search

DanJun6737 / TransFace

[ICCV 2023] TransFace: Calibrating Transformer Training for Face Recognition from a Data-Centric Perspective
52 stars 7 forks source link

Results for WebFace42M #1

Closed mnikitin closed 1 year ago

mnikitin commented 1 year ago

Hello,

Really nice work!

But why do you not provide any experimental results using WebFace42M as the training dataset? It is more "large-scale" dataset compared to Glint360K and that's why it is more natural to train ViT-based architectures using it.

Thank you!

DanJun6737 commented 1 year ago

Hi mnikitin~

Thanks for your recognition of our work!

Due to our current limited resources, we have temporarily not conducted experiments on the WF42M dataset. We will add this experiment in the future.

Best wishes!

TranThanh96 commented 3 months ago

any update on this? I tried with wf42m but not so good

DanJun6737 commented 3 months ago

Hi @TranThanh96 ~ Recently, we are conducting experiments on the wf42m dataset. It is worth mentioning that ViT is highly sensitive to some hyperparameters such as learning rate. We will upload the pre-trained models and configuration files in the future. Please be patient and thank you for your interest in our work.