
def _log_pdf(self, x):
try:
return -SigmoidCrossEntropy().forward(self.logit, x)
except AttributeError:
return x * log(self.mu) + (1 - x) * log(1 - self.mu)
def _log_pdf(self, x):
return log(self.pdf(x))
def _log_pdf(self, x):
return (
-np.log(np.pi)
- log(self.scale)
- log(1 + square((x - self.loc) / self.scale))
)
def _log_pdf(self, x):
return -self.rate * x + log(self.rate)
def _log_pdf(self, x):
return log(self.pdf(x))
def _log_pdf(self, x):
return np.log(0.5) - abs(x - self.loc) / self.scale - log(self.scale)
def _log_pdf(self, x):
try:
return -SoftmaxCrossEntropy(axis=self.axis).forward(self.logit, x)
except (KeyError, AttributeError):
return (x * log(self.mu)).sum(axis=self.axis)
def kl_gaussian(q, p):
qvar = square(q.std)
pvar = square(p.std)
return log(p.std) - log(q.std) + 0.5 * (qvar + square(p.mean - q.mean)) / pvar - 0.5
def _log_pdf(self, x):
return (log(gamma(self.alpha.sum(axis=self.axis))) + sum(
(self.alpha - 1) * log(x) - log(gamma(self.alpha)),
axis=self.axis))
