yingkaisha/keras-vision-transformer

swin transformer depth

atgc1984 opened this issue · 1 comments

atgc1984 commented

Hello,
Thank you for the example on MNIST classification. The demo seems to have only one downsample stage and it is enough for this case. How can I config a larger model to have more downsample stages? Shall I just add more patch_extract/patch_embedding and block loop with smaller patch size?

Thank you.

STU-ECHO commented

Hello, Thank you for the example on MNIST classification. The demo seems to have only one downsample stage and it is enough for this case. How can I config a larger model to have more downsample stages? Shall I just add more patch_extract/patch_embedding and block loop with smaller patch size?

Thank you.

hi have you solve the problem you proposed? Ｉ got the same question as you ,could you give me some advices