taki0112/SENet-Tensorflow

se should be added after inception block, right?

dpcross opened this issue · 4 comments

dpcross commented

hi, I think squeeze_excitation_layer should be added after inception block ,not inception_resnet block
Am I right ?

taki0112 commented

you mean Reduction A, B, C ?

dpcross commented

@taki0112 not reduction a,b,c . I mean it should be like :
x = squeeze_excatation_layer(x)
x = x * 0.1
x = init + x
should not be added after the inception_resnet_a ,b ,c
I change the structure and it performance much better

  • 👍2
taki0112 commented

The SE block does not matter where you put it.
In the case of keras code, he put it after the inception block you mentioned and put it after the reduction block.

dpcross commented

Actually I am not quite sure whether it matters or not , but the resnet module show here https://github.com/hujie-frank/SENet/blob/master/figures/SE-ResNet-module.jpg .
so I thought the module define like this might be better and it did work in my project.