ReFormer for textual semantic similarity

[srulikbd]

hey. I have 2 questions:

1. Is it good approach pooling output embeddings of the reformer model in order to use them for textual semantic similarity? (I believe a better approach is to train a classifier on top of it for STS like the approach in sentence-transformers repository..)

2. (noob question) how do I take the output embeddings of the reformer?

thanks a lot!