Paper Question: Use compression idea and generate same result as multi-branch model?

[oconnor127]

Hey,

I am just curious if I get this right. Could I just use your re-parametrization approach as a "compression approach" and thus make those multi-branch networks (e.g. ResNet, MobileNet etc.) plain and gain speed improvements while achieving similar accuracy ?

[DingXiaoH]

We cannot make an arbitrary multi-branch model single-branch because the conversion will not be equivalent if the model has nonlinearity before the addition of branches. We must design the structure properly to allow such equivalent conversion.