About loss function

[Mhaiyang]

Hi, I am confused by the "temperature" parameter in the "cross_entropy2d" and "KD_KLDivLoss". What is the function of this parameter? Thanks!

[TuesdayT]

Hi！The temperature parameter is a hyper-parameter in knowledge distillation. It can be used in the classification task (saliency detection can be seen as a pixel-wise binary classification) to increase the difficulty of classification and improve the ability of mining difficult samples. In detail, please refer to https://nervanasystems.github.io/distiller/knowledge_distillation.html On 07/23/2020 07:36, TaylorMei wrote: Hi, I am confused by the "temperature" parameter in the "cross_entropy2d" and "KD_KLDivLoss". What is the function of this parameter? Thanks! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.