How to use Multi-Code method for StyleGAN?

[zhufeida]

Hi, in derivable_generator.py, several configurations are listed. However, there is no multi-code configuration for StyleGAN. In fact, StyleGAN's design is different from pggan, where w is inserted into each middle layers of the generator by AdaIN. How to use multi-code scheme in StyleGAN scenario?

def get_derivable_generator(gan_model_name, generator_type, args):
    if generator_type == 'PGGAN-z':  # Single latent code
        return PGGAN(gan_model_name)
    elif generator_type == 'StyleGAN-z':
        return StyleGAN(gan_model_name, 'z')
    elif generator_type == 'StyleGAN-w':
        return StyleGAN(gan_model_name, 'w')
    elif generator_type == 'StyleGAN-w+':
        return StyleGAN(gan_model_name, 'w+')
    elif generator_type == 'PGGAN-Multi-Z':  # Multiple Latent Codes
        return PGGAN_multi_z(gan_model_name, args.composing_layer, args.z_number, args)
    else:
        raise Exception('Please indicate valid `generator_type`')

[yue95213]

Same issue!

[JasonGUTU]

Hi, @yue95213 and @zhufeida. Thank you for your question. It worth noticing that the stylegan has multi-code design natively (the style code for each layer) and the stylegan-w+ is for the inversion using these codes. Use multi-code strategy on stylegan is equivalent to an over-parameterization version of stylegan-w+ inversion. So in the released code, we recommend using stylegan-w+.

[betterze]

What parameters you are using for invert stylegan? I use

python multi_code_inversion.py --gan_model stylegan_bedroom 

--target_images ./examples/gan_inversion/bedroom 

 --outputs ./gan_inversion_bedroom2 

--inversion_type StyleGAN-w+

However, the results are not very realistic. What parameters should I use?

[JasonGUTU]

What parameters you are using for invert stylegan? I use

python multi_code_inversion.py --gan_model stylegan_bedroom 

--target_images ./examples/gan_inversion/bedroom 

 --outputs ./gan_inversion_bedroom2 

--inversion_type StyleGAN-w+

However, the results are not very realistic. What parameters should I use?

Have you tried on face images? How does the face experiment work?
And try to use Norm initialization. The initialization method is influential for stylegan.

[Xiaomao136]

python multi_code_inversion.py
--gan_model stylegan2_ffhq
--target_images ./examples/face
--outputs ./gan_inversion_face
--composing_layer 8
--z_number 20
--iterations 3000
--inversion_type StyleGAN-w+
--init_type Normal
--optimization Adam
--loss_type VGG

Traceback (most recent call last):
File "multi_code_inversion.py", line 112, in
main(args)
File "multi_code_inversion.py", line 37, in main
latent_estimates, history = inversion.invert(generator, y_gt, loss, batch_size=1, video=args.video)
File "/content/drive/My Drive/stylegan2/source code/mganprior/inversion/inversion_methods.py", line 24, in invert
input_size_list = generator.input_size()
File "/content/drive/My Drive/stylegan2/source code/mganprior/derivable_models/derivable_generator.py", line 77, in input_size
return [(self.stylegan.net.synthesis.num_layers, 512)]
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 772, in getattr
type(self).name, name))
torch.nn.modules.module.ModuleAttributeError: 'StyleGAN2GeneratorNet' object has no attribute 'net'

[liuliuliu11]

@Xiaomao136 Hello, have you solved this bug?

[Xiaomao136]

Please refer to genforce/idinvert project, I switched to genforce/idinvert project liuliuliu11 <[email protected]> 于2020年10月9日周五 下午4:47写道：

…

@Xiaomao136 <https://github.com/Xiaomao136> Hello, have you solved this bug? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#3 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMU4MTFFQ22ENHNWNQVB663SJ3E3NANCNFSM4MEXG7SA> .

[fungtion]

@JasonGUTU using stylegan_w+ for face is still far from ideal, can you show that how to split stylegan? I did that following PGGAN-Multi-Z, but it reported error that in x = self.epilogue(x, w), the shape of w is desired to be [1, 512], but actually [1, 18, 512], I've no idea what happened.