search

lucidrains / bottleneck-transformer-pytorch

Implementation of Bottleneck Transformer in Pytorch
MIT License
671 stars 82 forks source link

Feature map size #10

Closed benlee73 closed 3 years ago

benlee73 commented 3 years ago

hi In my case, the input images size are all different, so the feature map size keeps changing. In this case, how should the fmap_size parameter of BottleStack be set? Is it possible to learn with an unfixed feature map size?

lucidrains commented 3 years ago

unfortunately not with the way relative positional encoding is done atm

lucidrains commented 3 years ago

there is a way to do this with another type of relative positional encoding, but it would require deviating from the paper

benlee73 commented 3 years ago

thank you for the reply. We will preprocess the input image size through padding.