Open shamanez opened 2 years ago
Thanks for interested in this repo.
Yes, they're TF records. However, we convert them to .npy
format to prevent TF dependencies. You can see the data section on the readme and the converting script.
I also noticed that you asked which vector similarity searching library is used in the project in another issue, this is definitely worth to ask and not a spam. The current implementation in transformers
is using brute-force matrix searching, which is pure PyTorch operations and doesn't rely on other libraries. I'm considering to add a FAISS support instead of ScaNN as FAISS is more compatible with PyTorch and transformers
; on the other hand, ScaNN requires specific TF version installed, which is what I want to avoid.
Are they TF records?