Interpolation between dogs and cats about homointerpgan HOT 7 CLOSED

Hi ligaozhe, in our implementation, we create labels "is_cat" and "is_dog", so that the method can work like celeba or rafd.

from homointerpgan.

Thanks for your reply! But I still don't understand how to produce the intermediate result. If I only set two label, is_cat and is_dog, does the selective vector in interp_net only contain two items? Does the model only produce 2 result in this case? For example, If I want to produce 5 intermediate result like you did in paper, how should I do? Thank you so much.

from homointerpgan.

Although the labels are binary, with the Homomorphic loss, we have modeled the intermediate regions between cat and dog domains. So during testing, we can simply use floating-point labels to generate intermediate results.

Note that RaFD and CelebA are similar, i.e., both of them have binary labels, but we can synthesize intermediate results.

from homointerpgan.

OK, I will try it. Thanks for your answer!

from homointerpgan.

I have run the code. But I am still confused. What is the "floating-point labels"? Could you please describe it in the code? If I use only 2 attributes, I found selective_vector only contains 3 items in the function 'test_selected_curve', so which variable should I modify if I want to generate imterdediate results more than 3?

from homointerpgan.

OK，I got it. And I have reproduced your results in cat2dog dataset. Thanks for your work!

from homointerpgan.

Hi, I want to reimplement the cat2dog continuous translation results too.
Have you reproduced that? Could you share some experience with that?
@ligaozhe
Thanks!

from homointerpgan.