search

zhanglonghao1992 / One-Shot_Free-View_Neural_Talking_Head_Synthesis

Pytorch implementation of paper "One-Shot Free-View Neural Talking-Head Synthesis for Video Conferencing"
Other
735 stars 141 forks source link

GPU #46

Open LiangHui6 opened 2 years ago

LiangHui6 commented 2 years ago

hi,teacher,Using RTX3090 GPU can run this training code and testing code?

zhanglonghao1992 commented 2 years ago

@LiangHui6 I use 8 * A100 for training. RTX3090 is enough for inference.

DinoMan commented 2 years ago

@zhanglonghao1992 . Hi thanks for sharing the code. How long does the training take on your setup? I am trying something similar and it seems to be quite slow.

zhanglonghao1992 commented 2 years ago

@DinoMan Training on VoxCeleb-v1 took several days. I use 8 * A100 for training.

DinoMan commented 2 years ago

@zhanglonghao1992 Thanks, that lines up to what I am seeing. Do you happen to know which part of the network is causing the slowdown. I have tried reducing the number of channels in each layer but that does not make enough of a difference.

zhanglonghao1992 commented 2 years ago

@DinoMan

  1. DO NOT estimate jacobians.
  2. If you train the SPADE generator, just reduce the number of the SPADEResnetBlock.