Closed tulilin closed 11 months ago
Hi~Thanks for your interest in our work.
Actually, we extract the first attention map of the last block in Swin-Transformer and then interpolate to the original size of input images.
I have added the corresponding code (Pretrain/attn_vis_config.yaml
, Pretrain/main_attn_vis.py
, Pretrain/models/swin_transformer_with_attn.py
).
You may need to replace the original code in Pretrain/models/swin_transformer
with new Pretrain/models/swin_transformer_with_attn.py
.
This implementation is highly inspired and borrowed from DINO and iBOT. You may need to these repos for detailed information.
Closed as long periods of inactivity, feel free to reopen if there is any problem.
Hi~Thanks for your interest in our work. Actually, we extract the first attention map of the last block in Swin-Transformer and then interpolate to the original size of input images. I have added the corresponding code (
Pretrain/attn_vis_config.yaml
,Pretrain/main_attn_vis.py
,Pretrain/models/swin_transformer_with_attn.py
). You may need to replace the original code inPretrain/models/swin_transformer
with newPretrain/models/swin_transformer_with_attn.py
.This implementation is highly inspired and borrowed from DINO and iBOT. You may need to these repos for detailed information.
Got it! Thank you very much!
Thank you for your wonderful work of CMID and I have read your paper and code. In Fig.12 of the paper, the attention maps of CMID-Swin-B were analyzed. I am curious that how to obtain the attention maps of Swin Transformer considering that it does not have the [class] token similar to ViT? Will the code be provided? I am looking forward to your reply. Thank you!