changsn / STViT-R

This is an official implementation for "Making Vision Transformers Efficient from A Token Sparsification View".
MIT License
30 stars 3 forks source link

你好,我没有看到semantic token S1 加上 global center G的操作,是在实际操作中省略掉了吗? #2

Open 201722111112027 opened 11 months ago

changsn commented 11 months ago

写在这里了 https://github.com/changsn/STViT-R/blob/main/models/swin_transformer.py#L526C57-L526C77 但是这个操作在Swin上用处很小,因为Swin是带window的,在DeiT这种没window的用处很大。 代码很乱,是我做research时候直接拿过来的,一直到现在就一直没有卡来整理代码了,但是代码都是能直接跑的。

201722111112027 commented 11 months ago

感谢!懂了。

---原始邮件--- 发件人: @.> 发送时间: 2023年8月16日(星期三) 晚上10:42 收件人: @.>; 抄送: @.**@.>; 主题: Re: [changsn/STViT-R] 你好,我没有看到semantic token S1 加上 global center G的操作,是在实际操作中省略掉了吗? (Issue #2)

写在这里了 https://github.com/changsn/STViT-R/blob/main/models/swin_transformer.py#L526C57-L526C77 但是这个操作在Swin上用处很小,因为Swin是带window的,在DeiT这种没window的用处很大。 代码很乱,是我做research时候直接拿过来的,一直到现在就一直没有卡来整理代码了,但是代码都是能直接跑的。

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>