Why you did w = pospowbias(x/xn) instead of w = pospowbias(x) as your paper did

[shizenglin]

Hi, I'm confused about this code in positive bias.py, why did you do x/xn before geting w, I can't see this in your paper, in which you did w = pospowbias(x) instead of w = pospowbias(x/xn) It = F.upsample(F.avg_pool2d(I, 2), scale_factor=2, mode='nearest') x = ((I-It)**2)+1e-3 xn = F.upsample(F.avg_pool2d(x, 2), scale_factor=2, mode='nearest') w = pospowbias(x/xn) kp = F.avg_pool2d(w, 2) Iw = F.avg_pool2d(I*w, 2)

[farazsaeedan]

normalizing "x"s is done for increased numerical stability.