Training parameters

[UttaranB127]

Hello,

I am trying to train MeshTalk on the VOCA dataset, however, the loss value explodes if I use a learning rate 1e-4 or higher, and keeps oscillating in the range of 0.2 if I use a lower learning rate (this does not lead to realistic results). I was wondering what training parameters were used in the paper?

I am using the following parameters:
no. of frames, T = 128
optimizer SGD with lr=9e-5 (at the moment), momentum=0.9, nesterov=True
M_upper = 5 and M_lower = 5
batch_size = 16

Thanks for any help!

[alexanderrichard]

Hi,
here's what I used in the paper:

no. of frames T = 32
optimizer: Adam, lr=0.001 (everything else default)
upper/lower mask in [0,1] (1=vertex of interest, 0=vertex of no interest)
batch_size=8

I want to mention that the VOCA dataset has almost no upper face motion, particularly, the meshes do not have eye lids, so you can't get eye blinks etc. You might get a broader variety of mouth motion though.

[UttaranB127]

Thanks, let me try with these parameters! I agree, VOCA does not have too much upper face motion. I wonder if there are any other datasets that do have it? I have not found many datasets so far.

[alexanderrichard]

I don't believe so. We are curating some of our data for public release to close this gap, I will add a link to the repo once available.

[UttaranB127]

Yes, that would be super useful! Thanks for your response.

[xiapengchng]

Hello,

I am trying to train MeshTalk on the VOCA dataset, however, the loss value explodes if I use a learning rate 1e-4 or higher, and keeps oscillating in the range of 0.2 if I use a lower learning rate (this does not lead to realistic results). I was wondering what training parameters were used in the paper?

I am using the following parameters: no. of frames, T = 128 optimizer SGD with lr=9e-5 (at the moment), momentum=0.9, nesterov=True M_upper = 5 and M_lower = 5 batch_size = 16

Thanks for any help!

Do you get the result on VOCA, how to split the upper face and lower face?