The training hyper-parameters about white-box-cartoonization HOT 6 CLOSED

@Xinxiang7 maybe you can try to train the model with default parameters given in the code (1e4 for tv loss, 0.1 for d_loss_blur, 1 for d_loss_gray and 2e2 for content_loss, the ratio between each terms is the save). for the selective search, we at first normalize the pixel values to (0, 1). I guess there may also be some problem in terminology (maybe 1.25 in my setting means 0.8 in your setting)

from white-box-cartoonization.

@SystemErrorWang Thanks a lot. I will try sooner.
I found the function of texture in the paper is different from the one in the code
In the paper:

the code:
output1 = (b_weight * b1 + g_weight * g1 + r_weight * r1) / (b_weight + g_weight + r_weight)
Does the standard grayscale image Y is not important. So It can be removed away or some other reason?

from white-box-cartoonization.

@Xinxiang7 the code for random color shift is in utils.py and it strictly follows this formulation. please refer to it for details.

from white-box-cartoonization.

@SystemErrorWang Thanks very much.
I might be wrong. the code in utils.py as follows:

I can't find the Y in the code. Does I miss something?

from white-box-cartoonization.

@Xinxiang7 they are mathematically equivalent.

from white-box-cartoonization.

@SystemErrorWang Thanks Very much for your reply!

from white-box-cartoonization.