Does L_total = 0.5 * L_perceptual + 50.0 * L_style + 25.0 * L_tv only apply to optimize VGG?

[lovecambi]

https://github.com/HCIILAB/Scene-Text-Removal/blob/master/network.py#L78
https://github.com/HCIILAB/Scene-Text-Removal/blob/master/train.py#L167
The trainerV seems to only optimize parameters of VGG. So why this loss is needed?

https://github.com/HCIILAB/Scene-Text-Removal/blob/master/train.py#L157
This line I_comp = where(mask, real, fake) contradicts to the description in the paper where I_comp = where(mask, fake, real).

[zhangshuaitao]

@lovecambi，The trainerV,trainerD, and trainerG are trained alternately to optimize the results. for
I_comp = where(mask, real, fake)， 1 and 0 of the mask represent non-text and text areas, respectively.

[lovecambi]

@lovecambi，The trainerV,trainerD, and trainerG are trained alternately to optimize the results. for
I_comp = where(mask, real, fake)， 1 and 0 of the mask represent non-text and text areas, respectively.

So for this line https://github.com/HCIILAB/Scene-Text-Removal/blob/master/train.py#L119
the mask here also uses 1 to represent non-text region?