search

ThisisBillhe / BiViT

The official implementation of BiViT: Extremely Compressed Binary Vision Transformers
Apache License 2.0
9 stars 0 forks source link

Problems on reproducing the experimental results. #1

Open Alexrich961210 opened 6 months ago

Alexrich961210 commented 6 months ago

Could you please update some key codes or configs for helping reproducing the results of TinyImageNet in this paper? Thanks for your help.

ThisisBillhe commented 6 months ago

any specific config or problem you have met?

Alexrich961210 commented 6 months ago

Thanks for your reply. I have these specific problems:

  1. Is the data augmentation totally same as the default settings in DeiT repository?
  2. Which teacher model is used for training the BiViT on TinyImageNet and ImageNet?
ThisisBillhe commented 6 months ago
  1. yes.
  2. We use the corresponding FP model as the teacher model for BiViT.
Alexrich961210 commented 6 months ago

For Answer 2, does that mean we need to finetune a pretrained FP model on TinyImageNet first?