Why can't my training loss decrease with ArcFace? It drops to about 21 and stops falling.

Open LHH20000923 opened this issue 4 years ago • 3 comments

Jan 08 '22 07:01 LHH20000923

which database do you use for training, loss function, batch size etc...?

Jan 14 '22 13:01 fdbtrs

I use CASIA-WebFace, optim is SGD ,lr=0.1, batchsize=256/512, and didn't use knowledge distillation。 The loss curve is shown below。Is it a problem with the data set？

Jan 22 '22 07:01 LHH20000923

For casia dataset, you may need to train the model for at least 50 epochs with lr step of 20 30 and 40. The reported FR result in the paper is based on MS1Mv2

Jan 26 '22 20:01 fdbtrs