About the training on FFHQ

[AlbertHuyb]

Hello @Xuanmeng-Zhang , thanks for yoru excellent work.

I want to confirm some details of training on FFHQ.

1. Did you train with 8 GPU cards?
2. How long does it take to finish all the 3000 epochs?

Thanks

[Xuanmeng-Zhang]

Yes, 8 V100 for around ten days.

[AlbertHuyb]

Got it. Thank you.

[AlbertHuyb]

@Xuanmeng-Zhang , I'm also curious about the variables including h_mea=pi/2, h_stddev=0.3, v_mean=pi/2, v_stddev=0.155 and fov=12 degree in your metadata used when training on FFHQ dataset.
How did you get these statistics from FFHQ?

I'm not familiar with these variables. and I'd really appreciate it if you could help,

Thanks

[955999979]

@Xuanmeng-Zhang , I'm also curious about the variables including h_mea=pi/2, h_stddev=0.3, v_mean=pi/2, v_stddev=0.155 and fov=12 degree in your metadata used when training on FFHQ dataset. How did you get these statistics from FFHQ?

I'm not familiar with these variables. and I'd really appreciate it if you could help,

Thanks

Do you understand these parameters? I'm also having the same problem as you, looking forward to your reply, thank you！

[Xuanmeng-Zhang]

@AlbertHuyb @955999979 Hi, sorry for the late reply. We assume the face angle distribution is a Gaussian distribution, where the mean value is pi/2 (both in the vertical and horizontal directions). We set these parameters empirically. A more accurate way is to follow the eg3d to estimate the camera parameters from a pretrained face model. Please refer eg3d to for more details.

[955999979]

@Xuanmeng-Zhang Okay, thank you for your reply.