search

dmsm / HodgeNet

Code for HodgeNet: Learning Spectral Geometry on Triangle Meshes, in SIGGRAPH 2021.
MIT License
79 stars 9 forks source link

GPU training support #4

Closed kimihailv closed 3 years ago

kimihailv commented 3 years ago

Hello! Thank you for code. Is it possible to train HodgeNet on gpu partially? For instance do all calculations before eigenvalues computation on cuda, send result to cpu and then compute eigenvalues. Or there is the problem in backward pass?

dmsm commented 3 years ago

Thanks for your interest! Indeed, everything except the eigencomputation (including the backwards pass) is pretty GPU-friendly. However, when we tried partially training on GPU, we did not notice much of a difference in runtime compared to running everything on CPU. This is probably because our actual network is fairly small (e.g., the number/size of fully connected layers), and transferring data between CPU/GPU has overhead cost.

On Oct 20, 2021, at 4:08 PM, kimihailv @.***> wrote:

Hello! Thank you for code. Is it possible to train HodgeNet on gpu partially? For instance do all calculations before eigenvalues computation on cuda, send result to cpu and then compute eigenvalues. Or there is the problem in backward pass?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/dmsm/HodgeNet/issues/4, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAJ63XQNOEVEELZFUOTIW4LUH4OT7ANCNFSM5GMPHJAQ. Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.