search

clovaai / voxceleb_trainer

In defence of metric learning for speaker recognition
MIT License
1.05k stars 272 forks source link

about AMsoftmax #21

Closed DYJ1111 closed 4 years ago

DYJ1111 commented 4 years ago

Hi, I use same coda, just different loss function (softmax vs AMsoftmax). But the loss is reducing slowly and acc is increase slowly when i use amsoftmax . Have you ever meet this problem?

DYJ1111 commented 4 years ago

I use ResnetSe34 as the model

joonson commented 4 years ago

Yes, we have observed that it can take longer to converge and gets lower accuracy, but the verification EER is better.

DYJ1111 commented 4 years ago

Yes, we have observed that it can take longer to converge and gets lower accuracy, but the verification EER is better.

So could you please tell me the number of epoches? i set epoch is 40, but the loss is too high (7.x) and acc is too low(3.x%)

009deep commented 4 years ago

@DYJ1111 Check out issue10

@joonson one question I have for those logs is model mentioned there is

model ResNetSEFB34X

I am guessing it refers ResNetSE34L in this repo. If not please correct me what iis difference in that model.

DYJ1111 commented 4 years ago

@DYJ1111 Check out issue10

@joonson one question I have for those logs is model mentioned there is

model ResNetSEFB34X

I am guessing it refers ResNetSE34L in this repo. If not please correct me what iis difference in that model.

Thanks, I use ResNetSE34L and train 100 epoches, the TEER is 50% and veer is 4.8

009deep commented 4 years ago

@DYJ1111 Results Joon Soon has presented is at 500 epochs for different hyper parameters.

joonson commented 4 years ago

@009deep ResNetSEFB34X is the same as ResNetSE34L in the released version