style-based-gan-pytorch
style-based-gan-pytorch copied to clipboard
rosinality
what is the important of the below lines?
if mixing and random.random() < 0.9:
gen_in11, gen_in12, gen_in21, gen_in22 = torch.randn(4, b_size, code_size, device='cuda').chunk(4, 0)
gen_in1 = [gen_in11.squeeze(0), gen_in12.squeeze(0)]
gen_in2 = [gen_in21.squeeze(0), gen_in22.squeeze(0)]
else:
gen_in1, gen_in2 = torch.randn(2, b_size, code_size, device='cuda').chunk(2, 0)
gen_in1 = gen_in1.squeeze(0)
gen_in2 = gen_in2.squeeze(0)
It is mixing regularization in the paper. Mixing regularization makes model robust when more than 1 latent code used for generation. (style mixing.)
could you also explain what is inject_index and crossover do.
if len(style) < 2:
inject_index = [len(self.progression) + 1]
else:
inject_index = random.sample(list(range(step)), len(style) - 1)
crossover = 0
for i, (conv, to_rgb) in enumerate(zip(self.progression, self.to_rgb)):
if mixing_range == (-1, -1):
if crossover < len(inject_index) and i > inject_index[crossover]:
crossover = min(crossover + 1, len(style))
style_step = style[crossover]
To do mixing regularization, you should choice some layers (index), and use secondary latent codes after that layer. inject_index and crossover is for implement this.
style_step = style[crossover] gives learned style from latent code, rather layer where one need to mix.
I am sorry, I am very much stuck on the same line.
Yes, it works by get style codes using index of layers and inject it to that layer.