How to get sim_pos and sim_neg?

[CharlesChen24]

How to calculate sim_pos and sim_neg based on the feature embeddings output by the model and their labels?
What are the shapes of sim_pos and sim_neg?
Do they need L2 normalization?

[shuaitang]

the paper https://arxiv.org/pdf/1611.00822.pdf used cosine similarity to calculate the pair-wise similarity score. Bounded similarity scores tend to work better imo.

[CharlesChen24]

the paper https://arxiv.org/pdf/1611.00822.pdf used cosine similarity to calculate the pair-wise similarity score. Bounded similarity scores tend to work better imo.

so sim_pos represents the similarity score of positive pairs? The similarity score of positive pairs is the dot product of the feature embeddings of two samples with the same label？sim_neg (the similarity score of negative pairs) can be get in the same way?