About retraining EG3D

[hhuyujie]

Thank you for releasing the code. I use the FFHQ-image1024x1024 dataset and camera parameters predicted by Deep3DRecon to retrain EG3D. The following command is executed:
python train.py --outdir=training-runs --cfg=ffhq --data=FFHQ_512.zip --gpus=1 --batch=1 --mbstd-group=1 --gamma=1 --gen_pose_cond=True

So far I have trained 152k images, but the intermediate results are still quite strange, so I want to ask what might be the problem.

[Jingxiang Sun]

If you use such a small batch size, you should change gamma to a larger value such as 20 empirically。

[MartinMao1101]

hi there, i am trying to use this impressive model for my thesis and thank you for raising this issue. Till now i also have tried to run the training for quite a while and there seems to be some problems with my cam params json file.
I would like to ask about how did you use the Deep3DRecon to do the prediction? Looking forward to your reply and very appreciate that! Btw if it is ok for you, my email is [e-mail] and Wechat ningmao1101.
Thanks a lot in advance!
Ning