Hi everyone,
I wanted to fine tune a sentence transformers with a triplet (anchor, positive, negative) and so I choose the tripletLoss function (Losses — Sentence-Transformers documentation )
The default loss function is euclidean norm and I asked myself if there is an obvious benefits to get l2 (euclidean norm) instead of cosine similarity.
Because, I think that a huge majority of sentence transformers application goes with the cosine instead of the L2-norm and the standard seems being the cosine function.
So my point is why and when should I use L2 to compare vectors and is there a logical reason having L2 loss function instead of cosine?
It is just a random choice or is there a deeper reason?
Thanks