flmgjk comments

Repositories
Issues
Comments

Results 4 comments of


                                            flmgjk

loss_diff decrease to zero very fast

I also encountered the same problem, and found that private code learn nothing, they are all zeros....， i didn't directly use the original experimental datasets, but i think it doesn't...

[F-VLM] How to reproduce the experimental results of the paper?

I can not reproduce either, the unseen bbox AP50 still has a large margin with the paper result.

[F-VLM] precomputed embeddings for lvis 1.0?

i found the emb[:83] is different from each other, but should be 1 background and 80 classes?

Training stuck after index is created when training from checkpoint

I met the same