train.py fail (preloads images to 95%)

NVlabs / neuralangelo

Official implementation of "Neuralangelo: High-Fidelity Neural Surface Reconstruction" (CVPR 2023)

https://research.nvidia.com/labs/dir/neuralangelo/

Other

4.4k stars 390 forks source link

train.py fail (preloads images to 95%) #178

Closed MicKkmack closed 10 months ago

MicKkmack commented 10 months ago

Anyone able to help with the above train.py issue?

I've set the correct dict_size and dim for a 3070ti and I've tried reducing settings like iterations and batch size to see if that would help, but with no success.

MicKkmack commented 10 months ago

I've set preload in the base.yaml file to false and that seems to at least get it training, not sure if it'll cause issues

MicKkmack commented 10 months ago

Too many images , I had about 2400 images. I've tried around 100 images / 300 images / 600 images and that seems to be the sweat spot. At least for my 3070 graphics card