Unmatch between code and paper

[Rbrq03]

Hey there. Congrats for the great work!
My question is, in paper, you mentions the fusion in section 2.2that

we use 1x1 convolution and standard upsampling operation (e.g., bilinear/bicubic upsampling) to match their spatial and channel size and fuse them via addition.

However, it seems that in this repo, the code directly feed the feature produced by the final stage to the head, instead of using fused feature. Am i miss anything? If not so, what' s the reason for the different between code and paper?

[lizhe1531]

I have the same doubts. There is no upsampling process in the classification model. What is the reason?