metrics.py implemenation doesn't seem to align with the paper

[yl3829]

Hi, when reading the implementation of the evaluations, I found couple issues:

1. The code here is using 50s as collar instead of 0.25s as stated in the paper.
2. Also, a non-causal median filter of size 11 is used, which would increase the look ahead window and hence increase the latency to 1.5s instead of 1s in the paper.

Wondering if the results in the paper is evaluated by the command provided with the default median filter size of 11 and using 50s as collar? Thanks!

[yl3829]

More carefully reading the code: the diarization results are transformed in unit of frames without subsampling, instead of seconds. Maybe worth adding some comments to explain, as it is non-trivial.

[DiLiangWU]

Thank you for pointing this out. I have added the corresponding comments in metrics.py to avoid misunderstanding.