xiaolai-sqlai / mobilenetv3

mobilenetv3 with pytorch,provide pre-train model
MIT License
1.6k stars 340 forks source link

SE's position differs from paper. #26

Open TKONIY opened 4 years ago

TKONIY commented 4 years ago

image image

SE should be placed after Depthwise and before Pointwise. However, you set it after Pointwise. Is there any reason or just a mistake?

rwightman commented 4 years ago

@TKONIY This is one of several problems, see an earlier issue #8

rebeen commented 4 years ago

I changed the place of SE and I put before out = self.bn3(self.conv3(out)) but it makes an error

TKONIY commented 4 years ago

@rebeen It is not enough. Check the SE class and network class, we can see that SE takes a parameter of input channels, which should match the output channels of dw. It should be modified because now it matches the output channels of pw.

rebeen commented 4 years ago

Thank you very much I will try to tackle this issue