The wrn prototxt file in cls
zhanglonghao1992 opened this issue · 4 comments
zhanglonghao1992 commented
deploy_wrn50-2.prototxt doesn't seems like the real ‘wrn’ network,its just the nomal resnet-50 i
think..
soeaver commented
Check carefully, it is really wider than resnet50
zhanglonghao1992 commented
@soeaver WRN use dropout between two 3*3 conv in a residual block, but in this prototxt I just see nomal bottleneck conv with bn.
soeaver commented
Just for small dataset, please read the paper and source code again.
zhanglonghao1992 commented
@soeaver I see.. Its quite interesting that two 3*3 suits small datases well but Imagenet needs bottleneck..