SeNet forward process remove squeeze-and-excitation architecture？

[chaipangpang]

Hi，i have some consideration about squeeze-and-excitation architecture while forward process，when senet is trained， squeeze-and-excitation architecture had modify the weights of resnet，so，it is necessary for retain squeeze-and-excitation architecture when forward process？

[hujie-frank]

The channel-wise weights in each SE module depend on the input featuremaps. It is not able to squash SE modules by any modification on the weights of backbone from a well-trained SE model.