haibo-qiu / synface Goto Github PK

Hi, I'm trying to use your repository, but I have issues in complying with both the requirements for SynFace and DiscoGAN. For example, the CUDA requirements for tensorflow=1.12 (required by DisoGAN) and for pytorch=1.8.1 (SynFace) conflict with each other. How did you work around these issues? Do you have any advice for how to work with both SynFace and DiscoGAN? Thanks.

Could you please share network-snapshot-020126.pkl? Thanks.

Hi,

Would you prefer to share the synthetically generated dataset which is used in experiments?

Did you have any trouble generating face images with discogan?
I tried both their original code, and your test code, but I all I get is noise

The discoGan repo does not seem maintained at this time, so wondering if you can help me

Hi,

I downloaded the LFW dataset and aligned the faces using your script data/imgs_crop/face_align_crop.py.

Then I measured the accuracy using eval.sh.

However, I am getting the following accuracy, which is different from the values reported in the paper.

86.25% for model_10k_50_nomix_8898.pth.tar
89.27% for model_10k_50_idmix_9197.pth.tar
92.70% for model_10k_50_2k_10_9578.pth.tar
94.27% for model_10k_50_2k_20_9765.pth.tar

Are the models different from the ones used in the paper?

I'm questioning that does the method substantially improve the accuracy of face recognition.

If I understand your pipeline correctly,
the generative model was trained on WebFace which contains 500,000 samples. Then generate training images with the generative model. The recognition model is trained with synthetic images.

But why not directly train the recognition model on WebFace?

Does the whole pipeline substantially improve the accuracy?

There is some problem with your script，data root he sometimes doesn't work.And you did not specify the version of the downloaded dataset，This wastes a lot of unnecessary time。

Hi, @haibo-qiu
Thanks for this incredible work and also your code. I'm curious about why you set the scale s =30 in the arcface loss.
Based on arcface loss paper, they set s = 64. Is there any specific reason for you to set s = 30?

BR,
Ziv