Closed sumitpai closed 4 years ago
Multiclass_loss = log (exp(score_pos)/sum(exp(score_pos) + exp(score_neg)))
when score_pos goes above certain value, the loss becomes inf/inf which results in NaN loss. Currently we can use high regularization to avoid this. However it sometimes does occur
Description
Multiclass_loss = log (exp(score_pos)/sum(exp(score_pos) + exp(score_neg)))
when score_pos goes above certain value, the loss becomes inf/inf which results in NaN loss. Currently we can use high regularization to avoid this. However it sometimes does occur