Open Jasonbaby opened 5 years ago
I don't think so, maybe you need to tune your lr if you change the batchsize
and about convergence, it will not converge at all after several epoches
Thanks a lot. Actually, I have trained 3 epoches, with emore dataset. And the loss does not decrease at all. Is it ok? So I continue my training? I had changed some codes, but not changed the Arcface head.
I changed the backbone model, and the re-train the model with emore dataset.
that's not OK, the loss should start to decrease at very early stage
Thanks a lot. I will recheck my code.
@Jasonbaby Have u solved the problem? I am also training with small batch size 64 with r34 backbone and it is not converging.
I think the batch size need to be large, since there are a lot of batch normalization in mobileFaceNet.
Hi, I have tried to retrained the model with batch size 50, But it seems the model do not converge. I am wondering if the batch size is needed to be huge (more than 100)?