How to determine cond_fn in condition_mean about diffae HOT 4 OPEN

Can you please point to the line of the code in question?

from diffae.

I believe the line 403-416 was not used in DiffAE at all, and hence not related to DiffAE. It was a legacy code from the base repo that we built upon.

Regarding your second question, 0.3*sqrt(512). 0.3 is definitely tuned by hand based on qualitative results. sqrt(512) is related to the fact that F.normalize(cls_model.classifier.weight[cls_id][None, :], dim=1) is a unit vector (which is rather small, and depends on the number of dims). It's nice to have something that scales with dims (make the coefficient more robust). I think a Gaussian random vector with 512 dimensions has the norm of sqrt(512), multiplying a unit vector with sqrt(512) scales it as though it has the same size as a random Gaussian vector.

from diffae.

Thanks for your reply.

Yes of course. The condition_mean method is Line 403-416 in base.py

Also in the manipulate case:

cond2 = cond2 + 0.3 * math.sqrt(512) * F.normalize(cls_model.classifier.weight[cls_id][None, :], dim=1)

could you please share how you determine the coefficient (0.3*sqrt(512))?

Thanks for your help!

from diffae.

Really appreciate your reply.

But I think line 403-416 actually used in img = model.render(xT, cond2, T=100) from block[33] of manipulate.ipynb.

In which the render method can be conditioned based on cond_fn. But I don't find anywhere cond_fn be defined.

So could you please provide me some clue on how to implement this conditional render?

Thank you!

from diffae.