htyao89
commented
8 months ago Hi, In paper formula (3), L_kg compute the euclidean distance between w_i and w_clip_i. And in code, L_kg=1-cosine_similarity(.,.). Whether the two cases will affect the final effect? Thanks!
For the normalized features, the Euclidean distance is consistent with the cosine distance. Moreover, the smaller the Euclidean distance, the better. The higher the cosine similarity, the better. Therefore, (1-cosine) is the same as the Euclidean.
More detailed information can be see: https://stats.stackexchange.com/questions/146221/is-cosine-similarity-identical-to-l2-normalized-euclidean-distance
Hi, In paper formula (3), L_kg compute the euclidean distance between w_i and w_clip_i. And in code, L_kg=1-cosine_similarity(.,.). Whether the two cases will affect the final effect? Thanks!