Cuda memory run out about rignet HOT 8 CLOSED

Dear Zhan-xu,

I solved this issue, in pairwise_distances function I used torch.cdist function, and in meanshift_cluster function I used torch.mm instead of K = K * weights.

Also please clarify: what randomness is getting different run results in current one model every time?

Thanks and good luck!!!

from rignet.

Thanks for the provided information. I will test this later and consider to change my code to it if this is consistently more memory efficient.

from rignet.

Hello Zhan-xu,
One more question, are there any way (any option to set) to get fixed number of rig joints?

from rignet.

Hi @Tigran1983, so the RigNet project focuses on how to predict various number of joints. If you want fixed number of joints, you may need to redesign the network, similar to segmenting the mesh to a fixed number of parts. Another simple way is only preserving the first N joints based on the density order (density from meanshift, calculated based on the bandwidth). However if the code originally predicted M<N joints, we cannot add more.

from rignet.

Thank you, if I'll try the second way mentioned by you, I'll let you know.

Best Regards,
Tigran

from rignet.

Hello dear Zhan-xu,

One more question: did you try to rig character face?
In my opinion if we could find or collect dataset of rigged faces, then your model can learn face rig too.
What do you think?

Regards,
Tigran

from rignet.

I didn't try to rig faces only, and seems many people are interested in this!

In general I feel if the rig is very detailed (contains dense helper joints for small part of muscles), then RigNet sometimes fails to recover them. See our limitations in the paper. This however is more common on face rigs because people expect the fine-grained control over facial expression.

On the other hand, if the training data are more consistent (with similar rig structure), the performance might get better. We trained the model on data from all categories for better generalization. If only trained with face data, I assume the general rig structure could be recovered.

from rignet.

Thank you, I think no one feels this network better than you.

from rignet.