minyoungg / pix2latent Goto Github PK

Hello, sorry for asking too often.

I am curious about doing optimization in batch input in the code.

I think in actual use, batch size should be one and this code just show the result of diverse initialization results, is it right?
Or, is it better to do it in batch for the results in the case of basincma or adam?

Hello,

I want to try pix2latent on the FFHQ dataset on Google Colab. Due to RAM constraints, Colab won't run the optimization process with CMA or BasinCMA (unless I use the cars dataset), so I have to go with the faster (yet worse) option relying on Nevergrad.

I see that:

• Nevergrad is gradient-free optimization (CMA by default), followed by ADAM fine-tuning, so that would be similar to:

• HybridNevergrad alternates gradient-free optimization (CMA by default) and SGD optimization. That would be akin to the following, albeit with SGD instead of ADAM:

Between the two options (Nevergrad vs. HybridNevergrad), which one would you recommend?

Edit: Below are results obtained with Nevergrad .

Edit: Below are results obtained with HybridNevergrad.

I guess I would have to try another portrait, tweak parameters, or forget Colab and stick to CMA/BasinCMA on a local machine.

why do you apply remove_spectral_norm for biggan?

Thank you for sharing your code!! It's a cool work.

Is there a code using encoder network? I can't fine them..

https://github.com/richzhang/PerceptualSimilarity was recently updated,
this broke the import models that is called from the examples.

Traceback (most recent call last):
  File "invert_biggan_adam.py", line 44, in <module>
    loss_fn = LF.ProjectionLoss()
  File "/home/r/.local/share/virtualenvs/single_view_mpi-aYhVwZ1J/lib/python3.6/site-packages/pix2latent/loss_functions.py", line 90, in __init__
    self.ploss_fn = PerceptualLoss(net=lpips_net)
  File "/home/r/.local/share/virtualenvs/single_view_mpi-aYhVwZ1J/lib/python3.6/site-packages/pix2latent/loss_functions.py", line 134, in __init__
    import models
ModuleNotFoundError: No module named 'models'

I believe now you should call something like lpips.LPIPS(params) to instantiate a loss object.
(model='net-lin' is also not available anymore)
I'm going to try and clone an older version of the repo.

Update: git checkout 6abcdd1077b090cb9f0892103b45d56531d50689 seems to do the trick.

Can you please share your code and images for STYLEGAN2 FFHQ face example?
thanks

Hello,
I have a question，what is the mask in line 107 and line 108 in the file invert_biggan_with_transform.py,

target_transform_fn = SpatialTransform(pre_align=mask)
weight_transform_fn = SpatialTransform(pre_align=mask)

when I run this program, the following error occurs:

Traceback (most recent call last):
  File "invert_biggan_with_transform.py", line 107, in <module>
    target_transform_fn = SpatialTransform(pre_align=mask)
NameError: name 'mask' is not defined

How can I solve it？Thank you.

Hi there, excellent job! Thanks for providing the code!

I have a question about fine-tuning. In Section 3.5 of the paper, I think there is a regularization term that can also optimize the weights of the original GAN. Where can I find this part in your code? Currently, I just saw the optimization of z,c, and ф. Did I miss something? Please correct me if I was wrong.

Thank you again!

I want to invert images sized 256x256 into .pt through stylegan2, however it reports 【RuntimeError: mat1 dim 1 must match mat2 dim 0】. I know it is because original model is for images sized 512x512, but how should I change it?