facebookresearch/swav

Why setting two kinds of codes (Q, P)?

zhujiagang opened this issue · 3 comments

zhujiagang commented

A great work!
There seems to be two kinds of codes in this paper: pt in eq2, and qs in eq2 (Q in eq3). Can eq2 be ps * log(pt) instead of qs * log(pt)?

mathildecaron31 commented

Hi @zhujiagang

For the targets (i.e. Q in H(Q, P) = - Q log P) we need to impose an equipartition constraint to avoid all images to have the same representation. We enforce the constraint thanks to the Sinkhorn-Knopp algorithm.

Hope that helps

yuvaramsingh94 commented

Hi @mathildecaron31
following in the same question , i have trained the model on an unlabeled dataset , i am looking to cluster these images using this trained model . should i use the prototype values we get from the model while validating or should i run sinkhorn-knopp algorithm to get the clusters . pls suggest .

thanks

  • 👀3
shijianjian commented

@mathildecaron31 Hi, I have the same doubts for how to get the proper cluster assignments during the prediction?