Wrong with ResBlock?

[CuriousCat-7]

The original and advanced types of resnet blocks are list in there: http://torch.ch/blog/2016/02/04/resnets.html

x + R(I(C(R(I(C(x))))) should be changed to R(x + I(C(R(I(C(x)))) )

for

• R: relu
• I: InstanceNorm
• C: Conv2D

[leocvml]

Hi , Sorry for the late reply. i made experiment on Activation and InstanceNorm , but i forgeted to replace the Resblockv2 :(( in this paper : Identity Mappings in Deep Residual Networks it proposed model (e) (ResBlockv2):

i will change to ResblockV2 x + R(I(C(R(I(C(x))))) = > x + C(R(I(C(R(I(x))))))

thanks your commented

[CuriousCat-7]

Hi , Sorry for the late reply. i made experiment on Activation and InstanceNorm , but i forgeted to replace the Resblockv2 :(( in this paper : Identity Mappings in Deep Residual Networks it proposed model (e) (ResBlockv2):

i will change to ResblockV2 x + R(I(C(R(I(C(x))))) = > x + C(R(I(C(R(I(x))))))

thanks your commented

Thank you for your reply. it makes more sense. However, please note that other famous open sources cyclic gan projects use ResNetBlockV1【x + I(C(R(I(C(x)))) 】, and it seems accepted by others. It would be great to reproduce there method. As far as my experiment, the model acts better after I change the RestNetBlock in the code.

Best regard, Neo