def entropy(self):
# option 1:
output = (-self.scale * ive(self.__m / 2, self.scale) / ive(
(self.__m / 2) - 1, self.scale))
# option 2:
# output = - self.scale * ive_fraction_approx(torch.tensor(self.__m / 2), self.scale)
# option 3:
# output = - self.scale * ive_fraction_approx2(torch.tensor(self.__m / 2), self.scale)
return output.view(*(output.shape[:-1])) + self._log_normalization()
def _log_normalization(self):
output = -(
(self.__mf / 2 - 1) * math_ops.log(self.scale) -
(self.__mf / 2) * math.log(2 * math.pi) -
(self.scale + math_ops.log(ive(self.__mf / 2 - 1, self.scale))))
return array_ops.reshape(
output, ops.convert_to_tensor(array_ops.shape(output)[:-1]))
def mean(self):
return self.loc * (ive(self.__m / 2, self.scale) /
ive(self.__m / 2 - 1, self.scale))
def _log_normalization(self):
output = -((self.__m / 2 - 1) * torch.log(self.scale) -
(self.__m / 2) * math.log(2 * math.pi) -
(self.scale + torch.log(ive(self.__m / 2 - 1, self.scale))))
return output.view(*(output.shape[:-1]))
def entropy(self):
output = -self.scale * ive(self.__m / 2, self.scale) / ive(
(self.__m / 2) - 1, self.scale)
return output.view(*(output.shape[:-1])) + self._log_normalization()
def _mean(self):
return self._loc * (ive(self.__mf / 2, self.scale) /
ive(self.__mf / 2 - 1, self.scale))
def _entropy(self):
return -array_ops.reshape(
self.scale * ive(self.__mf / 2, self.scale) / ive(
(self.__mf / 2) - 1, self.scale),
ops.convert_to_tensor(array_ops.shape(
self.scale)[:-1])) + self._log_normalization()
def _kl_vmf_vmf(vmf1, vmf2):
rel_entropy = ((vmf2.dim * 0.5) - 1) * torch.log(
vmf2.scale) - (vmf2.dim * 0.5) * math.log(2 * math.pi) - torch.log(
ive(vmf2.dim / 2 - 1, vmf2.scale))
return vmf1.entropy() - rel_entropy
def _log_normalization(self):
output = -(self.__m / 2 - 1) * torch.log(self.scale) \
- (self.__m / 2) * math.log(2 * math.pi)\
- (self.scale + torch.log(ive(self.__m / 2 - 1, self.scale)))
return output.view(*([output.flatten().shape[0]]))
