search

nv-tlabs / NKSR

[CVPR 2023 Highlight] Neural Kernel Surface Reconstruction
https://research.nvidia.com/labs/toronto-ai/NKSR
Other
735 stars 43 forks source link

Questions about Training Process #74

Closed Gy920 closed 1 month ago

Gy920 commented 1 month ago

Hi,

Thank you for your outstanding work. I have a few questions regarding the training process:

  1. Training Time:

    • How long does it typically take to train on the ShapeNet and Points2Surf datasets?

  2. Batch Size Limitation:

    • I noticed that the batch size on a single GPU can only be set to 1. Are there any methods or configurations that could allow for a larger batch size?

Thanks for your help!

heiwang1997 commented 1 month ago

Training time -- it took ~4-5 days to train on ShapeNet on a single V100. Points2Surf is easier and takes ~2 days to converge.

Batch size -- good point and we release the batched version here at: https://github.com/nv-tlabs/XCube/tree/main/ext/nksr-cuda.

Currently we're pending an internal schedule so there's not a lot of documentation on that. I will update things hopefully in the beginning of August to incorporate that.

Thanks!!!

Gy920 commented 1 month ago

Thanks for your reply. I have no further questions.