Suppose a
is a tensor of size (1, 1, 256, 256, 256)
, it seems like
-torch.log(torch.exp(-|a|) + torch.exp(-|1-a|))
as the loss has computational stability issues like nan
, are there any other implementations that can avoid this?
Would this native function be applicable to your use case and would it offer better numerical stability?
https://pytorch.org/docs/stable/generated/torch.logaddexp.html
thanks, let me try this one