Added colab example

[sdtblck]

Added an example colab notebook exploring stylegan2 networks using ganspace results for those of us without beefy GPUs.

[woctezuma]

I would suggest to suppress some outputs, typically the very long logs, because it makes it more cumbersome than necessary to read the notebook: I have to scroll a lot.

[harskish]

Thank you @sdtblck for the contribution! A few comments:

• It would be better to programmatically create a new environment.yml instead of keeping a copy around in the notebook (e.g. !sed -e 's/ganspace/base/g' environment.yml > colab_env.yml)
• There is no need to install pycuda, cython, and glumpy, they are only used by the local interactive viewer. Removing these should speed up the initial setup
• The direction exploration is missing a crucial part of the algorithm: limiting the edits to a specific layer range. Leaving this out would give users an incorrect picture of the method - PCA alone only gets you so far.
• The exploration should use the pytorch models. This way, someone interested only in exploring the bundled models doesn't have to perform an unnecessary conversion to tensorflow.

Using the pytorch models is super easy:

from models import get_model

model = get_model('StyleGAN2', 'ffhq', torch.device('cuda'), use_w=True)
w = model.sample_latent(1, seed=seed).cpu().numpy()
w = [w]*model.get_max_latents() # one per layer

# Edit
for l in range(start, end):
    w[l] = w[l] + lat_comp[idx]*lat_std[idx]*delta

out = model.sample_np(w)

I'd be happy to merge once these things are fixed. Let me know if you have questions or need assistance!

[sdtblck]

Hey!

Thanks for the feedback. I put the notebook together very quickly just for personal use, hence the roughness. The exploration does look much simpler in pytorch I just wasn't at all familiar with the interface, so thanks for that snippet of code.

[sdtblck]

This should work!

No need to mess with conda/environment files at all in the end

[harskish]

I've modified the notebook to use (start, end) instead of (start, start + n_layers), since that matches the rest of the codebase better.

There's one more thing I still want to fix: the layer ranges of named edits are currently not stored, which means that the final widget where saved edits are used might not work as intended.

[sdtblck]

Ok, quite a simple fix - just added the layer ranges to the named directions dict. This should work fine. It would be nice also to have a slider controlling each direction, but I'm not handy enough with the ipnyb widgets to get that to work, and it's probably better to do that on your GUI anyway.

[harskish]

Awesome, looks good. Thanks a lot for your contribution Sid, will merge this now!