search

Kevinz-code / CSRA

Official code of ICCV2021 paper "Residual Attention: A Simple but Effective Method for Multi-Label Recognition"
GNU Affero General Public License v3.0
213 stars 37 forks source link

Specific improvements to vit #15

Open sure7018 opened 2 years ago

sure7018 commented 2 years ago

Hello, how to use CSRA in vit? There is no specific information in the paper？

Kevinz-code commented 2 years ago

Hi, thanks for reading. The 'main.py' in our codebase describes the setting of wider attribute dataset using VIT-CSRA models, say, step_size=5, lr_backbone=0.1, lr_fc=0.01. You can follow the same settings when using VIT-CSRA models in wider attribute dataset.

sure7018 commented 2 years ago

What improvements have your paper made to the original model of Vit?

Kevinz-code commented 2 years ago

Hi, @sure7018, I'm not sure I fully understand your question. You mean the improvement of the VIT structure, or the improvement of our VIT-CSRA model compared with orginal VIT model? In the later case, refer to the Table 3-4 in MS-COCO and wider-attribute datasets, which clearly demonstrate the performance gain of our method.

Best, Author