How to deploy in real recommender systems

[weilonghu]

I have several questions:

1. As I known, faiss does not support 'max' operation.
2. Fot i-th layer user representaion, we will compute each head pairwise to get the similarity score, So we need to retrieve H^2 times？If there are L layers, eventually we need to retrieve L* H^2 times?

Thanks!

[archersama]

thank your question.
1.we just save and get item embedding in faiss, 'max' operation is conducted in inference code.
2.In our code, for all heads, the Fe score is done in a parallel way. For a single layer, it only needs once to compute similarity score. For L layers, we only compute L times.

And in online inference, we find that only using last layer to compute MaxSim Score is also useful, so you can use last layer to
compute MaxSim score to get the similarity score.

[weilonghu]

Thanks for your reply, I'll try it later

[archersama]

It's ok, feel free to ask questions