ms_ssim calculation about pytorch-msssim HOT 3 CLOSED

Hi @assafge. Maybe there is a missing "**weights[-1]" at the end of your code? Then your code is equivalent to line_195,

Note that mcs[-1] = cs

msssim_val = torch.prod((mcs ** weights.unsqueeze(1)), dim=0) * (ssim_val / cs) ** weights[-1]
           # here we move the last cs to the end 
           = torch.prod((mcs[:-1] ** weights[:-1].unsqueeze(1)), dim=0) * (ssim_val / cs * cs)**weights[-1]
           = torch.prod((mcs[:-1] ** weights[:-1].unsqueeze(1)), dim=0) * ssim_val ** weights[-1]

if α=β，

from pytorch-msssim.

nop, all cs in the power of relevant weight. I had some negative numbers in my ssim values resulting nan values in the ms_ssim, for now my calculus is:
msssim_val = torch.prod((torch.sign(mcs) * abs(mcs) ** weights.unsqueeze(dim=1)), dim=0) * (ssim_val / cs)

according to exponents-less-than-1-return-nan

from pytorch-msssim.

Sorry for the late reply. Is it helpful to use a large C2? the negative ssim and cs is caused by the negative sigma_xy.

from pytorch-msssim.