ConvModule with layernorm #3227
Unanswered
Stephanie-Shen324
asked this question in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi, is it possible to use layernorm in convModule?
nn.layernorm expects a tensor with channel last (N, H, W, C) where normalized shape is (C). However the convModule's input/output are channel-first (N, C, H, W). hence running convModule with norm_cfg(type='LN') will give the error of "Given normalized_shape=[256], expected input with shape [*, 256], but got input of size[batch_size, 256, H, W]"
However, I did get it running with no errors with norm_cfg(type='GN', num_groups=1). And from the following image, group norm with 1 group should be equivalent to layernorm?
Please let me know if i made a mistake somewhere and how can I make it work for norm_cfg(type='LN') in ConvModule.
Beta Was this translation helpful? Give feedback.
All reactions