search

clovaai / voxceleb_trainer

In defence of metric learning for speaker recognition
MIT License
1.03k stars 272 forks source link

Differences between the two implementations in paper [1] and voxsrc 20 summary paper #88

Closed Evan0-0 closed 3 years ago

Evan0-0 commented 3 years ago

Hi @joonson Thank you so much for your contribution on this open source work, it really helps me a lot. I have read your voxsrc 20 summary paper.The EER of model Q/ASP using AP loss function in paper [1] is 2.05 on Vox1 test and 1.58 on Q3 model in voxsrc 20 summary paper.Both models have the same structure and loss function. So I wonder what is the difference between these two implementations?

joonson commented 3 years ago

There are differences such as the input filterbank dimension (40 vs 64), weight decay and batch size (due to half precision training).

joonson commented 3 years ago

Also, note the difference in test lists here