search

zma-c-137 / VarGFaceNet

310 stars 84 forks source link

请问你们有在megaface上测试过吗？ #7

Open olojuwin opened 4 years ago

olojuwin commented 4 years ago

我训练了一个模型，LFW：99.80，CFP: 0.98386, AGEDB: 0.97867，megaface 97左右，但感觉和文章有一些差距

jackchinor commented 4 years ago

请问你用knowledge distillation 训练了么？

olojuwin commented 4 years ago

请问你用knowledge distillation 训练了么？

@jackchinor 没有，

jackchinor commented 4 years ago

那可能有区别，作者说knowledge distillation 后，性能提升了很多。

olojuwin commented 4 years ago

那可能有区别，作者说knowledge distillation 后，性能提升了很多。

@jackchinor 我对比的是蒸馏之前的

clhne commented 4 years ago

@jackchinor Do you know how to do knowledge distillation? Thanks.

jackchinor commented 4 years ago

Do you know how to do knowledge distillation?

sorry,I‘ m also working on kd, still not fully understood.

yuskey commented 4 years ago

From my understanding you would run the images through both networks and use distillation loss to update the network for the student; however, I'm not sure if you train the teacher network with the student, pre train it on the same dataset, or use pre trained weights from imagenet and freeze the layers. Does anyone have any idea?

ykk648 commented 4 years ago

@olojuwin 可以放出训练参数吗