facebookresearch/meshtalk

Training on VOCA/BIWI

Doubiiu opened this issue · 4 comments

Doubiiu commented

Hi,
I am trying to train/test MeshTalk on BIWI and vocaset, but I found the training loss and validation loss keep oscillating in stage 1. And in stage 2, the ce loss decreases to 0.2 while the the val loss is a U shape around 0.5. I know it is somehow overfitting, but I test it on training/testing set, it always gives "still" results. (Kind of effect of teacher-forcing?) Can you give some advice on this? Thanks very much! And if someone has successfully train that on BIWI/vocaset, I will be very appreciated for any help~

loss

alexanderrichard commented

I have never trained on any of these datasets, so I can't really address this issue here.

xuzheyuan624 commented

@Doubiiu Do u get a good results on these two dataset (vocadataset and BIWI) now? I have meet the same problem, can u give some advice?

haonanhe commented

@xuzheyuan624 I've also met the same problem? Have you solved it? I'll be really grateful if you could share some training settings with me.

galib360 commented

I am facing the same problem training BIWI and VOCASET with meshtalk. The losses I get is similar and in inference, it produces static/still animation.
@Doubiiu any advice on how to solve it? Did you need to change the default hyperparams to get desired results? Would be great if you could share some insights. :)