def __getitem__(self, mask: torch.BoolTensor) -> nn.Module:
self.default = mask.to(self.default)
if self.hyper is not None:
self.hyper(self.net, self.default * 2. - 1.)
return self
def forward(
self, inputs: torch.Tensor, tags: torch.Tensor,
mask: torch.BoolTensor = None, reduction: str = None
) -> torch.Tensor:
"""
Computes the negative log likelihood.
"""
if mask is None:
mask = torch.ones(*tags.size(), dtype=torch.bool)
else:
# The code below fails in weird ways if this isn't a bool tensor, so we make sure.
mask = mask.to(torch.bool)
log_denominator = self._input_likelihood(inputs, mask)
log_numerator = self._joint_likelihood(inputs, tags, mask)
nll = log_denominator - log_numerator
if reduction is None:
reduction = self.reduction
if reduction == 'sum':
nll = torch.sum(nll)
elif reduction == 'mean':
nll = torch.mean(nll)
return nll
def __getitem__(self, mask: torch.BoolTensor) -> nn.Module:
mask = mask.to(self.masks)
select = torch.all(mask == self.masks, dim=-1)
indices = torch.nonzero(select).squeeze(-1).tolist()
for i in indices:
return self.nes[i]
return None
def forward(
self, inputs: torch.Tensor, tags: torch.Tensor, mask: torch.BoolTensor = None
) -> torch.Tensor:
"""
Computes the log likelihood.
"""
if mask is None:
mask = torch.ones(*tags.size(), dtype=torch.bool)
else:
# The code below fails in weird ways if this isn't a bool tensor, so we make sure.
mask = mask.to(torch.bool)
log_denominator = self._input_likelihood(inputs, mask)
log_numerator = self._joint_likelihood(inputs, tags, mask)
return torch.sum(log_numerator - log_denominator)
