先是核心网络,跟着head:
embeddings = self.model(imgs)
thetas = self.head(embeddings, labels)
然后求损失:
loss = conf.ce_loss(thetas, labels)
head就是Arcface网络,把512维特征映射到人脸识别分类上,如果直接10万分类,需要10万尺度的话,网络太大了,所以用了Arcface转化一下,再用交叉熵求损失。
各种loss:
https://cloud.tencent.com/developer/article/1058449
mobilefacenet:zb
https://github.com/TreB1eN/InsightFace_Pytorch
download the refined dataset: (emore recommended)
- emore dataset @ BaiduDrive, emore dataset @ Dropbox
- More Dataset please refer to the original post
下面两个作者是一个人:
https://github.com/foamliu/InsightFace-PyTorch
InsightFace v3: 权重:199m
Models MegaFace LFW Download
SE-LResNet101E-IR 98.06% 99.80% Link