Why use probabilities P_I and P_T instead of logits?

ChangyaoTian / VL-LTR

VL-LTR: Learning Class-wise Visual-Linguistic Representation for Long-Tailed Visual Recognition

62 stars 8 forks source link

Why use probabilities P_I and P_T instead of logits? #2

Closed madoka109 closed 2 years ago

madoka109 commented 2 years ago

And that's why your L_rec doesn't need to be divided by 2?

ChangyaoTian commented 2 years ago

Thank you for your comment. Here we just follow the standard implementation of CrossEntropy loss in Pytorch, please refer to the code directly，which means using the prob or the logits is nothing but a matter of expression. As for not adding the additional division of 2 for L_rec, well, you can see it is just a matter of scale, isn't it?