Nan problem for LOG

[jganzabal]

I guess I would be safe to add epsilon to the log. Something like:
return -K.sum(alpha * K.pow(1. - pt_1, gamma) * K.log(K.epsilon()+pt_1))-K.sum((1-alpha) * K.pow( pt_0, gamma) * K.log(1. - pt_0 + K.epsilon()))

[gulzainali98]

did you find the solution? i am getting nan loss

[jganzabal]

@gulzainali98 adding K.epsilon() fixed Nan

[gulzainali98]

can this be used on simple classification? i.e i am using CelebA dataset. Image just have to be classified into 40 different classes there is no object detection.

[roywang2011]

can this be used on simple classification? i.e i am using CelebA dataset. Image just have to be classified into 40 different classes there is no object detection.

did you fixed your problem? I add K.epsilon() and still get nan

[Fellfalla]

@roywang2011 you need to make sure pt_1 does not become 1 and pt_0 does not become 0 when using gamma<1.0. Clip both with K.epsilon().

[abc1044]

I add K.epsilon(),and it works! 👍