Reparameterization of RepVGG with SEBlocks #1809
Unanswered
Code-Gratefully
asked this question in
Q&A
Replies: 1 comment
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I tried to use the RepVGG.switch_to_deploy() function to convert the model. However, it does not seem to be working; the output of the converted model is dramatically different than the training version. I'm wondering if anyone has seen work to reparameterize the Squeeze and Excitation layers into the 3x3 convolution layers as well? Or is that simply impossible, and we need to keep the SE block there and cannot perform reparameterization?
Beta Was this translation helpful? Give feedback.
All reactions