hzzone / adatrans Goto Github PK

the det_face is always None . do you have any suggestions?

ps: The download link https://sota.nizhib.ai/pytorch-insightface/iresnet34-5b0d0e90.pth for the pre-trained IResNet model is invalid, causing an HTTP Error 404: Not Found. Could you please provide alternative methods to obtain the IResNet model. I really appreciate for your help and look forward to your reply!

Hello, thanks for your interesting work!
I have a question about strength of an edit (For example, in the teaser figure, where we can edit "old" attribute by specific age-number: 10 years old, 70+ years old, etc.)

As far as I understand, we need to manually edit the Facial Attribute vector $a$.
For example, given an initial image $I$ with corresponding facial attribute $a= [age, glasses,...] =[10, 1, ...]$.
I want to edit image $I$ into image $I'$ which age = 70.

Then I will have to edit the corresponding facial attribute $a$ into $a'= [age, glasses,...] =[70, 1, ...]$. Is that correct?

Looking forward for your response. Thank you!

Hello, thank you very much for providing the pre-trained models, which are very helpful.

I have a few additional questions. Firstly, I am unsure of the meaning of this sentence in section 4.2.1 of the paper: "Note that we train all attributes presented in Fig. 4 in a unified model". Aren't binary attributes supposed to be trained separately from one-hot conditions like hair color or style? Could you please give me some hints about how you conduct the training when performing the experiment?

Secondly, I can't find the code for some experiments mentioned in the paper, such as "Editing with 128 labeled samples only" and "nonlinear editing with a fixed step size at each step". Could you please provide the code or give me some guidance on the code implementation for these specific experiments?

Thank you very much, and have a great day :)

My E-mail: [email protected]

Hello, thank you so much for your previous help; it gives me a deeper understanding of the paper.

I'm confused by these two sentences in section 4.1 of the paper:

"For the fixed attribute classifier, we trained the last linear layer of ResNet-50 [15] on the binary attributes of the CelebA dataset [28] and the discrete age labels of FFHQ from [30]. We train another classifier from scratch for better classification performance and obtain the attribute labels of the FFHQ dataset".

1.Does that mean the weights of the classifier used to generate attribute labels are different from the one computing the Lmi loss?

2.Which classifier is the provided "Train the classifier" code aim to train? And could you please give me some guidance on the code implementation for training the other classifier?

3.In the "Editing with limited labeled data" experiment, are the attribute labels for the FFHQ dataset also generated by a classifier trained only on 128 samples, or by a classifier trained normally on the CelebA dataset?

Thank you very much, and have a great day :)

My E-mail: [email protected]