Er_Hall
(Er Hall)
1
Hi all,
I have a question I have a multiclass classification problem where each sample belongs only to one class, my targets are in [0, C-1] format.
My question is: as input the cross entropy loss must take logits? so as a last layer I need to have a sigmoid layer?
albanD
(Alban D)
2
You can see the doc for what exactly each one is doing: cross entropy, nll, bce and bce with logits should be the ones you want to check.
ricardborras
(Ricard BorrĂ s Navarra)
3
With cross entropy loss, it is enought to input class labels directly. Also, softmax is not needed for cross entropy loss.