Possible bug report and suggestion on how to fix it.

[hash2430]

I think you need to change utils.get_cossim() line 42
as follows
from
if speaker_num == centroid_num:
to
if (speaker_num == centroid_num) and (hp.training):

because you do not want to recalculate centroid from evaluation embeddings at the test time.
Wouldn't it be appropriate to use centroids that are calculated solely from enrollment embeddings?

Maybe my understanding on the paper and code is insufficient and I might be missing something. Let me know if I am wrong about this.

Of course it works without fixing it but in my understanding in strict sense, that centroid recalculation is only required for stabilizing training steps.