Why matrix multiplication when using queue?
karray opened this issue · 0 comments
karray commented
I'm trying to get my head around the multiplication of embeddings by model.module.prototypes.weight.t()
before passing them into distributed_sinkhorn
. It is not clear why do we do this multiplication only when using a queue (Line 298). Moreover, why do we mix output prototypes with multiplied embeddings (Line 301)? What's the motivation behind this?
Lines 294 to 307 in 06b1b7c