microsoft / CSWin-Transformer

CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped, CVPR 2022
MIT License
539 stars 78 forks source link

The num_heads of CSWin-B in the paper and code are not the same #18

Open 282857341 opened 2 years ago

282857341 commented 2 years ago

Hi!

Your work is great. But I’m confused with the num_heads in the CSWin-B. In the cswin.py, the num heads is [4,8,16,32] https://github.com/microsoft/CSWin-Transformer/blob/d8be74a7833898f7bd9c77eb8c051d1b8bd5d753/models/cswin.py#L392

But in your paper, the num heads of the CSWin-B is [2, 4, 8, 16].

LUO77123 commented 2 years ago

CSWin-B 96 2,4,32,2 1,2,7,7 4,8,16,32 78M 15.0G是一样的啊