texture loss weights

[r1ckya]

Сould you explain where these coefficients come from? Thanks.

SRNTT/SRNTT/model.py

Line 395 in 7b2178e

self.a1, self.a2, self.a3 = -20., -20, -20

[ZZUTK]

We got them from experiments, i.e., the center and slope of the sigmoid-like function are set according to the mean and variance of the similarity between matched patches. More specifically, the center could be considered as the threshold to separate similar and dissimilar matchings, and the slop controls the smoothness/sharpness of the transition between two ends. Actually, we considered learning those params during training, which should be more adaptive than manual setting.

[r1ckya]

I am confused about a_i being < 0, in the paper you propose to weight loss by the best matching score between normalized content and normalized conditional patches (as I understand). And it seems reasonable to penalize dissimilar patches less, but setting a_i < 0 seems to lead to opposite effect (sigma(-20weights + 0.65) is decreasing function), could you help me to understand your choise of setting a_i < 0?