def prior_prob(self, prior: torch.distributions = None, log=True):
if prior is None:
prior = self.prior_w
w_log = prior.log_prob(self.w.weights).sum()
b_log = prior.log_prob(
self.b.weights).sum() if self.b is not None else 1
if not log:
w_log = w_log.exp()
b_log = b_log.exp() if self.b is not None else 1
return w_log + b_log
def forward(self, output: torch.distributions, target: torch.Tensor):
"""
calculates NegativeLogLikelihood
"""
return -output.log_prob(target).sum()