search

mk-minchul / dcface

137 stars 17 forks source link

question about training synthesis data #20

Closed szlbiubiubiu closed 1 year ago

szlbiubiubiu commented 1 year ago

Hi, thank you for your excellent work.

I have a problem when trying to train the synthesis data you released. My training would converge in one epoch and always get 50% verification accuracy on validation sets, regardless of which loss function I used. But when I switch back to casia-webface, the results would be as normal(i.e. 90+ accuracy).

Could you please let me know if there is any difference between training synthesis data and real data?

The configuration is as follows: Start LR:0.1 weight decay: 5e-4 optimizer: SGD network: IResnet50 Loss: Cosface(s=64, m=0.4); Arcface(s=64,m=0.5), AdaFace(from https://github.com/mk-minchul/AdaFace)

szlbiubiubiu commented 1 year ago

Training works after I extract the images to image format(jpg/png). Sorry for the issue.