BichenWuUCB/squeezeDet

FLOPS counter

Opened this issue · 0 comments

emezh commented

Hello!
Number of FLOPS for conv layer is calculated as
num_flops = (1+2*int(channels)*size*size)*filters*out_shape[1]*out_shape[2]

"1+" I assume is for bias, correct?
What's the reason for 2x in the formula above: Do you assume that multiply-accumulate is 2 FLOPs: one for multiply and one for add?

Thanks