Where is the DINO self-supervised framework?
speaker-lover opened this issue · comments
Dear authors,
Thanks for this great work. In the paper, you proposed the RawNet3 architecture under the DINO self-supervised learning framework.
But in the code, i only see the RawNet3 architecture. I don‘t see the code for DINO self-supervised framework including loss function and training strategy.
So, is it that you haven't open sourced the DINO self-supervised framework or I missed it?
Looking forward to receiving your reply. Thanks.
Hi @yfchen97.
No we haven't open sourced the SSL part yet.
We had plans, however, we postponed it in order to match with another paper of ours entitledSelf-supervised curriculum learning for speaker verification
.
Our implementation is also based on the official DINO repository: https://github.com/facebookresearch/dino.
If you can adapt its dataloader and other necessary parts towards audio, I expect you to achieve similar results using RawNet3 architecture.
Thank you for reply. I think angleproto loss can be used as a self-supervised learning method, what do you think?