genforce/interfacegan

How to train a classifier to get attribute scores

UdonDa opened this issue · 2 comments

Hi,
I trained resnet18 using a celeba dataset with a multi-label setting.
And, I used it to find boundaries.

However, my results show that the generated images have an entangled boundary.
The image used an only 'makeup' class.
From left to right, the scalar weight of boundary is increased.

So, I guess that the training strategy of the classifier is different from your paper.
Would you provide the detail how to train a classifier?
Thanks.

image

Oh, I miss your conditional manipulation method...

can you send your repo link please ... ?