search

StevenBanama / C3AE

C3AE implement
BSD 2-Clause "Simplified" License
86 stars 16 forks source link

About facal loss? #6

Open rongyixiao opened 4 years ago

rongyixiao commented 4 years ago

Hello, does facal loss address data imbalance? if data is balance, is facal loss useless? In this case, loss in the original paper should be adopted?

rongyixiao commented 4 years ago

1 、Facal loss is supposed to increase learning of difficult samples, right? 2、Why is my facal loss almost the same? Thank you

StevenBanama commented 4 years ago

Hello, does facal loss address data imbalance? if data is balance, is facal loss useless? In this case, loss in the original paper should be adopted? @rongyixiao Q1:due to the unbalance distribution of wiki and imdb. Q2: you can use CE, when the dataset distribute balancely. Q3: origin papper may randomly select sample so that decrease the risk of overfitting and unbalance

StevenBanama commented 4 years ago

focal loss