def _kl_gumbel_gumbel(dist1, dist2):
scale_1d2 = dist1.scale / dist2.scale
return exponential.log(dist2.scale) - exponential.log(dist1.scale) \
+ EULER * (scale_1d2 - 1.) \
+ exponential.exp((dist2.loc - dist1.loc) / dist2.scale
+ lgamma.lgamma(scale_1d2 + 1.)) \
- 1 + (dist1.loc - dist2.loc) / dist2.scale
def log_prob(self, x):
if isinstance(x, chainer.Variable):
x = x.data
x = x.astype(self.lam.dtype)
xp1 = (x + 1).astype(self.lam.dtype)
x, xp1 = utils.force_array(x), utils.force_array(xp1)
return x * exponential.log(self.lam) - lgamma.lgamma(xp1) - self.lam
def _kl_gumbel_gumbel(dist1, dist2):
scale_1d2 = dist1.scale / dist2.scale
return dist2._log_scale - dist1._log_scale \
+ EULER * (scale_1d2 - 1.) \
+ exponential.exp((dist2.loc - dist1.loc) / dist2.scale
+ lgamma.lgamma(scale_1d2 + 1.)) \
- 1 + (dist1.loc - dist2.loc) / dist2.scale
def log_prob(self, x):
if isinstance(x, chainer.Variable):
x = x.data
x = x.astype(self.lam.dtype)
xp1 = (x + 1).astype(self.lam.dtype)
x, xp1 = utils.force_array(x), utils.force_array(xp1)
return x * exponential.log(self.lam) - lgamma.lgamma(xp1) - self.lam
def log_prob(self, x):
logp = - lgamma.lgamma(self.k) - self.k * exponential.log(self.theta) \
+ (self.k - 1) * exponential.log(x) - x / self.theta
xp = logp.xp
inf = xp.full_like(logp.array, xp.inf)
if isinstance(x, chainer.Variable):
x = x.array
return where.where(xp.asarray(x >= 0), logp, xp.asarray(-inf))
def log_prob(self, x):
logp = - lgamma.lgamma(self.k) - self.k * exponential.log(self.theta) \
+ (self.k - 1) * exponential.log(x) - x / self.theta
xp = logp.xp
inf = xp.full_like(logp.array, xp.inf)
if isinstance(x, chainer.Variable):
x = x.array
return where.where(xp.asarray(x >= 0), logp, xp.asarray(-inf))
def _kl_gamma_gamma(dist1, dist2):
return (dist1.k - dist2.k) * digamma.digamma(dist1.k) \
- (lgamma.lgamma(dist1.k) - lgamma.lgamma(dist2.k)) \
+ dist2.k\
* (exponential.log(dist2.theta) - exponential.log(dist1.theta)) \
+ dist1.k * (dist1.theta / dist2.theta - 1)
def entropy(self):
return (
self._half_k
+ numpy.log(2.)
+ lgamma.lgamma(self._half_k)
+ (1 - self._half_k) * digamma.digamma(self._half_k))
def _lbeta(x):
return sum_mod.sum(lgamma.lgamma(x), axis=-1) \
- lgamma.lgamma(sum_mod.sum(x, axis=-1))
def _kl_gumbel_gumbel(dist1, dist2):
scale_1d2 = dist1.scale / dist2.scale
return (dist2._log_scale - dist1._log_scale + EULER * (scale_1d2 - 1.) +
exponential.exp((dist2.loc - dist1.loc) / dist2.scale +
lgamma.lgamma(scale_1d2 + 1.)) - 1 +
(dist1.loc - dist2.loc) / dist2.scale)
def log_prob(self, x):
return (
- lgamma.lgamma(self._half_k)
- self._half_k * numpy.log(2.)
+ (self._half_k - 1) * exponential.log(x)
- 0.5 * x)
def _lbeta(a, b):
return lgamma.lgamma(a) + lgamma.lgamma(b) - lgamma.lgamma(a + b)
def _kl_gamma_gamma(dist1, dist2):
return (dist1.k - dist2.k) * digamma.digamma(dist1.k) \
- (lgamma.lgamma(dist1.k) - lgamma.lgamma(dist2.k)) \
+ dist2.k\
* (exponential.log(dist2.theta) - exponential.log(dist1.theta)) \
+ dist1.k * (dist1.theta / dist2.theta - 1)
def entropy(self):
return self.k + exponential.log(self.theta) + lgamma.lgamma(self.k) \
+ (1 - self.k) * digamma.digamma(self.k)
def _lbeta(x):
return sum_mod.sum(lgamma.lgamma(x), axis=-1) \
- lgamma.lgamma(sum_mod.sum(x, axis=-1))
def log_prob(self, x):
return - lgamma.lgamma(0.5 * self.k) - 0.5 * self.k * numpy.log(2.) \
+ (0.5 * self.k - 1) * exponential.log(x) - 0.5 * x
def entropy(self):
return self.k + exponential.log(self.theta) + lgamma.lgamma(self.k) \
+ (1 - self.k) * digamma.digamma(self.k)
def log_prob(self, x):
return - lgamma.lgamma(0.5 * self.k) - 0.5 * self.k * numpy.log(2.) \
+ (0.5 * self.k - 1) * exponential.log(x) - 0.5 * x
def entropy(self):
return 0.5 * self.k + numpy.log(2.) + lgamma.lgamma(0.5 * self.k) \
+ (1 - 0.5 * self.k) * digamma.digamma(0.5 * self.k)
def log_prob(self, x):
return - lgamma.lgamma(self._half_k) - self._half_k * numpy.log(2.) \
+ (self._half_k - 1) * exponential.log(x) - 0.5 * x
def entropy(self):
return self._half_k + numpy.log(2.) + lgamma.lgamma(self._half_k) \
+ (1 - self._half_k) * digamma.digamma(self._half_k)
def _lbeta(a, b):
return lgamma.lgamma(a) + lgamma.lgamma(b) - lgamma.lgamma(a + b)
def entropy(self):
return 0.5 * self.k + numpy.log(2.) + lgamma.lgamma(0.5 * self.k) \
+ (1 - 0.5 * self.k) * digamma.digamma(0.5 * self.k)
