search

THUFutureLab / gluon-face

An unofficial Gluon FR Toolkit for face recognition. https://gluon-face.readthedocs.io
MIT License
267 stars 51 forks source link

为什么没有做softmax的预训练? #9

Closed bigbao9494 closed 5 years ago

bigbao9494 commented 5 years ago

insightFace训练过程是分两步,先做softmax的预训练再做arcLoss训练

haoxintong commented 5 years ago

训练流程看个人经验, 怎么做都行, 最终精度差别不大

gehaocool commented 5 years ago

arcface的additive angular margin loss不需要做softmax预训练

bigbao9494 commented 5 years ago

arcface的additive angular margin loss不需要做softmax预训练

为什么？？

haoxintong commented 5 years ago

一般用softmax做预训练是因为改进的Loss在网络训练初期不稳定, 所以这是一种加快训练的策略, 对于ArcLoss来说, 属于可以做但是不必要.