Question about discriminator

[07hyx06]

Thanks for your great work..But i have some problems about the discriminator..

Does the discriminator (appears in encoder training process) is the StyleGAN discriminator or a discriminator trained from scratch with the encoder?

[07hyx06]

What about the result without training with a discriminator(just compute the pixelwise loss, remove the adverisal loss and the D loss)?

[zhujiapeng]

Here, the discriminator is just fine-tuned from the original StyleGAN's discriminator. The inverted images will be blurry if without a discriminator.

[07hyx06]

Here, the discriminator is just fine-tuned from the original StyleGAN's discriminator. The inverted images will be blurry if without a discriminator.

Thanks for your reply! Another question, in train_encoder.py, what's the meaning of total_kimg? Is it means the size of real image dataset is 14000 and train it 1000 times?

[ShenYujun]

@07hyx06 This follows the official StyleGAN implementation. Basically, it means how many thousands of images will be seen by the generator (or encoder in this repo). Suppose having batch_size = 10 for each iteration, then total_kimg = 2 means 2 * 1000 / 10 = 200 iterations.