search

fudan-zvg / SeaFormer

[ICLR 2023] SeaFormer: Squeeze-enhanced Axial Transformer for Mobile Semantic Segmentation
282 stars 21 forks source link

hi,I noticed that you comparison with different selfattention modules on ADE20K. val set based on Swin Transformer architecture in Table.4 #8

Open qqwewewejekl opened 1 year ago

qqwewewejekl commented 1 year ago

did you mean pretrain on imagenet and finetune on ade20k when you replace a new attention module?

wwqq commented 1 year ago

Hi, @qqwewewejekl Yes. We set the same training protocol, hyperparameters, and model architecture configurations as Swin.

qqwewewejekl commented 1 year ago

Thanks a lot.