Skip to content

Latest commit

 

History

History
89 lines (75 loc) · 3.66 KB

PCA.md

File metadata and controls

89 lines (75 loc) · 3.66 KB
Raw

PCA

Explore the latent space by checking components that affect a specific part through PCA
We used GANSpace!
Original Code


eyes
ears
mouth
face features

Requirements

Setting Up for GANSpace

Pretrained Model

Prepare models learned from freezeD experience.
Convert this pkl model to pt. (Converter)

Config

Open models/wrappers.py,
and edit the stylegan2 configs dict on line 110 to include your model and its corresponding resolution.

I.E from

# Image widths
configs = {
    'ffhq': 1024,
    'car': 512,
    'cat': 256,
    ...
}

to

# Image widths
configs = {
    'anti-cat': 256,
    'ffhq': 1024,
    'car': 512,
    'cat': 256,
    ...
}

Then copy your pytorch model over to your drive account or any other hosting platform, and add the direct download link to the checkpoints dict in the download_checkpoint function on line 136.

def download_checkpoint(self, outfile):
    checkpoints = {
        'anti-cat': 'https://drive.google.com/uc?export=download&id=1JxgW_zoVww4hXO0G4PO7e_w3rFtI7jsG',
        'ffhq': 'https://drive.google.com/uc?id=12yYXZymadSIj74Yue1Q7RrlbIqrXggo3',
        'car': 'https://drive.google.com/uc?export=download&id=1iRoWclWVbDBAy5iXYZrQnKYSbZUqXI6y',
        'cat': 'https://drive.google.com/uc?export=download&id=15vJP8GDr0FlRYpE8gD7CdeEz2mXrQMgN',
        ...
    }

Explore Latent Space

PCA Options
  ├  --model='StyleGAN2'
  ├  --class='anti-cat'
  ├  --layer='style'
  ├  --use_w
  ├  --components=32

We extracted 32 principal components and checked the effect of each component.

Result


eyes
ears

mouth
face features		 
We have identified Principal components that affect certain areas of the face!