hi everyone
can we obtain the BCEWithLogitsLoss as weighting
or it just BCE loss function combined with a sigmoid activation function?
Hi M!
No. They take different inputs – logits vs. probabilities – and you can’t
compensate for that difference with weighting.
Mathematically yes. In its numerical implementation, BCEWithLogitsLoss
uses
the (mathematically-equivalent) logsigmoid()
instead of log (sigmoid())
for
reasons of numerical stability.
Best.
K. Frank