How to use Learn_direction_in_latent_space.ipynb

[shouramo]

Hi,

I am just wondering, how am I supposed to use Learn_direction_in_latent_space.ipynb?

What should "data.tsv" be? How should I organize my data here?

Thanks!

[EvgenyKashin]

Hey, good question.

I agree that right now it is't clear. After using the stylegan2/run_generator.py (as in the readme) you will get images and latents with names like 100.npy and 100.png (according to this line).

After using your attributes classifier you have to create data.tsv with key column representing index of generated image (100 in the example above). And with label column that in our example in the notebook contains male/female labels.

Hope it would help!

[shouramo]

Thanks for such a quick response!

So, I would have 2 columns; the first indicating the index (just numbers), the second indicating the attribute label(gender, age, etc.)

For the attributes classifier, which one did you use? I know of deepFace and that works pretty well from my experience, but just curious which you have used to create your data.tsv!

I am using my own styleGAN for this that I trained to generate 256x256 images, corresponding to (14,512) latent outputs, so the directions provided [stylegan2directions] unfortunately do not work for me.

Thanks!

[EvgenyKashin]

Yep, I see. We used internal attributes classifier from my previous company Yandex. But for now I would recommend to use CLIP classifier. It's pretty easy to create a classifier by just describing it. For example, a prompt like "image of a smiling face" and "image of a neutral face" to classify emotions.

[shouramo]

Amazing! Thank you so much for all of this info.

Take care,
Moaz