search

instadeepai / protein-structure-tokenizer

Official implementation of "Learning the language of protein structures"
Apache License 2.0
6 stars 0 forks source link

FSQ training #1

Open andrewliu2001 opened 1 month ago

andrewliu2001 commented 1 month ago

Hello, I'm trying to reproduce FSQ VAE on a different point cloud dataset. Both the 64000 and 4096 vocab sizes don't produce good reconstructions. Is the structure module decoding also important (currently I'm just using 6 transformers)?

JeremieDona commented 3 weeks ago

Hello,

Thanks for your interest in our work. The decoding module is indeed very important as it provides the residue frames and atoms coordinates.

Are you trying to retrain from scratch or finetune from some checkpoints ?