neural-style-tf
neural-style-tf copied to clipboard
Vgg weights not normalized
The vgg weights have not been scaled as described in Section 2 of the paper.
This is probably what's making the results non-reproducible (if you try to use the style and content loss weights that Gatys et al have mentioned in their papers).
I noticed this as well, would normalization also help alleviate some of the issues with visual artififacts and chromatic aberrations at low iteration counts?
I'm also thinking if inputs shouldn't be normalized, pixel values for images are in range around [-150,150] and it makes the value of functions for f
and g
in lbfgs
enormous (around 1e11), on linux I'm getting WARNING: ROUNDING ERRORS PREVENT PROGRESS
warning on linux machine, and it does not converge.