DINO SELF-SUPERVISED LEARNING: dinossl

[JaejinCho]

This is a work in progress [WIP]. Below are the items in order:

• Fix parts to avoid using labels at all:

  • Sampler needs to change to avoid using labels at all.
  • Data preparation step (in shell level): NOT to concatenate the same speaker utterances. --> However, to utilize local and global view embedding contrast in training, using labels to concatenate might be better to have long utterances (need discussion). In that case, it is supervised embedding learning.

• Hyper-parameter tuning

• Code trimming:

  • I think this is NOT necessary to be done in this PR unless it hinders merging.
  • BTW, I have a private list related to this, which I may post later if needed.

• *_dinossl.py were copied and edited from .py files (also *dinossl.sh from .sh)

[jesus-villalba]

It could be good to create a new recipe directory for this separate from v1.1, for example v2, since you need to change the experimental setup

[jesus-villalba]

About concatenating or not utterances, we are concatenating by video id, not speaker id, the question is if we allow to know the video id in our setup. I think we should follow protocol from VoxSrc challenge last year, could you find out if they allow it? They are repeating challenge this year, we should participate