search

OliverRensu / TinyMIM

151 stars 7 forks source link

How to reproduce your segmentation result in paper? #3

Closed Gregory1994 closed 1 year ago

Gregory1994 commented 1 year ago

Hi, thanks for your great work!

I'm interested in your outstanding segentation result in TinyMIM-tiny, and try to reproduce it. But I only get 38.87 mIoU on ade20k dataset, considerably lower than 45.0 mIoU reported in your paper.

I did my experiment on mmsegmentation and use TinyMIM-FT-Tstar.pth as pretrained model.

So would you please tell me:

  1. how can I reproduce your segmentation result?
  2. can you provide training setting or training log or segmentation checkpoints?

Best wishes!

OliverRensu commented 1 year ago

There is slight difference between TinyMIM and TinyMIM-T* including distillation tokens, number of heads, and an extra fully connected layer. The code will be released soon. Please email me if you need the log and the code that I used but is not ready to be published in github for reference.

Gregory1994 commented 1 year ago

Thanks for your reply! i've emailed you and looking forward your code and log! I will close this issue.