Closed kuiran closed 3 years ago
In practice, I found that updating the centroids by the mean of the features is almost equal to updating the centroids by the corresponding features sequentially, achieving similar performances. So I use the latter one for brevity.
Okay, thank you!
In your paper, "For the source-domain class centroids {w}, the k-th centroid wk is updated by the mean of the encoded features belonging to class k in the mini-batch." But where is the mean operation in your code?