i spend some time implementing LSEP loss function for keras with classweights. It seems to be working great for me, i be glad to have some feedback on some improvments.
Thanks for sharing! I actually got a working version too but it was terribly slow (I had ~300 possible labels and I modified it so that it would work with MixUp and Spectral Masking, so there was a lot going on). Have you tested it against BCE or with a high number of labels? Curious if you had the same issues as me…