What does mean_style & n_source & n_target mean?

[brianw0924]

just wondering what is the parameter in get_mean_style means...

and the function style_mixing, what does n_target, n_source mean?

[rosinality]

get_mean_styles will calculate average latent codes over gaussian noises.

n_target and n_source means the number of styles originates from and transferred to.

[brianw0924]

Thanks for replying, I have another question:

I have about 70k real img, how many "number of samples used for each training phases" should I set is the best?

my final output img is 64*64, I set the parameter 35k, and the face still have some distortion (just like a monster)

should I increase "number of samples used for each training phases" ? or should I increase n_critics?

(btw, I'm drawing anime faces)

[rosinality]

I think it would be enough to use default phase steps. It is just for transitions between each resolutions, and progressive training. You can train more on your target resolutions. (64px) And it would be better to check FIDs.

[brianw0924]

thank you.

FID is really good, but there's another threshold in my homework, which is detecting faces for given images

and my output seems generate some abnormal faces (e.g. wrong eye, mouth position...etc)

[rosinality]

Are you using truncation and ema generators?

[brianw0924]

i didn't find truncation , ema generators in your codes, should I add it?

i just use your original setting, and change the image size

[rosinality]

If you are using generate.py, it will use ema generator and truncation. You can try to increase truncation by decreasing style_weight in here: https://github.com/rosinality/style-based-gan-pytorch/blob/master/generate.py#L33

I think there not much things that can be used to vastly increase image quality.

[brianw0924]

Thanks!!! it works, the image quality dramatically increase!!