BoilerPlate
BoilerPlate copied to clipboard
ReLU after residual branch missing?
trafficstars
In the paper you can see the diagram of Fixup which applies the residual branch, and adds the output to the original input. Finally it applies a ReLU to this sum.
https://i.stack.imgur.com/T67F3.png
In the code I cannot see this being applied anywhere (once for each block) - is this an oversight or am i missing something?
More specifically, why is it
return self.residual(x) + self.shortcut(x)
in https://github.com/ajbrock/BoilerPlate/blob/master/Models/fixup.py#L45
and not
return self.activation(self.residual(x) + self.shortcut(x))