search

Doubiiu / CodeTalker

[CVPR 2023] CodeTalker: Speech-Driven 3D Facial Animation with Discrete Motion Prior
MIT License
515 stars 57 forks source link

Not converge when larger batch size is used in Stage2 #56

Open NickChang97 opened 1 year ago

NickChang97 commented 1 year ago

Hi, the default batch size is 1, did you try the larger batch size. In my experiments, when the batch size >= 4, the loss can not be converged to a satisfied results despite tuning various hyper-parameters.

Doubiiu commented 1 year ago

Hi, I didn't try that in our experiments. I think this VQ-Stuff is unstable in terms of training, thus a 'proper' combination of these hyper-parameters is a must to make it work.