Warning about graph is not fully connected

With the provided ecapa_tdnn.yaml in AMI Diarization recipe, I was able to get 2.13% DER on Dev and 2.46% on Eval. However, I got the warning at the 10th step during evaluation on Dev:

speechbrain/speechbrain/processing/diarization.py:588: UserWarning: Graph is not fully connected, spectral embedding may not work as expected.

Can I ask if I need to worry about this? Thanks!

@nauman might be able to help here.

Hi,
This warning is similar as the sklearn warning and is shown when the affinity matrix has zero rows/columns, i.e., num of connected composnents exceeding 1. Kindly check once if that is the case for you. However, I think this is based on implementation details/affinity and can be ignored for this particular dataset as the affinity matrix may sometimes have very small values for some embeddings.